{
  "schema": "fak-llamacpp-index/1",
  "generated_at": "2026-08-26T22:20:00Z",
  "cutoff": "2026-08-26T22:20:00Z",
  "parent_issue": 9270,
  "method": {
    "scope": "Complete obtainable open issue/PR corpus at cutoff plus pinned recursive tree, recent commits, releases, and exact local full-text issue searches. Closed upstream history is represented by the recent-commit/release slices and named closed references, not claimed as a complete semantic classification.",
    "inclusion": "Every upstream open item is retained once. Categories are deterministic regex tags and are non-exclusive. Ticket construction requires a manual product-centrality and duplicate audit.",
    "exclusion": "Non-actionable UI/docs/platform-only records remain indexed but are not promoted. Related systems are referenced only where they alter a llama.cpp-derived decision.",
    "limitations": [
      "GitHub REST open_issues_count moved from 2226 at initial fetch to 2227 later; the corpus is pinned to the earlier complete 2226-item pagination and explicit cutoff.",
      "The recursive tree is pinned to HEAD 925e1179; later release b10642 was published at 22:00:33Z without a newer master commit.",
      "GitHub issue search is full-text but ranking/matching is forge-defined; exact query receipts and the union are retained.",
      "Semantic prioritization is reviewed, not inferred solely from categories."
    ]
  },
  "upstream": {
    "repository": "ggml-org/llama.cpp",
    "default_branch": "master",
    "commit": "925e1179947ea0c0ebfb0032df18af3a729822be",
    "commit_date": "2026-08-26T21:34:28Z",
    "commit_url": "https://github.com/ggml-org/llama.cpp/commit/925e1179947ea0c0ebfb0032df18af3a729822be",
    "tree": {
      "entries": 3871,
      "blobs": 3498,
      "trees": 373,
      "truncated": false,
      "top_level_counts": {
        ".clang-format": 1,
        ".clang-tidy": 1,
        ".devops": 26,
        ".dockerignore": 1,
        ".ecrc": 1,
        ".editorconfig": 1,
        ".flake8": 1,
        ".gemini": 2,
        ".github": 84,
        ".gitignore": 1,
        ".gitmodules": 1,
        ".pi": 3,
        ".pre-commit-config.yaml": 1,
        "AGENTS.md": 1,
        "AUTHORS": 1,
        "CLAUDE.md": 1,
        "CMakeLists.txt": 1,
        "CMakePresets.json": 1,
        "CODEOWNERS": 1,
        "CONTRIBUTING.md": 1,
        "LICENSE": 1,
        "Makefile": 1,
        "README.md": 1,
        "SECURITY.md": 1,
        "app": 4,
        "benches": 11,
        "build-xcframework.sh": 1,
        "ci": 4,
        "cmake": 14,
        "common": 75,
        "conversion": 90,
        "convert_hf_to_gguf.py": 1,
        "convert_hf_to_gguf_update.py": 1,
        "convert_llama_ggml_to_gguf.py": 1,
        "convert_lora_to_gguf.py": 1,
        "docs": 75,
        "examples": 302,
        "flake.nix": 1,
        "ggml": 1369,
        "gguf-py": 32,
        "grammars": 10,
        "include": 3,
        "licenses": 2,
        "media": 12,
        "models": 121,
        "mypy.ini": 1,
        "pocs": 6,
        "pyproject.toml": 1,
        "pyrightconfig.json": 1,
        "requirements.txt": 1,
        "requirements": 13,
        "scripts": 80,
        "skills": 5,
        "src": 217,
        "tests": 83,
        "tools": 1159,
        "ty.toml": 1,
        "vendor": 38
      }
    },
    "forge": {
      "open_pages": 23,
      "per_page": 100,
      "open_item_count": 2226,
      "open_issue_count": 830,
      "open_pr_count": 1396,
      "oldest_created_at": "2023-03-17T08:32:33Z",
      "newest_created_at": "2026-08-26T21:10:02Z",
      "category_counts": {
        "models_architectures": 2040,
        "scheduling_serving": 1336,
        "correctness_testing": 1684,
        "quantization": 1232,
        "memory_kv_context": 1109,
        "build_platform_ops": 1550,
        "api_ui_docs": 1139,
        "training_adapters": 232,
        "other": 15,
        "native_backends": 1436,
        "loader_conversion": 550,
        "multimodal_embedding": 476,
        "structured_generation_tools": 402,
        "speculative_generation": 306
      },
      "raw_sha256": "4855ab1dcc94b884dffcb437da4df87892c4821811b88029a376add8eb757510"
    },
    "items": [
      {
        "number": 231,
        "kind": "issue",
        "state": "open",
        "title": "Study how LM Evaluation Harness works and try to implement it",
        "labels": [
          "enhancement",
          "help wanted",
          "high priority",
          "generation quality",
          "research 🔬"
        ],
        "created_at": "2023-03-17T08:32:33Z",
        "updated_at": "2026-02-17T06:59:00Z",
        "closed_at": null,
        "comments": 10,
        "url": "https://github.com/ggml-org/llama.cpp/issues/231",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 744,
        "kind": "pr",
        "state": "open",
        "title": "Add \"-e\"/\"--eval-threads\" to distinguish thread counts for single-token eval and prompt eval",
        "labels": [
          "threading"
        ],
        "created_at": "2023-04-03T18:19:37Z",
        "updated_at": "2026-06-17T16:25:33Z",
        "closed_at": null,
        "comments": 11,
        "url": "https://github.com/ggml-org/llama.cpp/pull/744",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 813,
        "kind": "pr",
        "state": "open",
        "title": "Optimize locking behavior",
        "labels": [
          "threading"
        ],
        "created_at": "2023-04-06T14:19:10Z",
        "updated_at": "2025-12-11T01:30:03Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/pull/813",
        "categories": [
          "other"
        ]
      },
      {
        "number": 850,
        "kind": "pr",
        "state": "open",
        "title": "Run several single thread operators parellel",
        "labels": [
          "threading"
        ],
        "created_at": "2023-04-08T13:44:50Z",
        "updated_at": "2023-09-15T14:23:18Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/pull/850",
        "categories": [
          "correctness_testing"
        ]
      },
      {
        "number": 851,
        "kind": "pr",
        "state": "open",
        "title": "Use Threadpool to schedule the work",
        "labels": [
          "threading"
        ],
        "created_at": "2023-04-08T14:19:55Z",
        "updated_at": "2026-07-27T15:40:53Z",
        "closed_at": null,
        "comments": 9,
        "url": "https://github.com/ggml-org/llama.cpp/pull/851",
        "categories": [
          "native_backends",
          "scheduling_serving"
        ]
      },
      {
        "number": 1015,
        "kind": "pr",
        "state": "open",
        "title": "Add mmap pages stats (disabled by default)",
        "labels": [],
        "created_at": "2023-04-16T16:24:11Z",
        "updated_at": "2023-09-15T14:22:49Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/1015",
        "categories": [
          "other"
        ]
      },
      {
        "number": 1058,
        "kind": "pr",
        "state": "open",
        "title": "Add a option to force the token end of text apears even on interative, and also shows loading porcentage",
        "labels": [],
        "created_at": "2023-04-19T12:09:09Z",
        "updated_at": "2023-09-15T14:22:29Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/1058",
        "categories": [
          "loader_conversion"
        ]
      },
      {
        "number": 1103,
        "kind": "pr",
        "state": "open",
        "title": "llama : quantize attention results",
        "labels": [
          "demo"
        ],
        "created_at": "2023-04-21T14:45:59Z",
        "updated_at": "2025-12-03T11:12:37Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/1103",
        "categories": [
          "quantization",
          "models_architectures",
          "build_platform_ops"
        ]
      },
      {
        "number": 1132,
        "kind": "pr",
        "state": "open",
        "title": "main: add pledge call on OpenBSD",
        "labels": [],
        "created_at": "2023-04-22T21:13:22Z",
        "updated_at": "2023-09-15T14:21:21Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/1132",
        "categories": [
          "scheduling_serving",
          "api_ui_docs"
        ]
      },
      {
        "number": 1190,
        "kind": "pr",
        "state": "open",
        "title": "fix(LoRA): debugging",
        "labels": [],
        "created_at": "2023-04-26T15:33:52Z",
        "updated_at": "2023-09-15T14:21:12Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/1190",
        "categories": [
          "correctness_testing",
          "training_adapters"
        ]
      },
      {
        "number": 1204,
        "kind": "pr",
        "state": "open",
        "title": "Create run.py",
        "labels": [
          "enhancement",
          "script",
          "Review Complexity : Low",
          "python",
          "obsolete?"
        ],
        "created_at": "2023-04-27T14:39:49Z",
        "updated_at": "2024-05-25T14:05:43Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/1204",
        "categories": [
          "models_architectures",
          "api_ui_docs"
        ]
      },
      {
        "number": 1278,
        "kind": "pr",
        "state": "open",
        "title": "Implement get_num_physical_cores() for Windows",
        "labels": [],
        "created_at": "2023-05-02T14:25:53Z",
        "updated_at": "2023-09-15T14:20:40Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/pull/1278",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 1299,
        "kind": "issue",
        "state": "open",
        "title": "Test replit-code-v1-3b model",
        "labels": [
          "help wanted",
          "model"
        ],
        "created_at": "2023-05-03T15:47:01Z",
        "updated_at": "2023-05-05T20:12:38Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/issues/1299",
        "categories": [
          "models_architectures",
          "correctness_testing",
          "training_adapters"
        ]
      },
      {
        "number": 1472,
        "kind": "pr",
        "state": "open",
        "title": "[Research] Steering vectors",
        "labels": [
          "research 🔬"
        ],
        "created_at": "2023-05-16T00:14:50Z",
        "updated_at": "2023-09-18T13:59:24Z",
        "closed_at": null,
        "comments": 30,
        "url": "https://github.com/ggml-org/llama.cpp/pull/1472",
        "categories": [
          "models_architectures",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 1504,
        "kind": "pr",
        "state": "open",
        "title": "Upgrade v1/v2 format to v3 by leveraging quantize",
        "labels": [],
        "created_at": "2023-05-17T15:44:26Z",
        "updated_at": "2023-09-15T14:20:01Z",
        "closed_at": null,
        "comments": 11,
        "url": "https://github.com/ggml-org/llama.cpp/pull/1504",
        "categories": [
          "quantization",
          "models_architectures",
          "correctness_testing"
        ]
      },
      {
        "number": 1507,
        "kind": "pr",
        "state": "open",
        "title": "ggml : spread compute across threads in chunks",
        "labels": [
          "threading",
          "demo"
        ],
        "created_at": "2023-05-17T17:12:10Z",
        "updated_at": "2025-09-29T09:57:43Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/pull/1507",
        "categories": [
          "build_platform_ops"
        ]
      },
      {
        "number": 1660,
        "kind": "pr",
        "state": "open",
        "title": "Llama cpp low level python bindings",
        "labels": [],
        "created_at": "2023-06-01T06:18:02Z",
        "updated_at": "2025-06-01T09:14:33Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/1660",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 1714,
        "kind": "issue",
        "state": "open",
        "title": "Support CoreML like whisper.cpp?",
        "labels": [
          "help wanted",
          "performance",
          "macos"
        ],
        "created_at": "2023-06-06T09:23:08Z",
        "updated_at": "2025-02-20T14:52:01Z",
        "closed_at": null,
        "comments": 11,
        "url": "https://github.com/ggml-org/llama.cpp/issues/1714",
        "categories": [
          "models_architectures"
        ]
      },
      {
        "number": 1834,
        "kind": "pr",
        "state": "open",
        "title": "Added Arbitrary mixed quantization",
        "labels": [
          "research 🔬",
          "Less than 4 bits"
        ],
        "created_at": "2023-06-13T09:11:49Z",
        "updated_at": "2023-09-15T14:17:20Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/pull/1834",
        "categories": [
          "quantization",
          "models_architectures",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 1865,
        "kind": "issue",
        "state": "open",
        "title": "[IDEA] Global token enhancement/depression",
        "labels": [
          "help wanted",
          "research 🔬"
        ],
        "created_at": "2023-06-15T02:24:07Z",
        "updated_at": "2023-06-17T10:12:22Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/1865",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 1897,
        "kind": "pr",
        "state": "open",
        "title": "Disable _O_WTEXT when using main in MinGW",
        "labels": [],
        "created_at": "2023-06-16T16:10:46Z",
        "updated_at": "2024-07-10T14:55:48Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/1897",
        "categories": [
          "native_backends",
          "models_architectures"
        ]
      },
      {
        "number": 1986,
        "kind": "pr",
        "state": "open",
        "title": "Draft: #1776 making bos and eos available for user input",
        "labels": [],
        "created_at": "2023-06-24T23:54:43Z",
        "updated_at": "2023-09-15T14:16:35Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/1986",
        "categories": [
          "other"
        ]
      },
      {
        "number": 2026,
        "kind": "pr",
        "state": "open",
        "title": "Example work stealing chunked task allocator for issue #291",
        "labels": [],
        "created_at": "2023-06-27T20:21:10Z",
        "updated_at": "2023-09-15T14:16:15Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/2026",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 2029,
        "kind": "pr",
        "state": "open",
        "title": "Avoid unused constant warnings",
        "labels": [
          "refactoring",
          "Review Complexity : Low"
        ],
        "created_at": "2023-06-28T01:05:13Z",
        "updated_at": "2024-05-10T13:09:08Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/2029",
        "categories": [
          "other"
        ]
      },
      {
        "number": 2037,
        "kind": "pr",
        "state": "open",
        "title": "try to fix compile warnings on macOS, address issue #2036",
        "labels": [],
        "created_at": "2023-06-28T10:32:41Z",
        "updated_at": "2023-09-15T14:15:59Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/2037",
        "categories": [
          "models_architectures",
          "build_platform_ops"
        ]
      },
      {
        "number": 2069,
        "kind": "pr",
        "state": "open",
        "title": "Avoid allocating beyond MTLDevice.recommendedMaxWorkingSetSize",
        "labels": [],
        "created_at": "2023-07-02T04:51:55Z",
        "updated_at": "2023-09-15T14:15:22Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/2069",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "multimodal_embedding"
        ]
      },
      {
        "number": 2072,
        "kind": "pr",
        "state": "open",
        "title": "[llama] Add resegment post processing of tokenizer",
        "labels": [],
        "created_at": "2023-07-02T14:06:41Z",
        "updated_at": "2023-09-15T14:15:12Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/pull/2072",
        "categories": [
          "models_architectures",
          "loader_conversion"
        ]
      },
      {
        "number": 2164,
        "kind": "issue",
        "state": "open",
        "title": "mpi : attempt inference of 65B LLaMA on a cluster of Raspberry Pis",
        "labels": [
          "help wanted",
          "hardware",
          "research 🔬"
        ],
        "created_at": "2023-07-10T16:12:22Z",
        "updated_at": "2025-11-29T12:50:33Z",
        "closed_at": null,
        "comments": 54,
        "url": "https://github.com/ggml-org/llama.cpp/issues/2164",
        "categories": [
          "quantization",
          "models_architectures",
          "build_platform_ops"
        ]
      },
      {
        "number": 2168,
        "kind": "pr",
        "state": "open",
        "title": "Porting MPI PR to Darwin OpenMPI",
        "labels": [],
        "created_at": "2023-07-10T22:52:00Z",
        "updated_at": "2023-09-15T14:14:49Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/2168",
        "categories": [
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 2209,
        "kind": "issue",
        "state": "open",
        "title": "Using MPI w/ 65b model but each node uses the full RAM.",
        "labels": [
          "help wanted"
        ],
        "created_at": "2023-07-13T02:57:34Z",
        "updated_at": "2023-07-16T19:37:33Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/2209",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 2217,
        "kind": "pr",
        "state": "open",
        "title": "Add CFG to server",
        "labels": [],
        "created_at": "2023-07-13T19:55:29Z",
        "updated_at": "2025-01-13T17:31:48Z",
        "closed_at": null,
        "comments": 10,
        "url": "https://github.com/ggml-org/llama.cpp/pull/2217",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 2233,
        "kind": "pr",
        "state": "open",
        "title": "Add flag to make reverse prompt case insensitive",
        "labels": [
          "obsolete?"
        ],
        "created_at": "2023-07-15T20:00:10Z",
        "updated_at": "2024-06-09T05:33:12Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/2233",
        "categories": [
          "models_architectures"
        ]
      },
      {
        "number": 2246,
        "kind": "pr",
        "state": "open",
        "title": "Llama training finetuning interface",
        "labels": [],
        "created_at": "2023-07-17T01:50:08Z",
        "updated_at": "2023-09-15T14:14:06Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/2246",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "structured_generation_tools",
          "training_adapters"
        ]
      },
      {
        "number": 2296,
        "kind": "pr",
        "state": "open",
        "title": "Consider adding repo-review badge to README",
        "labels": [],
        "created_at": "2023-07-20T15:26:58Z",
        "updated_at": "2023-09-15T14:13:35Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/2296",
        "categories": [
          "api_ui_docs"
        ]
      },
      {
        "number": 2434,
        "kind": "pr",
        "state": "open",
        "title": "ggml : PoC for normalizing weights for better quantization packing",
        "labels": [
          "demo"
        ],
        "created_at": "2023-07-28T08:07:12Z",
        "updated_at": "2023-09-15T14:12:03Z",
        "closed_at": null,
        "comments": 16,
        "url": "https://github.com/ggml-org/llama.cpp/pull/2434",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 2593,
        "kind": "pr",
        "state": "open",
        "title": "Implementation of a sequence repetition penalty sampler",
        "labels": [
          "enhancement",
          "generation quality",
          "need feedback"
        ],
        "created_at": "2023-08-12T20:39:25Z",
        "updated_at": "2023-12-30T16:13:44Z",
        "closed_at": null,
        "comments": 49,
        "url": "https://github.com/ggml-org/llama.cpp/pull/2593",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 2623,
        "kind": "pr",
        "state": "open",
        "title": "one can now specify where ggml-metal.metal file is with en variable GGML_METAL_PATH",
        "labels": [],
        "created_at": "2023-08-15T14:51:10Z",
        "updated_at": "2023-09-15T14:10:06Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/2623",
        "categories": [
          "native_backends",
          "build_platform_ops"
        ]
      },
      {
        "number": 2728,
        "kind": "pr",
        "state": "open",
        "title": "Add date and commit hash to gguf metadata",
        "labels": [
          "obsolete?"
        ],
        "created_at": "2023-08-23T00:33:01Z",
        "updated_at": "2024-05-25T14:24:00Z",
        "closed_at": null,
        "comments": 18,
        "url": "https://github.com/ggml-org/llama.cpp/pull/2728",
        "categories": [
          "quantization",
          "models_architectures",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 2730,
        "kind": "pr",
        "state": "open",
        "title": "Fix docs",
        "labels": [
          "obsolete?"
        ],
        "created_at": "2023-08-23T01:09:04Z",
        "updated_at": "2024-05-25T14:22:29Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/2730",
        "categories": [
          "loader_conversion"
        ]
      },
      {
        "number": 2783,
        "kind": "issue",
        "state": "open",
        "title": "llama : tool for evaluating quantization results per layer",
        "labels": [
          "enhancement",
          "generation quality",
          "roadmap"
        ],
        "created_at": "2023-08-25T10:02:47Z",
        "updated_at": "2025-12-25T00:28:13Z",
        "closed_at": null,
        "comments": 9,
        "url": "https://github.com/ggml-org/llama.cpp/issues/2783",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 2809,
        "kind": "pr",
        "state": "open",
        "title": "cuda: 1.2x faster dequantization kernel",
        "labels": [
          "performance",
          "Review Complexity : High"
        ],
        "created_at": "2023-08-26T12:56:21Z",
        "updated_at": "2024-05-25T15:40:25Z",
        "closed_at": null,
        "comments": 16,
        "url": "https://github.com/ggml-org/llama.cpp/pull/2809",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 2891,
        "kind": "pr",
        "state": "open",
        "title": "metal: template for mat-vec multiplication kernels",
        "labels": [
          "performance"
        ],
        "created_at": "2023-08-30T03:04:47Z",
        "updated_at": "2023-09-15T13:58:38Z",
        "closed_at": null,
        "comments": 19,
        "url": "https://github.com/ggml-org/llama.cpp/pull/2891",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 3025,
        "kind": "pr",
        "state": "open",
        "title": "Model magic",
        "labels": [
          "obsolete?"
        ],
        "created_at": "2023-09-05T11:14:01Z",
        "updated_at": "2024-05-25T14:02:42Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/pull/3025",
        "categories": [
          "models_architectures",
          "scheduling_serving"
        ]
      },
      {
        "number": 3093,
        "kind": "pr",
        "state": "open",
        "title": "Adding SqueezeLLM Support",
        "labels": [
          "demo"
        ],
        "created_at": "2023-09-09T04:37:56Z",
        "updated_at": "2023-09-27T11:56:53Z",
        "closed_at": null,
        "comments": 11,
        "url": "https://github.com/ggml-org/llama.cpp/pull/3093",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "multimodal_embedding",
          "loader_conversion",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 3102,
        "kind": "pr",
        "state": "open",
        "title": "Don't highlight console session as Java.",
        "labels": [],
        "created_at": "2023-09-09T20:21:14Z",
        "updated_at": "2023-09-15T14:04:55Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/3102",
        "categories": [
          "api_ui_docs"
        ]
      },
      {
        "number": 3138,
        "kind": "pr",
        "state": "open",
        "title": "Adding Codcov badge",
        "labels": [
          "Review Complexity : Low",
          "devops"
        ],
        "created_at": "2023-09-12T08:05:23Z",
        "updated_at": "2024-05-25T14:17:44Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/3138",
        "categories": [
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 3219,
        "kind": "pr",
        "state": "open",
        "title": "log.h improvements",
        "labels": [
          "obsolete?"
        ],
        "created_at": "2023-09-16T17:41:25Z",
        "updated_at": "2024-05-25T14:13:16Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/3219",
        "categories": [
          "memory_kv_context",
          "build_platform_ops"
        ]
      },
      {
        "number": 3234,
        "kind": "pr",
        "state": "open",
        "title": "llama : store non-RoPEd K cache",
        "labels": [
          "demo"
        ],
        "created_at": "2023-09-17T20:50:12Z",
        "updated_at": "2024-01-25T22:57:09Z",
        "closed_at": null,
        "comments": 11,
        "url": "https://github.com/ggml-org/llama.cpp/pull/3234",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context"
        ]
      },
      {
        "number": 3294,
        "kind": "issue",
        "state": "open",
        "title": "Request: Nougat OCR Integration",
        "labels": [
          "help wanted",
          "model"
        ],
        "created_at": "2023-09-21T06:29:29Z",
        "updated_at": "2024-07-02T23:30:28Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/issues/3294",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 3334,
        "kind": "pr",
        "state": "open",
        "title": "[MPI] Add support for per-node options, thread counts, and layer allocations",
        "labels": [
          "build",
          "examples",
          "server",
          "ggml"
        ],
        "created_at": "2023-09-26T00:24:10Z",
        "updated_at": "2025-02-05T15:50:32Z",
        "closed_at": null,
        "comments": 44,
        "url": "https://github.com/ggml-org/llama.cpp/pull/3334",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 3431,
        "kind": "pr",
        "state": "open",
        "title": "Add basic support for function calls in oai python server",
        "labels": [
          "Review Complexity : Medium",
          "server/api"
        ],
        "created_at": "2023-10-01T21:49:26Z",
        "updated_at": "2025-11-29T12:50:32Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/3431",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 3565,
        "kind": "pr",
        "state": "open",
        "title": "Layer skipping/self-speculation demo",
        "labels": [
          "research 🔬",
          "demo"
        ],
        "created_at": "2023-10-10T01:08:48Z",
        "updated_at": "2023-12-14T03:02:30Z",
        "closed_at": null,
        "comments": 37,
        "url": "https://github.com/ggml-org/llama.cpp/pull/3565",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 3628,
        "kind": "issue",
        "state": "open",
        "title": "ci : fix Docker workflow",
        "labels": [
          "help wanted",
          "build"
        ],
        "created_at": "2023-10-15T06:37:47Z",
        "updated_at": "2024-03-20T01:07:19Z",
        "closed_at": null,
        "comments": 16,
        "url": "https://github.com/ggml-org/llama.cpp/issues/3628",
        "categories": [
          "models_architectures",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 3688,
        "kind": "pr",
        "state": "open",
        "title": "console: fix getwchar failing when LC_ALL undefined",
        "labels": [],
        "created_at": "2023-10-19T19:47:17Z",
        "updated_at": "2023-12-14T03:01:32Z",
        "closed_at": null,
        "comments": 11,
        "url": "https://github.com/ggml-org/llama.cpp/pull/3688",
        "categories": [
          "models_architectures",
          "multimodal_embedding"
        ]
      },
      {
        "number": 3777,
        "kind": "pr",
        "state": "open",
        "title": "AMX isa Native addition",
        "labels": [],
        "created_at": "2023-10-25T12:34:02Z",
        "updated_at": "2024-03-19T07:35:01Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/3777",
        "categories": [
          "quantization",
          "models_architectures",
          "build_platform_ops"
        ]
      },
      {
        "number": 3845,
        "kind": "pr",
        "state": "open",
        "title": "OpenCL: Pass src0 offset as kernel argument instead of global offset",
        "labels": [],
        "created_at": "2023-10-29T09:01:24Z",
        "updated_at": "2023-12-14T02:59:05Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/3845",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing"
        ]
      },
      {
        "number": 3957,
        "kind": "issue",
        "state": "open",
        "title": "GGUF endianness cannot be determined from GGUF itself",
        "labels": [
          "enhancement",
          "good first issue",
          "breaking change"
        ],
        "created_at": "2023-11-05T14:00:47Z",
        "updated_at": "2026-04-27T04:11:04Z",
        "closed_at": null,
        "comments": 20,
        "url": "https://github.com/ggml-org/llama.cpp/issues/3957",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "loader_conversion"
        ]
      },
      {
        "number": 3984,
        "kind": "pr",
        "state": "open",
        "title": "contrastive: PoC for improving reasoning via contrastive decoding",
        "labels": [],
        "created_at": "2023-11-08T01:18:30Z",
        "updated_at": "2023-12-14T02:55:47Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/3984",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "api_ui_docs"
        ]
      },
      {
        "number": 4012,
        "kind": "pr",
        "state": "open",
        "title": "added lto",
        "labels": [],
        "created_at": "2023-11-09T21:20:07Z",
        "updated_at": "2023-12-14T02:53:50Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/4012",
        "categories": [
          "models_architectures",
          "build_platform_ops"
        ]
      },
      {
        "number": 4028,
        "kind": "pr",
        "state": "open",
        "title": "Fix bug where POST /infill doesn't work without prompt argument",
        "labels": [],
        "created_at": "2023-11-10T20:29:06Z",
        "updated_at": "2023-12-14T02:52:51Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/4028",
        "categories": [
          "models_architectures",
          "correctness_testing"
        ]
      },
      {
        "number": 4062,
        "kind": "pr",
        "state": "open",
        "title": "gguf-py readme example fixes for keys: general.architecture and general.alignment",
        "labels": [],
        "created_at": "2023-11-13T17:47:06Z",
        "updated_at": "2023-12-14T02:52:25Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/4062",
        "categories": [
          "quantization",
          "models_architectures",
          "loader_conversion",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 4065,
        "kind": "pr",
        "state": "open",
        "title": "segfault with ebnf",
        "labels": [
          "obsolete?"
        ],
        "created_at": "2023-11-13T20:50:19Z",
        "updated_at": "2024-05-25T14:10:06Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/4065",
        "categories": [
          "models_architectures",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 4070,
        "kind": "pr",
        "state": "open",
        "title": "Update gpt2 preprocess and add deepseek coder preprocess",
        "labels": [],
        "created_at": "2023-11-14T09:51:12Z",
        "updated_at": "2024-04-18T06:52:46Z",
        "closed_at": null,
        "comments": 16,
        "url": "https://github.com/ggml-org/llama.cpp/pull/4070",
        "categories": [
          "quantization",
          "models_architectures",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 4085,
        "kind": "issue",
        "state": "open",
        "title": "metal : compile-time kernel args and params",
        "labels": [
          "performance",
          "research 🔬",
          "roadmap"
        ],
        "created_at": "2023-11-15T11:09:39Z",
        "updated_at": "2026-07-15T03:00:54Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/issues/4085",
        "categories": [
          "native_backends",
          "models_architectures",
          "build_platform_ops"
        ]
      },
      {
        "number": 4105,
        "kind": "pr",
        "state": "open",
        "title": "Update api_like_OAI.py server script to support text in mixed-mode data",
        "labels": [],
        "created_at": "2023-11-17T00:12:28Z",
        "updated_at": "2023-12-14T02:50:22Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/4105",
        "categories": [
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 4106,
        "kind": "pr",
        "state": "open",
        "title": "Token to piece (#1)",
        "labels": [],
        "created_at": "2023-11-17T01:17:01Z",
        "updated_at": "2023-11-20T15:43:46Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/4106",
        "categories": [
          "models_architectures",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 4218,
        "kind": "issue",
        "state": "open",
        "title": "llama : speed-up grammar sampling",
        "labels": [
          "performance",
          "refactoring",
          "roadmap"
        ],
        "created_at": "2023-11-25T17:04:06Z",
        "updated_at": "2025-11-08T03:59:23Z",
        "closed_at": null,
        "comments": 41,
        "url": "https://github.com/ggml-org/llama.cpp/issues/4218",
        "categories": [
          "models_architectures",
          "structured_generation_tools"
        ]
      },
      {
        "number": 4267,
        "kind": "pr",
        "state": "open",
        "title": "Input Temperature & Output Temperature",
        "labels": [],
        "created_at": "2023-11-30T13:14:13Z",
        "updated_at": "2023-12-14T02:46:15Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/4267",
        "categories": [
          "native_backends",
          "models_architectures",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 4313,
        "kind": "pr",
        "state": "open",
        "title": "A step ahead supporting NousResearch/Nous-Hermes-2-Vision",
        "labels": [
          "examples",
          "Review Complexity : Low"
        ],
        "created_at": "2023-12-03T19:54:26Z",
        "updated_at": "2024-05-10T15:09:10Z",
        "closed_at": null,
        "comments": 9,
        "url": "https://github.com/ggml-org/llama.cpp/pull/4313",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 4355,
        "kind": "pr",
        "state": "open",
        "title": "Fix typos in code.",
        "labels": [],
        "created_at": "2023-12-07T05:37:16Z",
        "updated_at": "2023-12-14T02:43:14Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/4355",
        "categories": [
          "models_architectures"
        ]
      },
      {
        "number": 4491,
        "kind": "issue",
        "state": "open",
        "title": "deprecate llama_batch_get_one and llama_get_logits",
        "labels": [
          "refactoring"
        ],
        "created_at": "2023-12-16T03:15:15Z",
        "updated_at": "2025-07-23T12:32:41Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/issues/4491",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "api_ui_docs"
        ]
      },
      {
        "number": 4504,
        "kind": "pr",
        "state": "open",
        "title": "Support for CUDA/HIP on OpenCL/Level0 via chipStar",
        "labels": [],
        "created_at": "2023-12-17T01:05:32Z",
        "updated_at": "2023-12-18T00:21:02Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/4504",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 4543,
        "kind": "pr",
        "state": "open",
        "title": "[Review] Merge PowerInfer with llama.cpp mainline",
        "labels": [],
        "created_at": "2023-12-20T08:02:29Z",
        "updated_at": "2026-01-28T03:41:58Z",
        "closed_at": null,
        "comments": 10,
        "url": "https://github.com/ggml-org/llama.cpp/pull/4543",
        "categories": [
          "models_architectures"
        ]
      },
      {
        "number": 4571,
        "kind": "pr",
        "state": "open",
        "title": "Initial import of OpenSHMEM support",
        "labels": [],
        "created_at": "2023-12-21T18:39:32Z",
        "updated_at": "2023-12-25T21:01:24Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/pull/4571",
        "categories": [
          "models_architectures",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 4574,
        "kind": "issue",
        "state": "open",
        "title": "llama : integer type consistency in `llama.h`",
        "labels": [
          "enhancement",
          "good first issue",
          "refactoring",
          "roadmap"
        ],
        "created_at": "2023-12-21T19:55:14Z",
        "updated_at": "2026-08-10T16:02:09Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/4574",
        "categories": [
          "models_architectures"
        ]
      },
      {
        "number": 4602,
        "kind": "pr",
        "state": "open",
        "title": "CUDA: fixed peer access toggle synchronization",
        "labels": [],
        "created_at": "2023-12-22T14:33:54Z",
        "updated_at": "2023-12-25T13:23:53Z",
        "closed_at": null,
        "comments": 10,
        "url": "https://github.com/ggml-org/llama.cpp/pull/4602",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 4612,
        "kind": "pr",
        "state": "open",
        "title": "Add flake app to run openai proxy",
        "labels": [
          "nix"
        ],
        "created_at": "2023-12-23T22:53:41Z",
        "updated_at": "2024-01-02T10:41:09Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/4612",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 4613,
        "kind": "pr",
        "state": "open",
        "title": "initial import of hpx support",
        "labels": [],
        "created_at": "2023-12-24T03:16:03Z",
        "updated_at": "2023-12-29T20:24:01Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/4613",
        "categories": [
          "models_architectures",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 4684,
        "kind": "pr",
        "state": "open",
        "title": "DRAFT: Issue#4638 General prompt template handling",
        "labels": [],
        "created_at": "2023-12-29T12:45:22Z",
        "updated_at": "2024-01-12T08:23:17Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/4684",
        "categories": [
          "api_ui_docs"
        ]
      },
      {
        "number": 4694,
        "kind": "pr",
        "state": "open",
        "title": "add missing completion params to chat - Issue#4429",
        "labels": [],
        "created_at": "2023-12-30T03:36:57Z",
        "updated_at": "2023-12-30T03:36:57Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/4694",
        "categories": [
          "other"
        ]
      },
      {
        "number": 4719,
        "kind": "pr",
        "state": "open",
        "title": "CUDA: parallelize graph evaluation across multiple streams",
        "labels": [],
        "created_at": "2023-12-31T17:15:55Z",
        "updated_at": "2024-01-01T16:05:15Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/4719",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 4752,
        "kind": "pr",
        "state": "open",
        "title": "Fix CUDA diag_mask_inf tests with LLAMA_FAST",
        "labels": [],
        "created_at": "2024-01-03T10:36:33Z",
        "updated_at": "2024-01-03T13:38:06Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/4752",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 5040,
        "kind": "issue",
        "state": "open",
        "title": "Include `eos_token` and `bos_token` from `tokenizer_config.json` for chat templating",
        "labels": [
          "enhancement",
          "help wanted"
        ],
        "created_at": "2024-01-19T20:31:05Z",
        "updated_at": "2024-03-20T01:06:12Z",
        "closed_at": null,
        "comments": 12,
        "url": "https://github.com/ggml-org/llama.cpp/issues/5040",
        "categories": [
          "quantization",
          "models_architectures",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 5089,
        "kind": "pr",
        "state": "open",
        "title": "ggllm tensor printf debug function",
        "labels": [],
        "created_at": "2024-01-22T23:24:56Z",
        "updated_at": "2024-02-08T04:38:41Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/5089",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 5143,
        "kind": "pr",
        "state": "open",
        "title": "feat: Introduce new GGUFValueType.OBJ virtual type🌠",
        "labels": [],
        "created_at": "2024-01-26T13:34:30Z",
        "updated_at": "2024-02-06T06:12:12Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/5143",
        "categories": [
          "quantization",
          "models_architectures",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 5215,
        "kind": "issue",
        "state": "open",
        "title": "llama : create llamax library",
        "labels": [
          "refactoring",
          "roadmap"
        ],
        "created_at": "2024-01-30T13:01:06Z",
        "updated_at": "2026-02-18T18:31:24Z",
        "closed_at": null,
        "comments": 22,
        "url": "https://github.com/ggml-org/llama.cpp/issues/5215",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 5276,
        "kind": "issue",
        "state": "open",
        "title": "MiniCPM 2b model support?",
        "labels": [
          "enhancement",
          "good first issue"
        ],
        "created_at": "2024-02-02T08:06:39Z",
        "updated_at": "2024-04-23T08:21:55Z",
        "closed_at": null,
        "comments": 26,
        "url": "https://github.com/ggml-org/llama.cpp/issues/5276",
        "categories": [
          "quantization",
          "models_architectures"
        ]
      },
      {
        "number": 5340,
        "kind": "pr",
        "state": "open",
        "title": "Documentation: Server Example Load Balancing Architecture",
        "labels": [],
        "created_at": "2024-02-05T10:36:30Z",
        "updated_at": "2024-02-05T10:36:30Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/5340",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 5365,
        "kind": "issue",
        "state": "open",
        "title": "Support for Sparse MoE models like Camelidae and Sparsetral",
        "labels": [
          "enhancement",
          "good first issue"
        ],
        "created_at": "2024-02-06T15:00:12Z",
        "updated_at": "2026-06-07T11:02:55Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/issues/5365",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 5385,
        "kind": "pr",
        "state": "open",
        "title": "fix bug make prompt with image always being default",
        "labels": [
          "need feedback"
        ],
        "created_at": "2024-02-07T08:46:33Z",
        "updated_at": "2024-02-14T08:51:06Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/5385",
        "categories": [
          "multimodal_embedding",
          "correctness_testing"
        ]
      },
      {
        "number": 5413,
        "kind": "pr",
        "state": "open",
        "title": "Fuse matrix multiplication + SiLU",
        "labels": [
          "performance",
          "refactoring",
          "Review Complexity : Medium"
        ],
        "created_at": "2024-02-08T13:34:25Z",
        "updated_at": "2024-06-29T13:37:06Z",
        "closed_at": null,
        "comments": 15,
        "url": "https://github.com/ggml-org/llama.cpp/pull/5413",
        "categories": [
          "native_backends",
          "build_platform_ops"
        ]
      },
      {
        "number": 5429,
        "kind": "pr",
        "state": "open",
        "title": "Add CUDA option to use the max release threshold for the default memory pool",
        "labels": [
          "demo"
        ],
        "created_at": "2024-02-09T11:01:47Z",
        "updated_at": "2024-02-13T12:10:44Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/pull/5429",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 5485,
        "kind": "pr",
        "state": "open",
        "title": "fix(server): infinite loop to inference",
        "labels": [],
        "created_at": "2024-02-13T23:44:38Z",
        "updated_at": "2024-02-14T10:49:08Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/5485",
        "categories": [
          "scheduling_serving"
        ]
      },
      {
        "number": 5554,
        "kind": "pr",
        "state": "open",
        "title": "CUDA: switch tile sizes based on binary version",
        "labels": [],
        "created_at": "2024-02-17T20:32:04Z",
        "updated_at": "2024-02-24T17:17:56Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/5554",
        "categories": [
          "native_backends",
          "quantization",
          "build_platform_ops"
        ]
      },
      {
        "number": 5561,
        "kind": "pr",
        "state": "open",
        "title": "Penalty threshold: A mechanism for improving repetition penalties",
        "labels": [
          "enhancement",
          "generation quality",
          "Review Complexity : Medium"
        ],
        "created_at": "2024-02-18T07:41:53Z",
        "updated_at": "2024-05-10T13:37:32Z",
        "closed_at": null,
        "comments": 13,
        "url": "https://github.com/ggml-org/llama.cpp/pull/5561",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 5612,
        "kind": "pr",
        "state": "open",
        "title": "[RFC] common, server : add top-a sampler",
        "labels": [
          "enhancement",
          "generation quality",
          "Review Complexity : High"
        ],
        "created_at": "2024-02-20T15:09:11Z",
        "updated_at": "2024-05-10T14:05:14Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/5612",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing"
        ]
      },
      {
        "number": 5613,
        "kind": "pr",
        "state": "open",
        "title": "Attempt to fix pre-tokenizer",
        "labels": [],
        "created_at": "2024-02-20T16:56:39Z",
        "updated_at": "2024-07-27T22:07:06Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/pull/5613",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "loader_conversion",
          "build_platform_ops"
        ]
      },
      {
        "number": 5675,
        "kind": "pr",
        "state": "open",
        "title": "P-Step Truncation Sampling",
        "labels": [
          "generation quality",
          "refactoring",
          "need feedback",
          "Review Complexity : High"
        ],
        "created_at": "2024-02-23T06:24:58Z",
        "updated_at": "2024-11-30T07:16:05Z",
        "closed_at": null,
        "comments": 12,
        "url": "https://github.com/ggml-org/llama.cpp/pull/5675",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 5679,
        "kind": "pr",
        "state": "open",
        "title": "llama : switch to floating-point token positions",
        "labels": [
          "demo",
          "refactoring",
          "Review Complexity : High"
        ],
        "created_at": "2024-02-23T10:35:39Z",
        "updated_at": "2024-07-22T10:34:03Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/5679",
        "categories": [
          "models_architectures",
          "memory_kv_context"
        ]
      },
      {
        "number": 5719,
        "kind": "pr",
        "state": "open",
        "title": "llama : add llama_kv_cache_compress",
        "labels": [
          "demo"
        ],
        "created_at": "2024-02-25T20:23:20Z",
        "updated_at": "2024-03-16T11:08:38Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/5719",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 5765,
        "kind": "issue",
        "state": "open",
        "title": "server : add \"token healing\" support",
        "labels": [
          "enhancement",
          "good first issue",
          "roadmap"
        ],
        "created_at": "2024-02-28T12:10:30Z",
        "updated_at": "2026-06-08T21:59:11Z",
        "closed_at": null,
        "comments": 9,
        "url": "https://github.com/ggml-org/llama.cpp/issues/5765",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 5823,
        "kind": "issue",
        "state": "open",
        "title": "persimmon crashes with CUDA: assertion failure `ggml_is_contiguous(src0)`",
        "labels": [
          "bug",
          "model"
        ],
        "created_at": "2024-03-01T19:27:09Z",
        "updated_at": "2024-04-16T20:57:28Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/5823",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing"
        ]
      },
      {
        "number": 5837,
        "kind": "pr",
        "state": "open",
        "title": "server: feature Add Admin key parameter for slots/health/metrics",
        "labels": [
          "demo",
          "Review Complexity : Medium",
          "server/api",
          "server"
        ],
        "created_at": "2024-03-02T20:09:58Z",
        "updated_at": "2024-05-10T13:53:42Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/5837",
        "categories": [
          "scheduling_serving",
          "api_ui_docs"
        ]
      },
      {
        "number": 5856,
        "kind": "issue",
        "state": "open",
        "title": "Regressions on IQ3_XXS over time",
        "labels": [
          "bug"
        ],
        "created_at": "2024-03-03T17:11:32Z",
        "updated_at": "2024-09-07T13:32:37Z",
        "closed_at": null,
        "comments": 16,
        "url": "https://github.com/ggml-org/llama.cpp/issues/5856",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 5867,
        "kind": "pr",
        "state": "open",
        "title": "IQ3_S: multiplier based code book",
        "labels": [
          "demo"
        ],
        "created_at": "2024-03-04T07:19:03Z",
        "updated_at": "2024-03-05T15:59:21Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/pull/5867",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "training_adapters"
        ]
      },
      {
        "number": 5891,
        "kind": "pr",
        "state": "open",
        "title": "llama : compute BERT graph with F16 K, V",
        "labels": [
          "demo",
          "Review Complexity : High"
        ],
        "created_at": "2024-03-05T19:25:35Z",
        "updated_at": "2025-09-29T09:57:52Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/pull/5891",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context"
        ]
      },
      {
        "number": 6259,
        "kind": "issue",
        "state": "open",
        "title": "split: allow --split-max-size option",
        "labels": [
          "enhancement",
          "help wanted",
          "good first issue",
          "split"
        ],
        "created_at": "2024-03-23T15:29:25Z",
        "updated_at": "2026-06-03T18:18:37Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/6259",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context"
        ]
      },
      {
        "number": 6268,
        "kind": "issue",
        "state": "open",
        "title": "llama cpp server cant open to public",
        "labels": [
          "enhancement",
          "good first issue",
          "server"
        ],
        "created_at": "2024-03-24T03:04:09Z",
        "updated_at": "2026-07-23T20:20:54Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/issues/6268",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "api_ui_docs"
        ]
      },
      {
        "number": 6287,
        "kind": "pr",
        "state": "open",
        "title": "Fix IQ1_S quantization",
        "labels": [
          "bugfix",
          "Review Complexity : High"
        ],
        "created_at": "2024-03-24T20:11:09Z",
        "updated_at": "2024-05-10T13:42:08Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/6287",
        "categories": [
          "quantization",
          "models_architectures",
          "correctness_testing"
        ]
      },
      {
        "number": 6289,
        "kind": "pr",
        "state": "open",
        "title": "Control vectors in server",
        "labels": [
          "demo",
          "Review Complexity : Medium"
        ],
        "created_at": "2024-03-24T21:53:12Z",
        "updated_at": "2024-07-01T07:38:42Z",
        "closed_at": null,
        "comments": 20,
        "url": "https://github.com/ggml-org/llama.cpp/pull/6289",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 6311,
        "kind": "issue",
        "state": "open",
        "title": "Add a new `llama_load_model_from_buffer()` method to compliment `llama_load_model_from_file()`",
        "labels": [
          "enhancement",
          "roadmap"
        ],
        "created_at": "2024-03-26T02:03:02Z",
        "updated_at": "2025-04-23T12:44:29Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/issues/6311",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 6312,
        "kind": "pr",
        "state": "open",
        "title": "server: fix system_tokens being erased in kv_cache;",
        "labels": [
          "bugfix",
          "Review Complexity : Medium"
        ],
        "created_at": "2024-03-26T02:47:16Z",
        "updated_at": "2024-05-10T05:14:13Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/6312",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 6358,
        "kind": "pr",
        "state": "open",
        "title": "Allow continuous batching to be disabled",
        "labels": [
          "enhancement",
          "Review Complexity : Medium"
        ],
        "created_at": "2024-03-28T02:58:43Z",
        "updated_at": "2024-05-12T00:59:24Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/6358",
        "categories": [
          "scheduling_serving"
        ]
      },
      {
        "number": 6400,
        "kind": "issue",
        "state": "open",
        "title": "Kompute-based Vulkan backend shows an GGML_OP_GET_ROWS error",
        "labels": [
          "bug"
        ],
        "created_at": "2024-03-30T21:56:07Z",
        "updated_at": "2024-06-03T04:00:21Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/6400",
        "categories": [
          "native_backends",
          "correctness_testing"
        ]
      },
      {
        "number": 6440,
        "kind": "pr",
        "state": "open",
        "title": "Xeon Phi (Knights Corner) Support.",
        "labels": [
          "enhancement",
          "performance",
          "Review Complexity : High",
          "ggml"
        ],
        "created_at": "2024-04-02T17:10:37Z",
        "updated_at": "2025-12-10T08:16:44Z",
        "closed_at": null,
        "comments": 9,
        "url": "https://github.com/ggml-org/llama.cpp/pull/6440",
        "categories": [
          "models_architectures",
          "build_platform_ops"
        ]
      },
      {
        "number": 6445,
        "kind": "pr",
        "state": "open",
        "title": "Smooth Sampling / Quadratic Sampling support",
        "labels": [
          "performance",
          "generation quality",
          "Review Complexity : High"
        ],
        "created_at": "2024-04-02T22:25:09Z",
        "updated_at": "2026-04-28T18:51:02Z",
        "closed_at": null,
        "comments": 13,
        "url": "https://github.com/ggml-org/llama.cpp/pull/6445",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 6453,
        "kind": "pr",
        "state": "open",
        "title": "Adding Support for Custom Qwen2moe Architectures with mergekit-qwen2",
        "labels": [
          "model",
          "Review Complexity : High"
        ],
        "created_at": "2024-04-03T07:25:44Z",
        "updated_at": "2024-05-10T05:02:15Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/6453",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 6502,
        "kind": "pr",
        "state": "open",
        "title": "Implement automatic NGL detection",
        "labels": [
          "enhancement",
          "need feedback",
          "Review Complexity : Medium"
        ],
        "created_at": "2024-04-05T10:09:56Z",
        "updated_at": "2024-05-30T22:11:01Z",
        "closed_at": null,
        "comments": 13,
        "url": "https://github.com/ggml-org/llama.cpp/pull/6502",
        "categories": [
          "build_platform_ops"
        ]
      },
      {
        "number": 6506,
        "kind": "issue",
        "state": "open",
        "title": "Question: How to generate an MPS gputrace",
        "labels": [
          "help wanted",
          "high priority"
        ],
        "created_at": "2024-04-05T14:08:32Z",
        "updated_at": "2025-10-16T10:08:01Z",
        "closed_at": null,
        "comments": 11,
        "url": "https://github.com/ggml-org/llama.cpp/issues/6506",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 6546,
        "kind": "issue",
        "state": "open",
        "title": "kubernetes example",
        "labels": [
          "enhancement",
          "help wanted",
          "kubernetes"
        ],
        "created_at": "2024-04-08T16:31:37Z",
        "updated_at": "2026-03-17T10:35:47Z",
        "closed_at": null,
        "comments": 25,
        "url": "https://github.com/ggml-org/llama.cpp/issues/6546",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "api_ui_docs"
        ]
      },
      {
        "number": 6685,
        "kind": "issue",
        "state": "open",
        "title": "parallel/server crashes with: ggml.c:16521: i != GGML_HASHTABLE_FULL when defragmentation is enabled",
        "labels": [
          "bug"
        ],
        "created_at": "2024-04-15T11:39:29Z",
        "updated_at": "2026-03-17T10:36:27Z",
        "closed_at": null,
        "comments": 28,
        "url": "https://github.com/ggml-org/llama.cpp/issues/6685",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 6717,
        "kind": "issue",
        "state": "open",
        "title": "Inference not running when using a tokenizer with word model",
        "labels": [
          "bug",
          "help wanted"
        ],
        "created_at": "2024-04-17T06:51:31Z",
        "updated_at": "2024-05-27T07:26:22Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/issues/6717",
        "categories": [
          "quantization",
          "models_architectures",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 6739,
        "kind": "pr",
        "state": "open",
        "title": "Update server.cpp example with correct startup sequence",
        "labels": [
          "examples",
          "Review Complexity : Medium"
        ],
        "created_at": "2024-04-18T10:09:17Z",
        "updated_at": "2025-12-05T17:33:25Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/pull/6739",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "loader_conversion",
          "api_ui_docs"
        ]
      },
      {
        "number": 6758,
        "kind": "issue",
        "state": "open",
        "title": "ggml : add GPU support for Mamba models",
        "labels": [
          "enhancement",
          "help wanted",
          "Nvidia GPU",
          "roadmap"
        ],
        "created_at": "2024-04-19T06:47:35Z",
        "updated_at": "2026-05-04T10:37:38Z",
        "closed_at": null,
        "comments": 39,
        "url": "https://github.com/ggml-org/llama.cpp/issues/6758",
        "categories": [
          "native_backends",
          "models_architectures",
          "build_platform_ops"
        ]
      },
      {
        "number": 6784,
        "kind": "pr",
        "state": "open",
        "title": "Added dependency needed for numa in numactl mode",
        "labels": [
          "enhancement",
          "Review Complexity : Low"
        ],
        "created_at": "2024-04-20T13:46:46Z",
        "updated_at": "2024-05-09T16:06:18Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/6784",
        "categories": [
          "other"
        ]
      },
      {
        "number": 6828,
        "kind": "pr",
        "state": "open",
        "title": "Server: enable lookup decoding",
        "labels": [
          "enhancement",
          "examples",
          "Review Complexity : Medium"
        ],
        "created_at": "2024-04-22T12:54:59Z",
        "updated_at": "2025-03-01T09:31:45Z",
        "closed_at": null,
        "comments": 38,
        "url": "https://github.com/ggml-org/llama.cpp/pull/6828",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 6831,
        "kind": "pr",
        "state": "open",
        "title": "common : fix parallel shard download interleaving output",
        "labels": [
          "bugfix",
          "Review Complexity : Low"
        ],
        "created_at": "2024-04-22T17:35:20Z",
        "updated_at": "2024-07-22T16:40:21Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/6831",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing"
        ]
      },
      {
        "number": 6834,
        "kind": "pr",
        "state": "open",
        "title": "Generic Chat templating code with text/json file based config; main chat updated to drive its in-prefix, in-suffix and reverse-prompt from same; chat-apply-template equivalent c-api to allow use by other codes also",
        "labels": [
          "enhancement",
          "Review Complexity : Medium"
        ],
        "created_at": "2024-04-22T20:46:33Z",
        "updated_at": "2024-07-03T11:57:18Z",
        "closed_at": null,
        "comments": 34,
        "url": "https://github.com/ggml-org/llama.cpp/pull/6834",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 6840,
        "kind": "pr",
        "state": "open",
        "title": "llamafile : improve moe prompt eval speed on cpu",
        "labels": [
          "enhancement",
          "Review Complexity : Medium",
          "ggml"
        ],
        "created_at": "2024-04-23T05:41:24Z",
        "updated_at": "2026-04-23T18:12:01Z",
        "closed_at": null,
        "comments": 21,
        "url": "https://github.com/ggml-org/llama.cpp/pull/6840",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "api_ui_docs"
        ]
      },
      {
        "number": 6888,
        "kind": "pr",
        "state": "open",
        "title": "Clamp out of range values in K quantizer",
        "labels": [
          "model",
          "bugfix",
          "Review Complexity : Medium"
        ],
        "created_at": "2024-04-25T00:00:13Z",
        "updated_at": "2024-05-18T05:21:51Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/6888",
        "categories": [
          "quantization",
          "models_architectures",
          "loader_conversion",
          "correctness_testing"
        ]
      },
      {
        "number": 6913,
        "kind": "issue",
        "state": "open",
        "title": "ggml : unified CMake build",
        "labels": [
          "enhancement",
          "build",
          "refactoring",
          "roadmap"
        ],
        "created_at": "2024-04-25T19:15:40Z",
        "updated_at": "2026-06-12T06:05:59Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/issues/6913",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "build_platform_ops"
        ]
      },
      {
        "number": 6919,
        "kind": "pr",
        "state": "open",
        "title": "support MiniCPM-V-2",
        "labels": [
          "enhancement",
          "demo",
          "examples",
          "python",
          "Review Complexity : High"
        ],
        "created_at": "2024-04-26T08:44:38Z",
        "updated_at": "2026-06-29T10:07:38Z",
        "closed_at": null,
        "comments": 43,
        "url": "https://github.com/ggml-org/llama.cpp/pull/6919",
        "categories": [
          "models_architectures",
          "multimodal_embedding",
          "api_ui_docs"
        ]
      },
      {
        "number": 6934,
        "kind": "pr",
        "state": "open",
        "title": "Fix clip build on windows + clang",
        "labels": [
          "bugfix",
          "Review Complexity : Low"
        ],
        "created_at": "2024-04-26T17:01:14Z",
        "updated_at": "2024-05-09T15:32:19Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/6934",
        "categories": [
          "models_architectures",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 6941,
        "kind": "pr",
        "state": "open",
        "title": "Updated server_queue to delete tasks from queue when server is shutdown. Feature Request #6421",
        "labels": [
          "demo",
          "Review Complexity : Low"
        ],
        "created_at": "2024-04-27T02:43:23Z",
        "updated_at": "2024-05-09T14:28:36Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/6941",
        "categories": [
          "scheduling_serving"
        ]
      },
      {
        "number": 6988,
        "kind": "pr",
        "state": "open",
        "title": "new tokenizer-verifier tool to check gguf tokenizer parameters",
        "labels": [
          "enhancement",
          "Review Complexity : Low"
        ],
        "created_at": "2024-04-29T20:08:16Z",
        "updated_at": "2024-05-14T12:15:56Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/6988",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "build_platform_ops"
        ]
      },
      {
        "number": 7058,
        "kind": "pr",
        "state": "open",
        "title": "Script to convert Grok-1 weights from raw JAX pickle files.",
        "labels": [
          "enhancement",
          "python",
          "Review Complexity : Medium"
        ],
        "created_at": "2024-05-03T11:36:27Z",
        "updated_at": "2024-05-27T09:23:42Z",
        "closed_at": null,
        "comments": 19,
        "url": "https://github.com/ggml-org/llama.cpp/pull/7058",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 7107,
        "kind": "pr",
        "state": "open",
        "title": "tokenization: no double BOS tokens",
        "labels": [
          "refactoring",
          "Review Complexity : Medium"
        ],
        "created_at": "2024-05-06T20:07:59Z",
        "updated_at": "2024-05-16T20:52:08Z",
        "closed_at": null,
        "comments": 16,
        "url": "https://github.com/ggml-org/llama.cpp/pull/7107",
        "categories": [
          "models_architectures",
          "loader_conversion"
        ]
      },
      {
        "number": 7187,
        "kind": "pr",
        "state": "open",
        "title": "Add token healing to `main` and `server`",
        "labels": [
          "enhancement",
          "help wanted",
          "need feedback",
          "examples",
          "Review Complexity : High",
          "server"
        ],
        "created_at": "2024-05-09T21:36:08Z",
        "updated_at": "2024-08-10T18:38:02Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/7187",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 7219,
        "kind": "issue",
        "state": "open",
        "title": "convert-hf-to-gguf.py breaks on phi-2",
        "labels": [
          "bug"
        ],
        "created_at": "2024-05-11T13:41:51Z",
        "updated_at": "2024-05-17T07:49:21Z",
        "closed_at": null,
        "comments": 11,
        "url": "https://github.com/ggml-org/llama.cpp/issues/7219",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "loader_conversion",
          "correctness_testing"
        ]
      },
      {
        "number": 7239,
        "kind": "pr",
        "state": "open",
        "title": "Fix self extend on the server.",
        "labels": [
          "help wanted",
          "examples",
          "Review Complexity : Low",
          "server"
        ],
        "created_at": "2024-05-12T10:16:59Z",
        "updated_at": "2024-05-15T01:51:47Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/7239",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 7269,
        "kind": "pr",
        "state": "open",
        "title": "ggml, ngram-cache, log: added const and const ref for function params",
        "labels": [
          "refactoring",
          "Review Complexity : Medium"
        ],
        "created_at": "2024-05-14T00:29:35Z",
        "updated_at": "2024-05-14T00:58:51Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/7269",
        "categories": [
          "memory_kv_context",
          "speculative_generation"
        ]
      },
      {
        "number": 7270,
        "kind": "pr",
        "state": "open",
        "title": "common, ngram_cache: added const reference for std::pair<> and std::tuple<> more 16 bytes:",
        "labels": [
          "refactoring",
          "Review Complexity : Low"
        ],
        "created_at": "2024-05-14T01:08:54Z",
        "updated_at": "2024-08-25T21:23:11Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/7270",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "speculative_generation"
        ]
      },
      {
        "number": 7315,
        "kind": "pr",
        "state": "open",
        "title": "sched : support async weight copy",
        "labels": [
          "performance",
          "Review Complexity : Medium"
        ],
        "created_at": "2024-05-15T22:54:23Z",
        "updated_at": "2024-05-17T19:36:06Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/7315",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing"
        ]
      },
      {
        "number": 7373,
        "kind": "pr",
        "state": "open",
        "title": "Add minimal python client example for the server, streaming callback",
        "labels": [
          "examples",
          "python",
          "Review Complexity : Medium",
          "server"
        ],
        "created_at": "2024-05-18T20:52:20Z",
        "updated_at": "2024-05-20T21:35:21Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/7373",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 7420,
        "kind": "pr",
        "state": "open",
        "title": "Direct I/O and Transparent HugePages",
        "labels": [
          "script",
          "demo",
          "examples",
          "python",
          "Review Complexity : Medium",
          "server"
        ],
        "created_at": "2024-05-20T20:33:18Z",
        "updated_at": "2024-09-08T17:07:43Z",
        "closed_at": null,
        "comments": 25,
        "url": "https://github.com/ggml-org/llama.cpp/pull/7420",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 7448,
        "kind": "pr",
        "state": "open",
        "title": "Check for llama_get_logits_ith() errors",
        "labels": [
          "android",
          "examples",
          "Review Complexity : Medium",
          "server"
        ],
        "created_at": "2024-05-21T21:20:16Z",
        "updated_at": "2024-05-22T22:53:16Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/7448",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 7455,
        "kind": "pr",
        "state": "open",
        "title": "Introduce ggml_syncthreads()",
        "labels": [
          "performance",
          "Review Complexity : Medium"
        ],
        "created_at": "2024-05-22T07:46:59Z",
        "updated_at": "2024-05-26T01:39:24Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/7455",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 7497,
        "kind": "pr",
        "state": "open",
        "title": "Introduce Q8_0 and Q4_0 with Bf16 delta values",
        "labels": [
          "examples",
          "python",
          "Review Complexity : High",
          "ggml",
          "Tensor Encoding Scheme"
        ],
        "created_at": "2024-05-23T16:50:15Z",
        "updated_at": "2024-08-22T07:10:57Z",
        "closed_at": null,
        "comments": 13,
        "url": "https://github.com/ggml-org/llama.cpp/pull/7497",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 7522,
        "kind": "pr",
        "state": "open",
        "title": "Rebalancing Metal threads workload in dot product kernel kernel_mul_mv_f16_f32_l4",
        "labels": [
          "Review Complexity : Medium",
          "Apple Metal"
        ],
        "created_at": "2024-05-24T19:41:35Z",
        "updated_at": "2025-03-08T19:39:23Z",
        "closed_at": null,
        "comments": 14,
        "url": "https://github.com/ggml-org/llama.cpp/pull/7522",
        "categories": [
          "native_backends",
          "quantization",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 7535,
        "kind": "pr",
        "state": "open",
        "title": "Portability: use the ccache path detected in cmake when setting the compiler launch rule",
        "labels": [
          "build",
          "Review Complexity : Low"
        ],
        "created_at": "2024-05-25T14:29:11Z",
        "updated_at": "2024-07-09T18:13:21Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/7535",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 7536,
        "kind": "issue",
        "state": "open",
        "title": "Is nix-publish-flake github actions broken?",
        "labels": [
          "bug",
          "bug-unconfirmed",
          "devops"
        ],
        "created_at": "2024-05-25T14:48:32Z",
        "updated_at": "2024-05-25T14:54:25Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/7536",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 7553,
        "kind": "pr",
        "state": "open",
        "title": "Add PaliGemma Support",
        "labels": [
          "examples",
          "Review Complexity : Low",
          "ggml"
        ],
        "created_at": "2024-05-27T02:49:36Z",
        "updated_at": "2025-06-01T21:06:55Z",
        "closed_at": null,
        "comments": 24,
        "url": "https://github.com/ggml-org/llama.cpp/pull/7553",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "multimodal_embedding",
          "loader_conversion",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 7555,
        "kind": "pr",
        "state": "open",
        "title": "fix performance regression on woa",
        "labels": [
          "build",
          "Review Complexity : Low"
        ],
        "created_at": "2024-05-27T04:48:15Z",
        "updated_at": "2024-05-27T11:08:29Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/7555",
        "categories": [
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 7559,
        "kind": "issue",
        "state": "open",
        "title": "Refactor: Existing examples refactoring opportunities",
        "labels": [
          "help wanted",
          "refactoring"
        ],
        "created_at": "2024-05-27T09:31:05Z",
        "updated_at": "2024-05-27T09:39:09Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/7559",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "structured_generation_tools",
          "api_ui_docs"
        ]
      },
      {
        "number": 7575,
        "kind": "issue",
        "state": "open",
        "title": "Bug: having more than one context doesn't work as expected with the Vulkan backend",
        "labels": [
          "bug",
          "medium severity"
        ],
        "created_at": "2024-05-27T22:24:43Z",
        "updated_at": "2024-08-10T01:07:04Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/7575",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 7576,
        "kind": "pr",
        "state": "open",
        "title": "ggml-threading.cpp",
        "labels": [
          "build",
          "Review Complexity : Medium",
          "ggml"
        ],
        "created_at": "2024-05-27T23:50:32Z",
        "updated_at": "2024-05-28T22:19:31Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/7576",
        "categories": [
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 7581,
        "kind": "pr",
        "state": "open",
        "title": "batched : make n_threads and n_threads_batch configurable in batched & batched-bench",
        "labels": [
          "examples",
          "Review Complexity : Medium"
        ],
        "created_at": "2024-05-28T02:53:35Z",
        "updated_at": "2024-05-29T06:45:17Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/7581",
        "categories": [
          "scheduling_serving",
          "api_ui_docs"
        ]
      },
      {
        "number": 7607,
        "kind": "pr",
        "state": "open",
        "title": "Readme: add HyperMink/inferenceable to HTTP server",
        "labels": [
          "Review Complexity : Low"
        ],
        "created_at": "2024-05-29T06:06:19Z",
        "updated_at": "2024-05-30T06:44:22Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/7607",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "api_ui_docs"
        ]
      },
      {
        "number": 7639,
        "kind": "issue",
        "state": "open",
        "title": "Bug: server crashes on startup is ckt ctv specified.",
        "labels": [
          "bug",
          "high severity"
        ],
        "created_at": "2024-05-30T13:03:41Z",
        "updated_at": "2024-05-30T14:57:27Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/7639",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 7651,
        "kind": "pr",
        "state": "open",
        "title": "Merging #7568 with #7430(Implementing LLaMA 3 torch to gguf conversion)",
        "labels": [
          "examples",
          "Review Complexity : Low",
          "python"
        ],
        "created_at": "2024-05-30T18:05:32Z",
        "updated_at": "2024-09-20T11:17:51Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/7651",
        "categories": [
          "quantization",
          "models_architectures",
          "multimodal_embedding",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 7705,
        "kind": "pr",
        "state": "open",
        "title": "PHI3-vision gguf conversion",
        "labels": [
          "examples",
          "Review Complexity : Low",
          "python",
          "ggml"
        ],
        "created_at": "2024-06-03T01:22:44Z",
        "updated_at": "2025-12-02T12:08:03Z",
        "closed_at": null,
        "comments": 11,
        "url": "https://github.com/ggml-org/llama.cpp/pull/7705",
        "categories": [
          "quantization",
          "models_architectures",
          "multimodal_embedding",
          "loader_conversion",
          "api_ui_docs"
        ]
      },
      {
        "number": 7720,
        "kind": "issue",
        "state": "open",
        "title": "Bug: Grammar readme seems incorrect",
        "labels": [
          "bug",
          "documentation",
          "low severity"
        ],
        "created_at": "2024-06-03T20:48:22Z",
        "updated_at": "2026-08-06T03:10:16Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/7720",
        "categories": [
          "models_architectures",
          "structured_generation_tools",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 7747,
        "kind": "pr",
        "state": "open",
        "title": "[ci] add LLAMA_CURL flags to the prebuilt binaries",
        "labels": [
          "Review Complexity : Low",
          "devops"
        ],
        "created_at": "2024-06-04T16:59:14Z",
        "updated_at": "2025-09-19T11:15:25Z",
        "closed_at": null,
        "comments": 9,
        "url": "https://github.com/ggml-org/llama.cpp/pull/7747",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 7772,
        "kind": "issue",
        "state": "open",
        "title": "ggml : add DirectML backend",
        "labels": [
          "help wanted",
          "research 🔬",
          "roadmap"
        ],
        "created_at": "2024-06-05T14:21:34Z",
        "updated_at": "2026-02-25T09:20:33Z",
        "closed_at": null,
        "comments": 20,
        "url": "https://github.com/ggml-org/llama.cpp/issues/7772",
        "categories": [
          "native_backends",
          "build_platform_ops"
        ]
      },
      {
        "number": 7795,
        "kind": "pr",
        "state": "open",
        "title": "feat: add changes to handle jina v2 chinese code",
        "labels": [
          "python",
          "Review Complexity : Medium"
        ],
        "created_at": "2024-06-06T08:19:02Z",
        "updated_at": "2024-09-30T02:03:28Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/7795",
        "categories": [
          "models_architectures",
          "api_ui_docs"
        ]
      },
      {
        "number": 7796,
        "kind": "pr",
        "state": "open",
        "title": "WIP: Use DirectStorage with CUDA interop to more efficient load tensors",
        "labels": [
          "build",
          "Nvidia GPU",
          "Review Complexity : Medium",
          "ggml"
        ],
        "created_at": "2024-06-06T08:27:49Z",
        "updated_at": "2024-06-11T00:37:10Z",
        "closed_at": null,
        "comments": 18,
        "url": "https://github.com/ggml-org/llama.cpp/pull/7796",
        "categories": [
          "native_backends",
          "models_architectures",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 7810,
        "kind": "issue",
        "state": "open",
        "title": "SIGSEGV on moderately complex grammar ",
        "labels": [
          "bug"
        ],
        "created_at": "2024-06-07T00:00:53Z",
        "updated_at": "2024-07-25T02:44:44Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/7810",
        "categories": [
          "quantization",
          "models_architectures",
          "structured_generation_tools",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 7884,
        "kind": "pr",
        "state": "open",
        "title": "Implemented Spellcheck for Llama.cpp",
        "labels": [
          "demo",
          "examples",
          "Review Complexity : Low"
        ],
        "created_at": "2024-06-11T17:16:14Z",
        "updated_at": "2024-10-07T17:59:53Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/7884",
        "categories": [
          "models_architectures",
          "api_ui_docs"
        ]
      },
      {
        "number": 7915,
        "kind": "pr",
        "state": "open",
        "title": "rpc : enable async operations",
        "labels": [
          "Review Complexity : Low"
        ],
        "created_at": "2024-06-13T08:00:39Z",
        "updated_at": "2024-06-20T10:50:02Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/7915",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving"
        ]
      },
      {
        "number": 7926,
        "kind": "pr",
        "state": "open",
        "title": "Fix #7924: Error while converting BERT to GGUF: Can not map tensor 'bert.embeddings.LayerNorm.beta'",
        "labels": [
          "bugfix",
          "Review Complexity : Low",
          "python"
        ],
        "created_at": "2024-06-13T21:08:53Z",
        "updated_at": "2024-08-05T12:51:37Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/7926",
        "categories": [
          "quantization",
          "models_architectures",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 8032,
        "kind": "pr",
        "state": "open",
        "title": "rpc : copy tensors across servers",
        "labels": [
          "Review Complexity : Medium"
        ],
        "created_at": "2024-06-20T10:37:34Z",
        "updated_at": "2025-04-23T21:19:33Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/pull/8032",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 8050,
        "kind": "issue",
        "state": "open",
        "title": "Bug: ABI problem in binary file \"llama-b3187-bin-win-msvc-arm64.zip\"",
        "labels": [
          "bug",
          "medium severity"
        ],
        "created_at": "2024-06-21T07:11:57Z",
        "updated_at": "2024-07-02T21:27:00Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/8050",
        "categories": [
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 8079,
        "kind": "pr",
        "state": "open",
        "title": "Resolving the problem in chktxt",
        "labels": [
          "Review Complexity : Low",
          "python"
        ],
        "created_at": "2024-06-23T15:51:28Z",
        "updated_at": "2024-06-24T09:11:30Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/8079",
        "categories": [
          "quantization",
          "models_architectures",
          "loader_conversion",
          "api_ui_docs"
        ]
      },
      {
        "number": 8109,
        "kind": "issue",
        "state": "open",
        "title": "Bug: abort on Android (pixel 8 pro)",
        "labels": [
          "bug",
          "android",
          "high severity"
        ],
        "created_at": "2024-06-25T11:28:31Z",
        "updated_at": "2024-07-08T07:04:12Z",
        "closed_at": null,
        "comments": 9,
        "url": "https://github.com/ggml-org/llama.cpp/issues/8109",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 8121,
        "kind": "pr",
        "state": "open",
        "title": "Embed files",
        "labels": [
          "Review Complexity : Low",
          "python"
        ],
        "created_at": "2024-06-25T20:16:30Z",
        "updated_at": "2024-08-10T11:17:18Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/8121",
        "categories": [
          "quantization",
          "models_architectures",
          "api_ui_docs"
        ]
      },
      {
        "number": 8133,
        "kind": "pr",
        "state": "open",
        "title": "`json`: unified properties order across optional & required",
        "labels": [
          "testing",
          "examples",
          "python",
          "Review Complexity : Medium",
          "server"
        ],
        "created_at": "2024-06-26T08:46:26Z",
        "updated_at": "2025-05-26T16:30:25Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/8133",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 8136,
        "kind": "pr",
        "state": "open",
        "title": "Performance Tuning for Q4_K matmul CUDA kernel",
        "labels": [
          "Nvidia GPU",
          "Review Complexity : Low",
          "ggml"
        ],
        "created_at": "2024-06-26T13:38:30Z",
        "updated_at": "2024-06-30T10:14:37Z",
        "closed_at": null,
        "comments": 14,
        "url": "https://github.com/ggml-org/llama.cpp/pull/8136",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing"
        ]
      },
      {
        "number": 8188,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Installable package via winget",
        "labels": [
          "enhancement",
          "help wanted"
        ],
        "created_at": "2024-06-28T13:27:20Z",
        "updated_at": "2026-08-21T08:35:26Z",
        "closed_at": null,
        "comments": 23,
        "url": "https://github.com/ggml-org/llama.cpp/issues/8188",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 8199,
        "kind": "pr",
        "state": "open",
        "title": "`json`: $ref + object overhaul (https & recursive $refs, mix properties & allOf)",
        "labels": [
          "testing",
          "breaking change",
          "examples",
          "python",
          "Review Complexity : Medium",
          "server"
        ],
        "created_at": "2024-06-28T21:31:33Z",
        "updated_at": "2024-08-28T19:27:51Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/8199",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 8208,
        "kind": "pr",
        "state": "open",
        "title": "server: Windows 7 compatibility",
        "labels": [
          "build",
          "examples",
          "Review Complexity : Low",
          "server"
        ],
        "created_at": "2024-06-29T14:04:28Z",
        "updated_at": "2025-03-20T11:22:16Z",
        "closed_at": null,
        "comments": 10,
        "url": "https://github.com/ggml-org/llama.cpp/pull/8208",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 8210,
        "kind": "pr",
        "state": "open",
        "title": "clip: don't throw exceptions from llava functions compiled as extern \"C\"",
        "labels": [
          "examples",
          "Review Complexity : Low"
        ],
        "created_at": "2024-06-29T15:05:30Z",
        "updated_at": "2024-07-12T13:34:21Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/8210",
        "categories": [
          "models_architectures",
          "multimodal_embedding",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 8249,
        "kind": "pr",
        "state": "open",
        "title": "Fixed the model conversion bug caused by minicpm's GQA structure。After testing minicpm's GQA, the converted model generates all <h>. This is because the number of k and v matrices of Gqa should be the same as kv_head, not the same as head/kv_head.",
        "labels": [
          "Review Complexity : Low",
          "python"
        ],
        "created_at": "2024-07-02T07:45:01Z",
        "updated_at": "2024-07-05T04:15:30Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/pull/8249",
        "categories": [
          "models_architectures",
          "loader_conversion",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 8252,
        "kind": "pr",
        "state": "open",
        "title": "Fix false flags used in termux demo in README.md",
        "labels": [
          "bugfix",
          "Review Complexity : Low"
        ],
        "created_at": "2024-07-02T08:58:30Z",
        "updated_at": "2024-08-05T12:51:25Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/8252",
        "categories": [
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 8339,
        "kind": "pr",
        "state": "open",
        "title": "build example/main.cpp as shared library and intercept token printing using FFI",
        "labels": [
          "demo",
          "examples",
          "Review Complexity : Medium"
        ],
        "created_at": "2024-07-06T17:49:19Z",
        "updated_at": "2024-07-13T14:42:26Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/8339",
        "categories": [
          "models_architectures",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 8362,
        "kind": "pr",
        "state": "open",
        "title": "server: Update public_simplechat/datautils.mjs",
        "labels": [
          "examples",
          "Review Complexity : Low",
          "server"
        ],
        "created_at": "2024-07-08T05:33:47Z",
        "updated_at": "2024-07-13T14:41:18Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/8362",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "api_ui_docs"
        ]
      },
      {
        "number": 8379,
        "kind": "pr",
        "state": "open",
        "title": "Tokenizer fixes",
        "labels": [
          "testing",
          "Review Complexity : Low",
          "python"
        ],
        "created_at": "2024-07-08T23:34:38Z",
        "updated_at": "2024-07-19T15:24:38Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/8379",
        "categories": [
          "quantization",
          "models_architectures",
          "loader_conversion",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 8435,
        "kind": "pr",
        "state": "open",
        "title": "convert_hf_to_gguf.py, convert_hf_to_gguf_update.py: Added Ukrainian tokens into string",
        "labels": [
          "Review Complexity : Low",
          "python"
        ],
        "created_at": "2024-07-11T10:13:05Z",
        "updated_at": "2024-07-13T14:40:02Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/8435",
        "categories": [
          "quantization",
          "models_architectures",
          "loader_conversion",
          "api_ui_docs"
        ]
      },
      {
        "number": 8540,
        "kind": "pr",
        "state": "open",
        "title": "README.md:  include steps to run cmake [no ci]",
        "labels": [
          "examples",
          "Review Complexity : Low"
        ],
        "created_at": "2024-07-17T13:31:53Z",
        "updated_at": "2024-07-19T13:58:03Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/8540",
        "categories": [
          "models_architectures",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 8606,
        "kind": "pr",
        "state": "open",
        "title": "llama : tokenizer unicode codepoint categories",
        "labels": [
          "script",
          "testing",
          "python",
          "Review Complexity : Medium"
        ],
        "created_at": "2024-07-20T21:52:33Z",
        "updated_at": "2024-08-13T19:19:28Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/8606",
        "categories": [
          "models_architectures",
          "loader_conversion",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 8633,
        "kind": "pr",
        "state": "open",
        "title": "Python - Pre-compiled CFFI module for CPU and CUDA",
        "labels": [
          "Review Complexity : Low"
        ],
        "created_at": "2024-07-22T20:57:12Z",
        "updated_at": "2024-08-01T12:09:00Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/8633",
        "categories": [
          "native_backends",
          "models_architectures",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 8634,
        "kind": "issue",
        "state": "open",
        "title": "Bug: non-chat completions not respecting the max_tokens parameter using the OpenAI api",
        "labels": [
          "bug",
          "bug-unconfirmed",
          "high severity"
        ],
        "created_at": "2024-07-22T22:29:09Z",
        "updated_at": "2024-07-27T10:40:24Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/8634",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 8637,
        "kind": "pr",
        "state": "open",
        "title": "llava-cli: format batch --image descriptions according to --template",
        "labels": [
          "examples",
          "Review Complexity : Low"
        ],
        "created_at": "2024-07-22T23:42:40Z",
        "updated_at": "2024-08-30T22:36:28Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/8637",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 8648,
        "kind": "pr",
        "state": "open",
        "title": "lookup: Use tree of sequences instead of single sequence",
        "labels": [
          "examples",
          "Review Complexity : Medium"
        ],
        "created_at": "2024-07-23T15:06:01Z",
        "updated_at": "2024-07-23T15:44:31Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/8648",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 8726,
        "kind": "pr",
        "state": "open",
        "title": "Add support for loongarch backend in sgemm.cpp",
        "labels": [
          "Review Complexity : High"
        ],
        "created_at": "2024-07-27T16:27:16Z",
        "updated_at": "2024-08-01T12:10:38Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/8726",
        "categories": [
          "native_backends",
          "models_architectures",
          "multimodal_embedding",
          "correctness_testing"
        ]
      },
      {
        "number": 8742,
        "kind": "pr",
        "state": "open",
        "title": "Swift: Fix Windows build",
        "labels": [
          "bugfix",
          "Review Complexity : Low"
        ],
        "created_at": "2024-07-29T01:25:01Z",
        "updated_at": "2024-08-05T12:51:02Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/8742",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 8795,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Support Zyphra/Zamba2-2.7B",
        "labels": [
          "enhancement",
          "model"
        ],
        "created_at": "2024-07-31T12:58:18Z",
        "updated_at": "2026-08-02T12:06:39Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/issues/8795",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 8836,
        "kind": "pr",
        "state": "open",
        "title": "Changes for the existing quant strategies / FTYPEs and new ones",
        "labels": [
          "examples",
          "Review Complexity : Low",
          "python"
        ],
        "created_at": "2024-08-02T18:40:55Z",
        "updated_at": "2025-03-08T00:14:43Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/8836",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 8837,
        "kind": "pr",
        "state": "open",
        "title": "vulkan: fix storageBuffer16BitAccess detection on some adreno driver",
        "labels": [
          "Vulkan",
          "bugfix",
          "Review Complexity : Low",
          "ggml"
        ],
        "created_at": "2024-08-02T18:53:02Z",
        "updated_at": "2024-08-14T00:37:34Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/pull/8837",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing"
        ]
      },
      {
        "number": 8878,
        "kind": "pr",
        "state": "open",
        "title": "ggml : make GeLU faster and more accurate on CPU",
        "labels": [
          "Review Complexity : High",
          "ggml"
        ],
        "created_at": "2024-08-05T16:39:20Z",
        "updated_at": "2024-08-18T13:37:52Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/pull/8878",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 8917,
        "kind": "pr",
        "state": "open",
        "title": "Quantize: specify each major tensor quant in CLI for common LLMs",
        "labels": [
          "demo",
          "examples",
          "Review Complexity : Medium"
        ],
        "created_at": "2024-08-07T20:12:06Z",
        "updated_at": "2024-10-10T00:34:16Z",
        "closed_at": null,
        "comments": 13,
        "url": "https://github.com/ggml-org/llama.cpp/pull/8917",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 8924,
        "kind": "pr",
        "state": "open",
        "title": "Added perplexity metrics for llama 3.1 with different quantization se…",
        "labels": [
          "documentation",
          "examples",
          "Review Complexity : Low"
        ],
        "created_at": "2024-08-08T08:56:40Z",
        "updated_at": "2024-08-09T12:47:12Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/8924",
        "categories": [
          "quantization",
          "models_architectures",
          "api_ui_docs"
        ]
      },
      {
        "number": 8962,
        "kind": "pr",
        "state": "open",
        "title": "Added support to select GPU using metal on Apple Intel or Apple Silicon using --main-gpu index",
        "labels": [
          "examples",
          "Review Complexity : Low",
          "ggml",
          "Apple Metal"
        ],
        "created_at": "2024-08-10T03:11:23Z",
        "updated_at": "2024-08-10T13:33:58Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/8962",
        "categories": [
          "native_backends",
          "models_architectures",
          "api_ui_docs"
        ]
      },
      {
        "number": 8986,
        "kind": "pr",
        "state": "open",
        "title": "Revert \"ggml : remove OpenCL (#7735) + (#8235)\"",
        "labels": [
          "documentation",
          "build",
          "script",
          "nix",
          "examples",
          "python",
          "devops",
          "ggml",
          "SYCL",
          "Apple Metal"
        ],
        "created_at": "2024-08-11T15:47:50Z",
        "updated_at": "2025-07-17T16:42:26Z",
        "closed_at": null,
        "comments": 17,
        "url": "https://github.com/ggml-org/llama.cpp/pull/8986",
        "categories": [
          "native_backends",
          "models_architectures",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 8989,
        "kind": "pr",
        "state": "open",
        "title": "feat: whitelist jina bert v2 for llama-server embedding",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2024-08-12T00:06:09Z",
        "updated_at": "2024-08-12T07:21:34Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/8989",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 8992,
        "kind": "pr",
        "state": "open",
        "title": "llama : support raw NUL bytes in tokens",
        "labels": [
          "need feedback",
          "examples",
          "python",
          "ggml"
        ],
        "created_at": "2024-08-12T01:33:35Z",
        "updated_at": "2024-08-12T06:59:38Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/8992",
        "categories": [
          "quantization",
          "models_architectures",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 9016,
        "kind": "issue",
        "state": "open",
        "title": "Bug: llama-server scales default context incorrectly for multiple slots",
        "labels": [
          "bug",
          "medium severity"
        ],
        "created_at": "2024-08-13T08:55:05Z",
        "updated_at": "2024-09-22T16:06:24Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/9016",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 9022,
        "kind": "pr",
        "state": "open",
        "title": "add hf2gguf conv format of q4_0 q4_1 q5_0 q5_1",
        "labels": [
          "python"
        ],
        "created_at": "2024-08-14T06:45:54Z",
        "updated_at": "2024-08-15T06:28:09Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/9022",
        "categories": [
          "quantization",
          "models_architectures",
          "api_ui_docs"
        ]
      },
      {
        "number": 9076,
        "kind": "pr",
        "state": "open",
        "title": "server: add repeat penalty sigmoid",
        "labels": [
          "testing",
          "examples",
          "server"
        ],
        "created_at": "2024-08-18T12:57:14Z",
        "updated_at": "2024-09-25T18:02:14Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/9076",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 9078,
        "kind": "pr",
        "state": "open",
        "title": "llava : fix occasional undefined behavior crash",
        "labels": [
          "examples"
        ],
        "created_at": "2024-08-18T14:18:37Z",
        "updated_at": "2024-08-20T13:05:31Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/9078",
        "categories": [
          "models_architectures",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 9095,
        "kind": "issue",
        "state": "open",
        "title": "Refactor: Add more typechecking to GGUFWriter.add_key_value",
        "labels": [
          "help wanted",
          "refactoring"
        ],
        "created_at": "2024-08-19T21:08:18Z",
        "updated_at": "2026-06-09T05:59:32Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/9095",
        "categories": [
          "quantization",
          "models_architectures",
          "correctness_testing"
        ]
      },
      {
        "number": 9102,
        "kind": "pr",
        "state": "open",
        "title": "Fix todo: avoid relying on `logits_all == true` in perplexity_v2",
        "labels": [
          "android",
          "examples",
          "server"
        ],
        "created_at": "2024-08-20T09:33:37Z",
        "updated_at": "2024-08-27T04:35:44Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/9102",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 9123,
        "kind": "pr",
        "state": "open",
        "title": "Add to README: ConfiChat - a lightweight, standalone, multi-platform,…",
        "labels": [
          "Review Complexity : Low"
        ],
        "created_at": "2024-08-21T18:59:35Z",
        "updated_at": "2024-08-30T12:16:12Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/9123",
        "categories": [
          "models_architectures",
          "api_ui_docs"
        ]
      },
      {
        "number": 9131,
        "kind": "pr",
        "state": "open",
        "title": "sync ggml-vocab-qwen2",
        "labels": [
          "Review Complexity : Low"
        ],
        "created_at": "2024-08-22T09:46:50Z",
        "updated_at": "2024-08-30T12:15:26Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/9131",
        "categories": [
          "quantization",
          "models_architectures",
          "structured_generation_tools",
          "loader_conversion"
        ]
      },
      {
        "number": 9177,
        "kind": "pr",
        "state": "open",
        "title": "ggml: skip excess iteration for pair whose vars same element when i2 == i1",
        "labels": [
          "Review Complexity : High",
          "ggml"
        ],
        "created_at": "2024-08-25T23:46:32Z",
        "updated_at": "2024-08-30T12:08:42Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/9177",
        "categories": [
          "quantization",
          "models_architectures"
        ]
      },
      {
        "number": 9179,
        "kind": "pr",
        "state": "open",
        "title": "llama: changed default type IQ2_XS to IQ2_S for LLAMA_FTYPE_MOSTLY_IQ2_S",
        "labels": [
          "Review Complexity : Low"
        ],
        "created_at": "2024-08-26T00:19:55Z",
        "updated_at": "2024-09-01T02:41:03Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/9179",
        "categories": [
          "quantization",
          "models_architectures"
        ]
      },
      {
        "number": 9206,
        "kind": "pr",
        "state": "open",
        "title": "Fix for Debian CMake package creation",
        "labels": [
          "build",
          "Review Complexity : Low"
        ],
        "created_at": "2024-08-27T20:49:01Z",
        "updated_at": "2024-08-30T12:07:57Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/9206",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "build_platform_ops"
        ]
      },
      {
        "number": 9209,
        "kind": "pr",
        "state": "open",
        "title": "Add support for Phi-3.5-vision-instruct",
        "labels": [
          "examples",
          "Review Complexity : Medium"
        ],
        "created_at": "2024-08-27T22:24:56Z",
        "updated_at": "2025-12-01T05:54:36Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/9209",
        "categories": [
          "quantization",
          "models_architectures",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 9220,
        "kind": "pr",
        "state": "open",
        "title": "nix: fix regression: asking for deprecated autoAddOpenGLRunpath",
        "labels": [
          "nix",
          "Review Complexity : Low",
          "devops"
        ],
        "created_at": "2024-08-28T11:29:51Z",
        "updated_at": "2024-08-30T18:47:27Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/9220",
        "categories": [
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 9286,
        "kind": "issue",
        "state": "open",
        "title": "Bug: llama_print_timings seems to accumulate load_time/total_time in `llama-bench`",
        "labels": [
          "bug",
          "low severity"
        ],
        "created_at": "2024-09-03T06:04:18Z",
        "updated_at": "2024-09-19T03:50:33Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/9286",
        "categories": [
          "quantization",
          "models_architectures",
          "correctness_testing"
        ]
      },
      {
        "number": 9289,
        "kind": "issue",
        "state": "open",
        "title": "changelog : `libllama` API",
        "labels": [
          "documentation",
          "roadmap"
        ],
        "created_at": "2024-09-03T06:48:45Z",
        "updated_at": "2026-08-05T16:33:18Z",
        "closed_at": null,
        "comments": 12,
        "url": "https://github.com/ggml-org/llama.cpp/issues/9289",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 9291,
        "kind": "issue",
        "state": "open",
        "title": "changelog : `llama-server` REST API",
        "labels": [
          "documentation",
          "roadmap"
        ],
        "created_at": "2024-09-03T06:56:11Z",
        "updated_at": "2026-05-27T11:38:03Z",
        "closed_at": null,
        "comments": 19,
        "url": "https://github.com/ggml-org/llama.cpp/issues/9291",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 9316,
        "kind": "issue",
        "state": "open",
        "title": "Bug: llama-perplexity error using multiple-choice binary data",
        "labels": [
          "bug",
          "medium severity"
        ],
        "created_at": "2024-09-04T19:41:26Z",
        "updated_at": "2025-08-14T08:02:33Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/9316",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 9325,
        "kind": "pr",
        "state": "open",
        "title": "cmake : use ggml-metal.metal from source dir to build default.metallib",
        "labels": [],
        "created_at": "2024-09-05T16:39:00Z",
        "updated_at": "2024-09-06T07:50:55Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/9325",
        "categories": [
          "native_backends",
          "models_architectures",
          "build_platform_ops"
        ]
      },
      {
        "number": 9328,
        "kind": "pr",
        "state": "open",
        "title": "ggml/kompute: Add multi-GPU support",
        "labels": [
          "ggml",
          "Kompute"
        ],
        "created_at": "2024-09-06T03:07:59Z",
        "updated_at": "2025-01-12T00:04:03Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/9328",
        "categories": [
          "native_backends",
          "models_architectures"
        ]
      },
      {
        "number": 9391,
        "kind": "issue",
        "state": "open",
        "title": "Bug: cannot create std::vector larger than max_size()",
        "labels": [
          "bug",
          "medium severity"
        ],
        "created_at": "2024-09-09T15:52:21Z",
        "updated_at": "2024-09-21T00:43:29Z",
        "closed_at": null,
        "comments": 10,
        "url": "https://github.com/ggml-org/llama.cpp/issues/9391",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 9403,
        "kind": "pr",
        "state": "open",
        "title": "feat: support internvl",
        "labels": [
          "examples",
          "python"
        ],
        "created_at": "2024-09-10T07:15:24Z",
        "updated_at": "2025-02-22T13:41:54Z",
        "closed_at": null,
        "comments": 10,
        "url": "https://github.com/ggml-org/llama.cpp/pull/9403",
        "categories": [
          "models_architectures",
          "api_ui_docs"
        ]
      },
      {
        "number": 9455,
        "kind": "issue",
        "state": "open",
        "title": "Bug: loading llava models fails",
        "labels": [
          "bug",
          "critical severity"
        ],
        "created_at": "2024-09-12T16:59:37Z",
        "updated_at": "2024-09-12T17:05:28Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/9455",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 9492,
        "kind": "issue",
        "state": "open",
        "title": "Bug: llama-server api first query very slow",
        "labels": [
          "bug",
          "medium severity"
        ],
        "created_at": "2024-09-15T06:37:09Z",
        "updated_at": "2026-03-17T10:42:27Z",
        "closed_at": null,
        "comments": 16,
        "url": "https://github.com/ggml-org/llama.cpp/issues/9492",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 9493,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: RDMA support for rpc back ends",
        "labels": [
          "enhancement",
          "research 🔬"
        ],
        "created_at": "2024-09-15T09:45:09Z",
        "updated_at": "2026-08-24T08:29:10Z",
        "closed_at": null,
        "comments": 20,
        "url": "https://github.com/ggml-org/llama.cpp/issues/9493",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 9603,
        "kind": "pr",
        "state": "open",
        "title": "keep the minimum `min_keep` value to 1 in sampling",
        "labels": [],
        "created_at": "2024-09-23T08:37:14Z",
        "updated_at": "2024-09-23T10:47:22Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/9603",
        "categories": [
          "models_architectures",
          "structured_generation_tools"
        ]
      },
      {
        "number": 9648,
        "kind": "pr",
        "state": "open",
        "title": "[Draft] Tensor Parallel support to llama.cpp",
        "labels": [
          "ggml",
          "SYCL"
        ],
        "created_at": "2024-09-26T02:36:27Z",
        "updated_at": "2025-12-27T18:39:15Z",
        "closed_at": null,
        "comments": 10,
        "url": "https://github.com/ggml-org/llama.cpp/pull/9648",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving"
        ]
      },
      {
        "number": 9659,
        "kind": "pr",
        "state": "open",
        "title": "Introduce Graph Profiler",
        "labels": [
          "ggml"
        ],
        "created_at": "2024-09-26T16:45:38Z",
        "updated_at": "2025-12-18T01:29:14Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/pull/9659",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 9833,
        "kind": "pr",
        "state": "open",
        "title": "llama : adds llama-grammar memoization stacks (#4218)",
        "labels": [
          "testing",
          "examples"
        ],
        "created_at": "2024-10-11T01:39:51Z",
        "updated_at": "2024-12-16T00:50:21Z",
        "closed_at": null,
        "comments": 10,
        "url": "https://github.com/ggml-org/llama.cpp/pull/9833",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "structured_generation_tools",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 9855,
        "kind": "pr",
        "state": "open",
        "title": "New quant strategy / FTYPE IQ3_XL 4bpw",
        "labels": [
          "examples",
          "python"
        ],
        "created_at": "2024-10-12T01:03:26Z",
        "updated_at": "2024-10-12T01:20:22Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/9855",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 9910,
        "kind": "pr",
        "state": "open",
        "title": "llama : bump max layers from 512 to 1024",
        "labels": [],
        "created_at": "2024-10-16T12:29:39Z",
        "updated_at": "2024-10-16T12:29:39Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/9910",
        "categories": [
          "models_architectures",
          "correctness_testing"
        ]
      },
      {
        "number": 9916,
        "kind": "pr",
        "state": "open",
        "title": "consolidated.safetensors",
        "labels": [
          "python"
        ],
        "created_at": "2024-10-16T21:29:51Z",
        "updated_at": "2024-10-17T05:54:51Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/9916",
        "categories": [
          "models_architectures",
          "loader_conversion",
          "api_ui_docs"
        ]
      },
      {
        "number": 9922,
        "kind": "pr",
        "state": "open",
        "title": "sample: maintain token count in penalty sampler context",
        "labels": [],
        "created_at": "2024-10-17T10:13:48Z",
        "updated_at": "2024-10-18T11:09:21Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/9922",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "structured_generation_tools"
        ]
      },
      {
        "number": 9949,
        "kind": "issue",
        "state": "open",
        "title": "Bug: Segmentation fault when running speculative decoding",
        "labels": [
          "bug",
          "critical severity"
        ],
        "created_at": "2024-10-19T04:03:33Z",
        "updated_at": "2025-03-12T15:26:22Z",
        "closed_at": null,
        "comments": 16,
        "url": "https://github.com/ggml-org/llama.cpp/issues/9949",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 9967,
        "kind": "pr",
        "state": "open",
        "title": "LLamaCausalLM add support for tokenizer.json",
        "labels": [
          "python"
        ],
        "created_at": "2024-10-20T22:58:09Z",
        "updated_at": "2024-10-22T13:56:09Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/9967",
        "categories": [
          "native_backends",
          "models_architectures",
          "loader_conversion",
          "api_ui_docs"
        ]
      },
      {
        "number": 10004,
        "kind": "pr",
        "state": "open",
        "title": "llama : rename batch.logits to batch.output",
        "labels": [
          "breaking change",
          "android",
          "examples",
          "server"
        ],
        "created_at": "2024-10-22T13:29:16Z",
        "updated_at": "2025-02-06T07:01:40Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/10004",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 10005,
        "kind": "issue",
        "state": "open",
        "title": "llama : enable FA by default and disable it per-layer",
        "labels": [
          "enhancement",
          "roadmap"
        ],
        "created_at": "2024-10-22T14:07:59Z",
        "updated_at": "2025-02-04T07:34:55Z",
        "closed_at": null,
        "comments": 18,
        "url": "https://github.com/ggml-org/llama.cpp/issues/10005",
        "categories": [
          "models_architectures",
          "build_platform_ops"
        ]
      },
      {
        "number": 10008,
        "kind": "pr",
        "state": "open",
        "title": "feat(convert_hf_to_gguf): support q4_0 and q4_1 quantifications",
        "labels": [
          "python"
        ],
        "created_at": "2024-10-22T16:53:10Z",
        "updated_at": "2024-10-31T19:11:56Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/10008",
        "categories": [
          "quantization",
          "models_architectures",
          "loader_conversion",
          "api_ui_docs"
        ]
      },
      {
        "number": 10044,
        "kind": "pr",
        "state": "open",
        "title": "Fix logging from llama-llava-cli",
        "labels": [
          "examples"
        ],
        "created_at": "2024-10-25T15:10:15Z",
        "updated_at": "2024-10-29T08:32:00Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/10044",
        "categories": [
          "models_architectures",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 10048,
        "kind": "pr",
        "state": "open",
        "title": "sampling: add K-Shift sampler",
        "labels": [
          "testing",
          "examples",
          "server"
        ],
        "created_at": "2024-10-25T18:51:04Z",
        "updated_at": "2024-11-09T10:09:40Z",
        "closed_at": null,
        "comments": 12,
        "url": "https://github.com/ggml-org/llama.cpp/pull/10048",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 10055,
        "kind": "pr",
        "state": "open",
        "title": "add FP8 support to gguf/llama:",
        "labels": [
          "build",
          "script",
          "testing",
          "examples",
          "ggml",
          "Tensor Encoding Scheme"
        ],
        "created_at": "2024-10-26T02:30:39Z",
        "updated_at": "2026-01-12T01:11:38Z",
        "closed_at": null,
        "comments": 32,
        "url": "https://github.com/ggml-org/llama.cpp/pull/10055",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 10101,
        "kind": "pr",
        "state": "open",
        "title": "llama : enable flash attn automatically when supported (WIP)",
        "labels": [],
        "created_at": "2024-10-30T22:32:13Z",
        "updated_at": "2024-10-30T22:32:13Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/10101",
        "categories": [
          "native_backends",
          "models_architectures"
        ]
      },
      {
        "number": 10102,
        "kind": "pr",
        "state": "open",
        "title": "`main`: add test-cli + ensure completion goes to stdout even w/ --log-disable",
        "labels": [
          "testing",
          "examples",
          "ggml"
        ],
        "created_at": "2024-10-30T23:11:14Z",
        "updated_at": "2024-10-31T08:51:45Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/10102",
        "categories": [
          "models_architectures",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 10131,
        "kind": "pr",
        "state": "open",
        "title": "Add user-provided tokenizer/detokenizer functionality",
        "labels": [
          "examples"
        ],
        "created_at": "2024-11-02T01:35:47Z",
        "updated_at": "2024-11-02T11:28:41Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/10131",
        "categories": [
          "models_architectures",
          "loader_conversion",
          "api_ui_docs"
        ]
      },
      {
        "number": 10142,
        "kind": "pr",
        "state": "open",
        "title": "Fix docker locale issue (#6267)",
        "labels": [
          "devops"
        ],
        "created_at": "2024-11-02T17:35:43Z",
        "updated_at": "2024-11-04T19:00:44Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/10142",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 10180,
        "kind": "issue",
        "state": "open",
        "title": "ggml : refactor ggml-cpu.c into multiple C++ source files",
        "labels": [
          "refactoring",
          "roadmap"
        ],
        "created_at": "2024-11-05T07:12:48Z",
        "updated_at": "2025-03-31T16:19:13Z",
        "closed_at": null,
        "comments": 17,
        "url": "https://github.com/ggml-org/llama.cpp/issues/10180",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving"
        ]
      },
      {
        "number": 10181,
        "kind": "pr",
        "state": "open",
        "title": "Introduce New Lookup-Table(LUT)-Based Matrix Multiplication Method",
        "labels": [
          "python",
          "ggml",
          "Tensor Encoding Scheme"
        ],
        "created_at": "2024-11-05T08:36:26Z",
        "updated_at": "2025-05-15T11:03:42Z",
        "closed_at": null,
        "comments": 29,
        "url": "https://github.com/ggml-org/llama.cpp/pull/10181",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 10209,
        "kind": "pr",
        "state": "open",
        "title": "docs: add doxygen documentation",
        "labels": [
          "build"
        ],
        "created_at": "2024-11-08T00:22:17Z",
        "updated_at": "2024-11-08T06:42:10Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/10209",
        "categories": [
          "models_architectures",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 10220,
        "kind": "pr",
        "state": "open",
        "title": "metal : use F16 math in mul_mat kernels",
        "labels": [],
        "created_at": "2024-11-08T13:06:22Z",
        "updated_at": "2024-12-12T09:10:52Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/10220",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 10275,
        "kind": "pr",
        "state": "open",
        "title": "Test tokenizer-0.py rewrite",
        "labels": [
          "testing",
          "python"
        ],
        "created_at": "2024-11-13T06:05:39Z",
        "updated_at": "2024-11-13T06:05:56Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/10275",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 10276,
        "kind": "pr",
        "state": "open",
        "title": "Add try/except to test-tokenizer-random.py",
        "labels": [
          "testing",
          "python"
        ],
        "created_at": "2024-11-13T06:29:53Z",
        "updated_at": "2024-11-17T16:38:30Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/10276",
        "categories": [
          "models_architectures",
          "loader_conversion",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 10310,
        "kind": "pr",
        "state": "open",
        "title": "chore : Fix the error when compiling rocm build on windows using cmake",
        "labels": [
          "documentation"
        ],
        "created_at": "2024-11-15T12:41:27Z",
        "updated_at": "2024-11-16T00:57:18Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/10310",
        "categories": [
          "native_backends",
          "models_architectures",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 10315,
        "kind": "pr",
        "state": "open",
        "title": "Add complete implementation of the classical PCA algorithm with covar…",
        "labels": [
          "examples"
        ],
        "created_at": "2024-11-15T17:45:02Z",
        "updated_at": "2026-03-25T11:24:51Z",
        "closed_at": null,
        "comments": 9,
        "url": "https://github.com/ggml-org/llama.cpp/pull/10315",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 10321,
        "kind": "issue",
        "state": "open",
        "title": "Bug: llama-gbnf-validator parses grammar but gets a seg fault when validating an input string against the grammar",
        "labels": [
          "bug",
          "critical severity"
        ],
        "created_at": "2024-11-15T20:27:54Z",
        "updated_at": "2025-01-31T08:53:13Z",
        "closed_at": null,
        "comments": 10,
        "url": "https://github.com/ggml-org/llama.cpp/issues/10321",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 10343,
        "kind": "pr",
        "state": "open",
        "title": "Refactor/tinyblas",
        "labels": [
          "documentation",
          "build",
          "demo",
          "ggml"
        ],
        "created_at": "2024-11-16T22:21:10Z",
        "updated_at": "2024-11-17T06:49:13Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/10343",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 10353,
        "kind": "pr",
        "state": "open",
        "title": "common: compile shared lib, and export some c functions",
        "labels": [],
        "created_at": "2024-11-17T06:33:52Z",
        "updated_at": "2024-11-18T13:29:10Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/10353",
        "categories": [
          "models_architectures",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 10420,
        "kind": "issue",
        "state": "open",
        "title": "Bug: Vulkan vk::DeviceLostError with multithreaded environment",
        "labels": [
          "bug",
          "low severity"
        ],
        "created_at": "2024-11-20T06:45:19Z",
        "updated_at": "2025-02-25T08:12:51Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/10420",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 10435,
        "kind": "issue",
        "state": "open",
        "title": "Bug: Severe Performance Degradation on Q4_0 CPU-only with MacOS / Apple Silicon M2, after PR#9921 / Version 4081",
        "labels": [
          "bug"
        ],
        "created_at": "2024-11-20T17:06:21Z",
        "updated_at": "2024-11-25T12:08:04Z",
        "closed_at": null,
        "comments": 12,
        "url": "https://github.com/ggml-org/llama.cpp/issues/10435",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 10453,
        "kind": "issue",
        "state": "open",
        "title": "ggml : add ANE backend",
        "labels": [
          "help wanted",
          "research 🔬",
          "roadmap"
        ],
        "created_at": "2024-11-22T08:20:22Z",
        "updated_at": "2026-05-10T09:27:27Z",
        "closed_at": null,
        "comments": 21,
        "url": "https://github.com/ggml-org/llama.cpp/issues/10453",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 10510,
        "kind": "pr",
        "state": "open",
        "title": "speculative : add infill mode",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2024-11-26T09:15:55Z",
        "updated_at": "2024-11-26T09:32:42Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/10510",
        "categories": [
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "api_ui_docs"
        ]
      },
      {
        "number": 10528,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Inconsistent Vulkan segfault",
        "labels": [
          "bug"
        ],
        "created_at": "2024-11-26T19:54:03Z",
        "updated_at": "2026-06-18T02:27:33Z",
        "closed_at": null,
        "comments": 82,
        "url": "https://github.com/ggml-org/llama.cpp/issues/10528",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 10571,
        "kind": "pr",
        "state": "open",
        "title": "llama: Add generic abort to token_decode_internal",
        "labels": [],
        "created_at": "2024-11-28T20:49:07Z",
        "updated_at": "2024-11-29T20:09:34Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/10571",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 10582,
        "kind": "pr",
        "state": "open",
        "title": "Example: Add support T5 in swift.ui example",
        "labels": [
          "examples"
        ],
        "created_at": "2024-11-29T15:38:38Z",
        "updated_at": "2024-11-29T15:38:54Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/10582",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "api_ui_docs"
        ]
      },
      {
        "number": 10614,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: interface for model quantization is not fully C-compatible",
        "labels": [
          "bug"
        ],
        "created_at": "2024-12-01T13:03:11Z",
        "updated_at": "2024-12-26T22:33:54Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/10614",
        "categories": [
          "quantization",
          "models_architectures",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 10660,
        "kind": "pr",
        "state": "open",
        "title": "server: add request aggregation functionallity ",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2024-12-04T17:51:24Z",
        "updated_at": "2024-12-04T22:09:12Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/10660",
        "categories": [
          "scheduling_serving",
          "multimodal_embedding",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 10683,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: softmax may get error answer when src0->ne[3]!=1 on cuda",
        "labels": [
          "bug"
        ],
        "created_at": "2024-12-06T08:08:13Z",
        "updated_at": "2024-12-06T15:01:02Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/10683",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing"
        ]
      },
      {
        "number": 10685,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: llama-server hot swapping cvectors via API like we can do with LoRA adapters now",
        "labels": [
          "enhancement",
          "good first issue"
        ],
        "created_at": "2024-12-06T09:10:27Z",
        "updated_at": "2026-06-24T23:47:51Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/10685",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 10706,
        "kind": "pr",
        "state": "open",
        "title": "cmake: fix for building with no internet connection",
        "labels": [
          "examples"
        ],
        "created_at": "2024-12-07T11:50:56Z",
        "updated_at": "2024-12-07T18:48:15Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/10706",
        "categories": [
          "models_architectures",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 10718,
        "kind": "pr",
        "state": "open",
        "title": "server : chunked prefill support",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2024-12-08T07:55:39Z",
        "updated_at": "2024-12-08T07:55:55Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/10718",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "api_ui_docs"
        ]
      },
      {
        "number": 10732,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: server provides strutured output for response_format: json_object, but not for response_format: json_schema",
        "labels": [
          "enhancement",
          "good first issue"
        ],
        "created_at": "2024-12-09T04:39:47Z",
        "updated_at": "2026-08-19T17:21:41Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/issues/10732",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 10742,
        "kind": "pr",
        "state": "open",
        "title": "server: Add timeout to stop the server automatically when idling for too long.",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2024-12-09T23:42:45Z",
        "updated_at": "2025-01-18T08:47:22Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/pull/10742",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 10743,
        "kind": "pr",
        "state": "open",
        "title": "Cuda build doc",
        "labels": [
          "documentation"
        ],
        "created_at": "2024-12-10T01:44:38Z",
        "updated_at": "2024-12-12T09:15:06Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/10743",
        "categories": [
          "native_backends",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 10744,
        "kind": "pr",
        "state": "open",
        "title": "add verbosity -1 to log token, so can output only tokens with -lv -1",
        "labels": [
          "examples"
        ],
        "created_at": "2024-12-10T02:00:59Z",
        "updated_at": "2024-12-10T02:01:13Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/10744",
        "categories": [
          "api_ui_docs"
        ]
      },
      {
        "number": 10747,
        "kind": "issue",
        "state": "open",
        "title": "Compile bug: ios swift xcode build error when upgrade to llama : use cmake for swift build ",
        "labels": [
          "help wanted",
          "good first issue",
          "build"
        ],
        "created_at": "2024-12-10T05:12:25Z",
        "updated_at": "2025-03-21T15:58:29Z",
        "closed_at": null,
        "comments": 41,
        "url": "https://github.com/ggml-org/llama.cpp/issues/10747",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 10835,
        "kind": "pr",
        "state": "open",
        "title": "Fix compilation on Pop!_OS 22.04 LTS CUDA",
        "labels": [
          "Nvidia GPU",
          "ggml"
        ],
        "created_at": "2024-12-15T08:46:44Z",
        "updated_at": "2025-04-25T17:08:48Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/10835",
        "categories": [
          "native_backends",
          "models_architectures",
          "build_platform_ops"
        ]
      },
      {
        "number": 10873,
        "kind": "pr",
        "state": "open",
        "title": "llama: Ensure KV cache is fully defragmented.",
        "labels": [],
        "created_at": "2024-12-17T20:46:35Z",
        "updated_at": "2024-12-17T20:46:35Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/10873",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving"
        ]
      },
      {
        "number": 10928,
        "kind": "pr",
        "state": "open",
        "title": "Allow user to compile with any cuda version using github actions",
        "labels": [
          "devops"
        ],
        "created_at": "2024-12-21T07:30:22Z",
        "updated_at": "2026-04-24T21:25:39Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/10928",
        "categories": [
          "native_backends",
          "build_platform_ops"
        ]
      },
      {
        "number": 10972,
        "kind": "pr",
        "state": "open",
        "title": "Removed unnecessary iteration of batch n_tokens on sequence embedding…",
        "labels": [
          "examples"
        ],
        "created_at": "2024-12-25T11:21:21Z",
        "updated_at": "2024-12-31T15:10:20Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/10972",
        "categories": [
          "scheduling_serving",
          "multimodal_embedding",
          "api_ui_docs"
        ]
      },
      {
        "number": 10982,
        "kind": "issue",
        "state": "open",
        "title": "Research: Performance differences between Metal (macOS) and Vulkan (Linux)",
        "labels": [
          "research 🔬",
          "Vulkan",
          "Apple Metal"
        ],
        "created_at": "2024-12-26T11:12:21Z",
        "updated_at": "2026-08-07T00:52:46Z",
        "closed_at": null,
        "comments": 49,
        "url": "https://github.com/ggml-org/llama.cpp/issues/10982",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 11070,
        "kind": "pr",
        "state": "open",
        "title": "server : POC OAI-compat TTS using OuteTTS",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2025-01-03T22:42:48Z",
        "updated_at": "2025-01-08T16:23:42Z",
        "closed_at": null,
        "comments": 10,
        "url": "https://github.com/ggml-org/llama.cpp/pull/11070",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 11183,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cuda : add TQ2_0 kernels, for ternary inference on GPU",
        "labels": [
          "enhancement",
          "performance",
          "testing",
          "Nvidia GPU",
          "python",
          "Review Complexity : High",
          "ggml",
          "Apple Metal"
        ],
        "created_at": "2025-01-10T20:10:49Z",
        "updated_at": "2026-05-25T08:36:29Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/pull/11183",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 11185,
        "kind": "pr",
        "state": "open",
        "title": "Fix GGML not compiling on macOS with GCC",
        "labels": [
          "ggml"
        ],
        "created_at": "2025-01-11T09:06:37Z",
        "updated_at": "2025-01-11T09:06:52Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/11185",
        "categories": [
          "native_backends",
          "build_platform_ops"
        ]
      },
      {
        "number": 11187,
        "kind": "pr",
        "state": "open",
        "title": "ggml doesn't use sse42, specify only up to sse4.1",
        "labels": [
          "ggml"
        ],
        "created_at": "2025-01-11T09:16:14Z",
        "updated_at": "2025-01-11T12:44:10Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/11187",
        "categories": [
          "multimodal_embedding",
          "build_platform_ops"
        ]
      },
      {
        "number": 11188,
        "kind": "pr",
        "state": "open",
        "title": "Fix ggml-cuda using a driver symbol in NO_VMM mode",
        "labels": [
          "Nvidia GPU",
          "ggml"
        ],
        "created_at": "2025-01-11T09:19:04Z",
        "updated_at": "2025-01-11T13:32:15Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/11188",
        "categories": [
          "native_backends",
          "api_ui_docs"
        ]
      },
      {
        "number": 11189,
        "kind": "pr",
        "state": "open",
        "title": "CUDA op getrows fails for long sequences",
        "labels": [
          "Nvidia GPU",
          "ggml"
        ],
        "created_at": "2025-01-11T09:25:00Z",
        "updated_at": "2025-01-11T10:21:38Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/11189",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 11190,
        "kind": "pr",
        "state": "open",
        "title": "Allow compiling cuda without mmq and flash attention",
        "labels": [
          "Nvidia GPU",
          "ggml"
        ],
        "created_at": "2025-01-11T09:29:26Z",
        "updated_at": "2025-01-11T10:09:17Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/11190",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "build_platform_ops"
        ]
      },
      {
        "number": 11226,
        "kind": "pr",
        "state": "open",
        "title": "Build linux CUDA releases suitable for Colab & other platforms on 12.2",
        "labels": [
          "devops",
          "ggml"
        ],
        "created_at": "2025-01-14T02:13:56Z",
        "updated_at": "2025-02-23T19:06:57Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/11226",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 11246,
        "kind": "pr",
        "state": "open",
        "title": "fix makefile and cmake logic for AARCH64",
        "labels": [
          "ggml"
        ],
        "created_at": "2025-01-15T03:48:22Z",
        "updated_at": "2025-01-26T22:51:17Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/11246",
        "categories": [
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 11291,
        "kind": "pr",
        "state": "open",
        "title": "Align structures for 64bit, reorder params and ignore error-warn for Clang 19",
        "labels": [
          "ggml"
        ],
        "created_at": "2025-01-18T17:51:58Z",
        "updated_at": "2025-01-21T19:06:32Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/11291",
        "categories": [
          "other"
        ]
      },
      {
        "number": 11294,
        "kind": "pr",
        "state": "open",
        "title": "Removed const references for simple types and structures less 16 bytes",
        "labels": [
          "ggml"
        ],
        "created_at": "2025-01-18T20:54:18Z",
        "updated_at": "2025-01-21T18:58:58Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/11294",
        "categories": [
          "build_platform_ops"
        ]
      },
      {
        "number": 11297,
        "kind": "pr",
        "state": "open",
        "title": "ggml: reserve in gguf_writer and added const pointers as params",
        "labels": [
          "ggml"
        ],
        "created_at": "2025-01-18T21:56:04Z",
        "updated_at": "2025-01-18T22:02:42Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/11297",
        "categories": [
          "quantization"
        ]
      },
      {
        "number": 11308,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: segfault on Alpine linux docker image",
        "labels": [
          "bug"
        ],
        "created_at": "2025-01-20T11:12:22Z",
        "updated_at": "2025-01-28T21:20:05Z",
        "closed_at": null,
        "comments": 11,
        "url": "https://github.com/ggml-org/llama.cpp/issues/11308",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 11326,
        "kind": "pr",
        "state": "open",
        "title": "cpu_pnp_strategy changes",
        "labels": [
          "ggml"
        ],
        "created_at": "2025-01-21T09:16:21Z",
        "updated_at": "2025-01-21T09:17:43Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/11326",
        "categories": [
          "build_platform_ops"
        ]
      },
      {
        "number": 11371,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Error running multiple contexts from multiple threads at the same time with Vulkan",
        "labels": [
          "bug"
        ],
        "created_at": "2025-01-23T13:32:49Z",
        "updated_at": "2025-06-17T05:21:54Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/11371",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 11410,
        "kind": "pr",
        "state": "open",
        "title": "Llama-bench: allow benchmarking lora impact",
        "labels": [
          "examples"
        ],
        "created_at": "2025-01-25T10:12:59Z",
        "updated_at": "2025-01-28T14:01:49Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/11410",
        "categories": [
          "quantization",
          "models_architectures",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 11477,
        "kind": "pr",
        "state": "open",
        "title": "Readme Update: Added IRIS under UI section",
        "labels": [],
        "created_at": "2025-01-29T06:55:03Z",
        "updated_at": "2025-01-29T12:26:34Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/11477",
        "categories": [
          "models_architectures",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 11514,
        "kind": "issue",
        "state": "open",
        "title": "Move gguf fuzzers to the llama.cpp repository",
        "labels": [
          "enhancement",
          "testing",
          "roadmap"
        ],
        "created_at": "2025-01-30T15:57:53Z",
        "updated_at": "2025-05-13T17:18:23Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/issues/11514",
        "categories": [
          "quantization",
          "models_architectures",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 11534,
        "kind": "pr",
        "state": "open",
        "title": "readme : add UI of  ChatPDFLocal which is based by llama.cpp for running LLMs on Mac",
        "labels": [],
        "created_at": "2025-01-31T06:43:55Z",
        "updated_at": "2025-02-16T14:56:01Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/11534",
        "categories": [
          "models_architectures",
          "api_ui_docs"
        ]
      },
      {
        "number": 11557,
        "kind": "pr",
        "state": "open",
        "title": "Add support for Deepseek-R1 flash attention",
        "labels": [
          "Nvidia GPU",
          "ggml"
        ],
        "created_at": "2025-01-31T18:53:05Z",
        "updated_at": "2025-03-26T20:04:10Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/pull/11557",
        "categories": [
          "native_backends",
          "models_architectures"
        ]
      },
      {
        "number": 11558,
        "kind": "pr",
        "state": "open",
        "title": "Update CMakeLists.txt",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2025-01-31T19:04:03Z",
        "updated_at": "2025-02-01T16:16:23Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/11558",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 11577,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: resize an existing context",
        "labels": [
          "enhancement",
          "roadmap"
        ],
        "created_at": "2025-02-01T15:51:53Z",
        "updated_at": "2026-03-04T14:18:31Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/issues/11577",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 11597,
        "kind": "pr",
        "state": "open",
        "title": "scripts: added inline script metadata per PEP 723",
        "labels": [
          "script",
          "python"
        ],
        "created_at": "2025-02-02T16:54:30Z",
        "updated_at": "2025-02-02T16:55:46Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/11597",
        "categories": [
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 11610,
        "kind": "pr",
        "state": "open",
        "title": "Clean up Test Script + Update it to work on Instruct Tuned Models",
        "labels": [
          "examples",
          "SYCL"
        ],
        "created_at": "2025-02-03T05:19:37Z",
        "updated_at": "2025-02-03T05:19:52Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/11610",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 11642,
        "kind": "pr",
        "state": "open",
        "title": "[Important] Added README to the Qwen2VL implementation",
        "labels": [
          "examples"
        ],
        "created_at": "2025-02-04T05:20:41Z",
        "updated_at": "2025-02-09T08:46:19Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/11642",
        "categories": [
          "models_architectures",
          "multimodal_embedding",
          "api_ui_docs"
        ]
      },
      {
        "number": 11660,
        "kind": "pr",
        "state": "open",
        "title": "Add information for Podman as well as Docker",
        "labels": [
          "documentation"
        ],
        "created_at": "2025-02-04T14:34:52Z",
        "updated_at": "2025-02-04T19:23:18Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/11660",
        "categories": [
          "models_architectures",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 11692,
        "kind": "pr",
        "state": "open",
        "title": "fix: ensure proper cleanup of img_res_v.data in all code paths",
        "labels": [
          "examples"
        ],
        "created_at": "2025-02-06T02:50:49Z",
        "updated_at": "2025-02-06T02:53:14Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/11692",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 11693,
        "kind": "pr",
        "state": "open",
        "title": "fix: free meta memory in clip model loading",
        "labels": [
          "examples"
        ],
        "created_at": "2025-02-06T03:22:35Z",
        "updated_at": "2025-02-06T03:23:06Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/11693",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 11694,
        "kind": "pr",
        "state": "open",
        "title": "Use named LOG_COL_* colors in examples",
        "labels": [
          "examples"
        ],
        "created_at": "2025-02-06T03:37:58Z",
        "updated_at": "2025-02-06T03:38:13Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/11694",
        "categories": [
          "api_ui_docs"
        ]
      },
      {
        "number": 11716,
        "kind": "pr",
        "state": "open",
        "title": "Supporting Velvet model",
        "labels": [
          "testing",
          "python"
        ],
        "created_at": "2025-02-06T16:08:01Z",
        "updated_at": "2025-05-16T13:09:28Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/11716",
        "categories": [
          "models_architectures",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 11722,
        "kind": "pr",
        "state": "open",
        "title": "ggml : rpc replace reallocation to reuse vector",
        "labels": [
          "ggml"
        ],
        "created_at": "2025-02-06T20:37:29Z",
        "updated_at": "2026-08-20T18:02:21Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/11722",
        "categories": [
          "native_backends",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 11745,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cpu-aarch64: Fix compilation issues",
        "labels": [
          "ggml"
        ],
        "created_at": "2025-02-08T02:08:09Z",
        "updated_at": "2025-06-17T11:49:50Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/11745",
        "categories": [
          "models_architectures"
        ]
      },
      {
        "number": 11785,
        "kind": "pr",
        "state": "open",
        "title": "ggml: move kvalues_iq4nl definition to ggml-common.h",
        "labels": [
          "Nvidia GPU",
          "ggml",
          "SYCL"
        ],
        "created_at": "2025-02-10T08:17:52Z",
        "updated_at": "2025-02-11T01:44:33Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/11785",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures"
        ]
      },
      {
        "number": 11815,
        "kind": "pr",
        "state": "open",
        "title": "Improved KV cache loading performance for Vulkan, resulting in a 20x …",
        "labels": [],
        "created_at": "2025-02-12T06:30:04Z",
        "updated_at": "2025-02-16T21:49:15Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/11815",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 11867,
        "kind": "pr",
        "state": "open",
        "title": "Overlap CUDA graph building and processing to minimize GPU idle time and improve tokens per seconds performance.",
        "labels": [
          "Nvidia GPU",
          "ggml"
        ],
        "created_at": "2025-02-14T10:03:32Z",
        "updated_at": "2025-08-07T07:24:25Z",
        "closed_at": null,
        "comments": 18,
        "url": "https://github.com/ggml-org/llama.cpp/pull/11867",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 11916,
        "kind": "pr",
        "state": "open",
        "title": "Update ggml-backend.cpp",
        "labels": [
          "ggml"
        ],
        "created_at": "2025-02-17T03:27:06Z",
        "updated_at": "2025-02-17T03:27:19Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/11916",
        "categories": [
          "native_backends",
          "models_architectures"
        ]
      },
      {
        "number": 11939,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cpu: add arm64 CPU feature check for OpenBSD, FreeBSD",
        "labels": [
          "ggml"
        ],
        "created_at": "2025-02-18T07:11:36Z",
        "updated_at": "2026-03-25T11:23:31Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/11939",
        "categories": [
          "other"
        ]
      },
      {
        "number": 11970,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: The KV cache is sometimes truncated incorrectly when making v1/chat/completions API calls",
        "labels": [
          "bug",
          "high priority"
        ],
        "created_at": "2025-02-20T11:20:01Z",
        "updated_at": "2026-06-07T23:51:09Z",
        "closed_at": null,
        "comments": 49,
        "url": "https://github.com/ggml-org/llama.cpp/issues/11970",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 12052,
        "kind": "pr",
        "state": "open",
        "title": "Updated readme file for rpc server",
        "labels": [
          "examples"
        ],
        "created_at": "2025-02-24T18:41:51Z",
        "updated_at": "2025-02-28T16:19:47Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/12052",
        "categories": [
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 12063,
        "kind": "pr",
        "state": "open",
        "title": "[WIP]backend: Integrating QNN (Qualcomm AI Engine Direct) as a dedicated backend for Qualcomm NPUs",
        "labels": [
          "documentation",
          "build",
          "ggml"
        ],
        "created_at": "2025-02-25T07:19:54Z",
        "updated_at": "2026-06-15T09:26:28Z",
        "closed_at": null,
        "comments": 14,
        "url": "https://github.com/ggml-org/llama.cpp/pull/12063",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 12067,
        "kind": "pr",
        "state": "open",
        "title": "Cache based tokenization for the server input prompts",
        "labels": [
          "demo",
          "examples",
          "server"
        ],
        "created_at": "2025-02-25T13:08:35Z",
        "updated_at": "2025-07-03T07:49:15Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/12067",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 12073,
        "kind": "pr",
        "state": "open",
        "title": "llama : expose API to retrieve devices associated with the model.",
        "labels": [],
        "created_at": "2025-02-25T20:10:26Z",
        "updated_at": "2025-06-10T00:41:56Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/12073",
        "categories": [
          "models_architectures",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 12127,
        "kind": "pr",
        "state": "open",
        "title": "Server: openai-style lookup decoding",
        "labels": [
          "examples",
          "python",
          "server"
        ],
        "created_at": "2025-03-01T09:25:46Z",
        "updated_at": "2025-03-13T05:56:24Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/12127",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "speculative_generation",
          "api_ui_docs"
        ]
      },
      {
        "number": 12171,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: The inference speed of llama-server is one-third of that of llama-cli",
        "labels": [
          "bug"
        ],
        "created_at": "2025-03-04T08:33:49Z",
        "updated_at": "2025-10-23T15:55:13Z",
        "closed_at": null,
        "comments": 9,
        "url": "https://github.com/ggml-org/llama.cpp/issues/12171",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 12189,
        "kind": "pr",
        "state": "open",
        "title": "libfuse3 supported mounting split gguf's to a single in-memory file",
        "labels": [
          "examples"
        ],
        "created_at": "2025-03-05T05:07:28Z",
        "updated_at": "2025-03-05T17:06:41Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/12189",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "api_ui_docs"
        ]
      },
      {
        "number": 12236,
        "kind": "pr",
        "state": "open",
        "title": "tests: use adaptive number of threads",
        "labels": [
          "testing"
        ],
        "created_at": "2025-03-06T22:04:55Z",
        "updated_at": "2025-03-08T21:49:25Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/12236",
        "categories": [
          "native_backends",
          "correctness_testing"
        ]
      },
      {
        "number": 12243,
        "kind": "pr",
        "state": "open",
        "title": "Fix rocWMMA build documentation",
        "labels": [
          "documentation"
        ],
        "created_at": "2025-03-07T08:30:52Z",
        "updated_at": "2025-06-13T15:27:33Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/12243",
        "categories": [
          "native_backends",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 12260,
        "kind": "pr",
        "state": "open",
        "title": "vulkan: optimization proposals for coopmat1 mul_mm",
        "labels": [
          "Vulkan",
          "ggml"
        ],
        "created_at": "2025-03-07T20:51:02Z",
        "updated_at": "2025-05-10T11:17:30Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/12260",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 12288,
        "kind": "pr",
        "state": "open",
        "title": "`tool-call`: Phi-4 support",
        "labels": [
          "documentation",
          "testing",
          "android",
          "Nvidia GPU",
          "Vulkan",
          "examples",
          "python",
          "devops",
          "server",
          "ggml",
          "SYCL",
          "Apple Metal"
        ],
        "created_at": "2025-03-09T17:52:45Z",
        "updated_at": "2025-03-24T00:01:02Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/pull/12288",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 12329,
        "kind": "pr",
        "state": "open",
        "title": "Fixed Eval Bug: 12163 : Fallback to CPU when loading model: vk::PhysicalDevice::createDevice: ErrorExtensionNotPresent.",
        "labels": [
          "Vulkan",
          "ggml"
        ],
        "created_at": "2025-03-11T08:44:37Z",
        "updated_at": "2025-03-11T09:10:22Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/12329",
        "categories": [
          "native_backends",
          "models_architectures",
          "loader_conversion",
          "correctness_testing"
        ]
      },
      {
        "number": 12333,
        "kind": "pr",
        "state": "open",
        "title": "server : improve infill stop criteria",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2025-03-11T13:44:11Z",
        "updated_at": "2025-03-12T13:29:02Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/12333",
        "categories": [
          "scheduling_serving",
          "api_ui_docs"
        ]
      },
      {
        "number": 12383,
        "kind": "pr",
        "state": "open",
        "title": "[WIP] MUSA: enable fastfp16, correct warp reduce impl and perf tuning",
        "labels": [
          "Nvidia GPU",
          "ggml"
        ],
        "created_at": "2025-03-14T09:36:50Z",
        "updated_at": "2025-03-30T11:41:27Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/12383",
        "categories": [
          "native_backends",
          "models_architectures"
        ]
      },
      {
        "number": 12459,
        "kind": "pr",
        "state": "open",
        "title": "[Issue #12458] Temporarily Clamp inf Values in ggml-cpu.c to Prevent Garbled Output(or coredump) on RK3588",
        "labels": [
          "ggml"
        ],
        "created_at": "2025-03-19T02:14:06Z",
        "updated_at": "2025-03-19T02:16:18Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/12459",
        "categories": [
          "native_backends",
          "models_architectures",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 12476,
        "kind": "issue",
        "state": "open",
        "title": "tts : add support for Orpheus",
        "labels": [
          "good first issue",
          "tts"
        ],
        "created_at": "2025-03-20T08:11:43Z",
        "updated_at": "2026-08-17T16:52:09Z",
        "closed_at": null,
        "comments": 10,
        "url": "https://github.com/ggml-org/llama.cpp/issues/12476",
        "categories": [
          "models_architectures",
          "multimodal_embedding",
          "api_ui_docs"
        ]
      },
      {
        "number": 12485,
        "kind": "pr",
        "state": "open",
        "title": "Metal TQ2_0",
        "labels": [
          "ggml",
          "Apple Metal"
        ],
        "created_at": "2025-03-20T21:51:24Z",
        "updated_at": "2026-07-30T01:20:30Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/12485",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures"
        ]
      },
      {
        "number": 12487,
        "kind": "pr",
        "state": "open",
        "title": "(draft) tts: Orpheus support",
        "labels": [
          "examples",
          "python",
          "ggml"
        ],
        "created_at": "2025-03-21T04:19:48Z",
        "updated_at": "2026-08-17T16:52:10Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/pull/12487",
        "categories": [
          "quantization",
          "models_architectures",
          "loader_conversion",
          "api_ui_docs"
        ]
      },
      {
        "number": 12488,
        "kind": "pr",
        "state": "open",
        "title": "Evenly and stably pinning thread pool",
        "labels": [
          "ggml"
        ],
        "created_at": "2025-03-21T05:20:14Z",
        "updated_at": "2025-03-23T20:32:57Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/12488",
        "categories": [
          "native_backends",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 12552,
        "kind": "pr",
        "state": "open",
        "title": "llama-map to support hugepage feature of pagesize 2M or 1G which can …",
        "labels": [],
        "created_at": "2025-03-24T20:45:49Z",
        "updated_at": "2025-03-31T01:51:17Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/12552",
        "categories": [
          "native_backends",
          "models_architectures",
          "loader_conversion",
          "build_platform_ops"
        ]
      },
      {
        "number": 12557,
        "kind": "pr",
        "state": "open",
        "title": "ggml-quants : weighted rounding algorithms with cumulative search",
        "labels": [
          "generation quality",
          "research 🔬",
          "Less than 4 bits",
          "Review Complexity : Medium",
          "ggml",
          "Tensor Encoding Scheme"
        ],
        "created_at": "2025-03-25T01:39:26Z",
        "updated_at": "2025-08-18T02:04:55Z",
        "closed_at": null,
        "comments": 16,
        "url": "https://github.com/ggml-org/llama.cpp/pull/12557",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "speculative_generation",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 12648,
        "kind": "pr",
        "state": "open",
        "title": "tts : implement sesame CSM + Mimi decoder",
        "labels": [
          "examples",
          "python"
        ],
        "created_at": "2025-03-29T23:36:28Z",
        "updated_at": "2026-07-29T23:40:43Z",
        "closed_at": null,
        "comments": 31,
        "url": "https://github.com/ggml-org/llama.cpp/pull/12648",
        "categories": [
          "quantization",
          "models_architectures",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 12679,
        "kind": "pr",
        "state": "open",
        "title": "WIP: Add support for CogAgent",
        "labels": [
          "examples",
          "python",
          "server"
        ],
        "created_at": "2025-03-31T22:57:21Z",
        "updated_at": "2025-06-24T22:45:00Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/12679",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 12712,
        "kind": "pr",
        "state": "open",
        "title": "Fix: Abnormal exit on Android devices",
        "labels": [
          "ggml"
        ],
        "created_at": "2025-04-02T09:11:49Z",
        "updated_at": "2025-04-02T09:42:59Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/12712",
        "categories": [
          "build_platform_ops"
        ]
      },
      {
        "number": 12727,
        "kind": "pr",
        "state": "open",
        "title": "Update llama-quant.cpp llama_tensor_get_type with DeepSeek friendly modifications",
        "labels": [
          "ggml"
        ],
        "created_at": "2025-04-03T02:48:49Z",
        "updated_at": "2026-05-25T18:18:53Z",
        "closed_at": null,
        "comments": 19,
        "url": "https://github.com/ggml-org/llama.cpp/pull/12727",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 12794,
        "kind": "pr",
        "state": "open",
        "title": "Support for OuteTTS 1.0",
        "labels": [
          "examples",
          "python"
        ],
        "created_at": "2025-04-07T09:54:32Z",
        "updated_at": "2025-05-20T04:04:59Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/12794",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 12835,
        "kind": "pr",
        "state": "open",
        "title": "llamax : add a possible implementation of a simple API for llama.cpp …",
        "labels": [
          "build"
        ],
        "created_at": "2025-04-08T18:07:15Z",
        "updated_at": "2025-04-08T18:07:27Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/12835",
        "categories": [
          "models_architectures",
          "structured_generation_tools",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 12837,
        "kind": "pr",
        "state": "open",
        "title": "convert : write tensors in parallel",
        "labels": [
          "performance",
          "python"
        ],
        "created_at": "2025-04-08T21:04:22Z",
        "updated_at": "2026-04-13T20:23:38Z",
        "closed_at": null,
        "comments": 19,
        "url": "https://github.com/ggml-org/llama.cpp/pull/12837",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 12902,
        "kind": "pr",
        "state": "open",
        "title": "Get CPU model in ggml_backend_cpu_device_context on FreeBSD",
        "labels": [
          "ggml"
        ],
        "created_at": "2025-04-11T17:41:59Z",
        "updated_at": "2025-04-14T21:37:07Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/12902",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context"
        ]
      },
      {
        "number": 12940,
        "kind": "pr",
        "state": "open",
        "title": "set b = ub when b > ub with embedding",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2025-04-14T09:30:54Z",
        "updated_at": "2025-04-15T17:43:34Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/12940",
        "categories": [
          "scheduling_serving",
          "multimodal_embedding",
          "api_ui_docs"
        ]
      },
      {
        "number": 12976,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Vulkan performance depends on thread priority",
        "labels": [
          "bug"
        ],
        "created_at": "2025-04-16T14:13:31Z",
        "updated_at": "2025-06-09T01:08:05Z",
        "closed_at": null,
        "comments": 16,
        "url": "https://github.com/ggml-org/llama.cpp/issues/12976",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 12992,
        "kind": "pr",
        "state": "open",
        "title": "Fix convert script for non-hf GLM4 checkpoints",
        "labels": [
          "python"
        ],
        "created_at": "2025-04-17T05:35:40Z",
        "updated_at": "2025-04-20T04:44:43Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/12992",
        "categories": [
          "quantization",
          "models_architectures",
          "loader_conversion",
          "api_ui_docs"
        ]
      },
      {
        "number": 13005,
        "kind": "pr",
        "state": "open",
        "title": "Nix portability improvements",
        "labels": [
          "nix",
          "devops",
          "ggml"
        ],
        "created_at": "2025-04-18T06:41:52Z",
        "updated_at": "2025-04-25T19:24:39Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/13005",
        "categories": [
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 13026,
        "kind": "pr",
        "state": "open",
        "title": "Bitnet: directly use scale instead of inverting it twice",
        "labels": [
          "python"
        ],
        "created_at": "2025-04-19T16:22:56Z",
        "updated_at": "2025-04-19T16:23:08Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/13026",
        "categories": [
          "quantization",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 13028,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Ability to pack multiple GGUFs into single one",
        "labels": [
          "enhancement",
          "research 🔬"
        ],
        "created_at": "2025-04-19T19:09:58Z",
        "updated_at": "2026-01-27T04:32:36Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/issues/13028",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 13056,
        "kind": "pr",
        "state": "open",
        "title": "Update README.md for tts example to use afplay on MacOS",
        "labels": [
          "examples"
        ],
        "created_at": "2025-04-22T02:29:34Z",
        "updated_at": "2025-04-22T02:29:47Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/13056",
        "categories": [
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 13058,
        "kind": "pr",
        "state": "open",
        "title": "Fix ChatGLMModel for glm-4-9b cannot find tokenizer merges in model file",
        "labels": [
          "python"
        ],
        "created_at": "2025-04-22T03:57:54Z",
        "updated_at": "2025-05-06T10:56:09Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/13058",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "speculative_generation",
          "loader_conversion",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 13079,
        "kind": "pr",
        "state": "open",
        "title": "ggml: Implement yield barrier using futex for improved thread scheduling efficiency",
        "labels": [
          "ggml"
        ],
        "created_at": "2025-04-23T15:53:57Z",
        "updated_at": "2026-08-24T15:05:57Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/pull/13079",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 13106,
        "kind": "pr",
        "state": "open",
        "title": "llama : try loading tensors with pre-computed hashes",
        "labels": [
          "Nvidia GPU",
          "Vulkan",
          "examples",
          "ggml",
          "SYCL",
          "Apple Metal",
          "Kompute"
        ],
        "created_at": "2025-04-25T11:00:12Z",
        "updated_at": "2025-05-25T08:38:09Z",
        "closed_at": null,
        "comments": 29,
        "url": "https://github.com/ggml-org/llama.cpp/pull/13106",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 13206,
        "kind": "pr",
        "state": "open",
        "title": "Introduce New Lookup-Table(LUT)-Based Matrix Multiplication Method (TMAC)",
        "labels": [
          "python",
          "ggml"
        ],
        "created_at": "2025-04-30T10:28:41Z",
        "updated_at": "2025-09-26T11:27:04Z",
        "closed_at": null,
        "comments": 26,
        "url": "https://github.com/ggml-org/llama.cpp/pull/13206",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 13293,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: -TS doesn't support more than ? Devices",
        "labels": [
          "bug"
        ],
        "created_at": "2025-05-04T05:49:56Z",
        "updated_at": "2025-06-04T23:10:24Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/13293",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 13295,
        "kind": "pr",
        "state": "open",
        "title": "Added dynamic context size. This is perfect for servers running llama models as a service.",
        "labels": [],
        "created_at": "2025-05-04T08:53:04Z",
        "updated_at": "2026-03-04T14:04:24Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/pull/13295",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving"
        ]
      },
      {
        "number": 13319,
        "kind": "pr",
        "state": "open",
        "title": "[Perf] [CPU] eliminate redundant memory access in group query attention",
        "labels": [
          "ggml"
        ],
        "created_at": "2025-05-05T17:29:31Z",
        "updated_at": "2025-06-11T01:19:01Z",
        "closed_at": null,
        "comments": 12,
        "url": "https://github.com/ggml-org/llama.cpp/pull/13319",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 13334,
        "kind": "pr",
        "state": "open",
        "title": "add AMD Genoa",
        "labels": [
          "ggml"
        ],
        "created_at": "2025-05-06T07:49:37Z",
        "updated_at": "2025-05-06T07:50:10Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/13334",
        "categories": [
          "models_architectures"
        ]
      },
      {
        "number": 13339,
        "kind": "pr",
        "state": "open",
        "title": "Support  Sp token Function Call Token Implementation",
        "labels": [],
        "created_at": "2025-05-06T10:50:03Z",
        "updated_at": "2025-05-06T12:30:53Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/13339",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 13348,
        "kind": "pr",
        "state": "open",
        "title": "Add mistral-chat-7b preset for llama-server",
        "labels": [
          "examples"
        ],
        "created_at": "2025-05-07T01:48:46Z",
        "updated_at": "2025-05-07T01:48:59Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/13348",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "api_ui_docs"
        ]
      },
      {
        "number": 13352,
        "kind": "pr",
        "state": "open",
        "title": "common: add default reranker presets",
        "labels": [],
        "created_at": "2025-05-07T10:28:12Z",
        "updated_at": "2025-05-09T10:19:46Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/13352",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 13360,
        "kind": "pr",
        "state": "open",
        "title": "CUDA: update build CTK version to 12.8",
        "labels": [
          "Nvidia GPU",
          "devops",
          "ggml"
        ],
        "created_at": "2025-05-07T17:00:58Z",
        "updated_at": "2025-11-29T23:36:09Z",
        "closed_at": null,
        "comments": 12,
        "url": "https://github.com/ggml-org/llama.cpp/pull/13360",
        "categories": [
          "native_backends",
          "build_platform_ops"
        ]
      },
      {
        "number": 13378,
        "kind": "pr",
        "state": "open",
        "title": "gguf-py: Optimize `GGUFReader` read-only mode performance",
        "labels": [
          "python"
        ],
        "created_at": "2025-05-08T07:54:53Z",
        "updated_at": "2025-06-18T22:34:37Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/13378",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 13391,
        "kind": "pr",
        "state": "open",
        "title": "grammar: handle misplaced special regex chars [*+?]",
        "labels": [],
        "created_at": "2025-05-08T20:47:13Z",
        "updated_at": "2025-05-09T11:14:29Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/13391",
        "categories": [
          "structured_generation_tools",
          "build_platform_ops"
        ]
      },
      {
        "number": 13401,
        "kind": "pr",
        "state": "open",
        "title": "Update README.md for using llama.cpp in Microsoft Word locally",
        "labels": [],
        "created_at": "2025-05-09T11:02:27Z",
        "updated_at": "2025-05-20T01:22:52Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/13401",
        "categories": [
          "models_architectures",
          "api_ui_docs"
        ]
      },
      {
        "number": 13404,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: llama-cli, Qwen3 jinja template will break CLI multiturn conversation",
        "labels": [
          "bug"
        ],
        "created_at": "2025-05-09T13:10:48Z",
        "updated_at": "2025-06-29T18:00:26Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/13404",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing"
        ]
      },
      {
        "number": 13462,
        "kind": "pr",
        "state": "open",
        "title": "llama: Add configuration presets for chat and reranking servers",
        "labels": [],
        "created_at": "2025-05-12T00:36:13Z",
        "updated_at": "2025-05-13T17:09:20Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/13462",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding"
        ]
      },
      {
        "number": 13474,
        "kind": "pr",
        "state": "open",
        "title": "docker : enable RPC for docker images",
        "labels": [
          "devops"
        ],
        "created_at": "2025-05-12T08:58:58Z",
        "updated_at": "2026-03-01T03:05:07Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/pull/13474",
        "categories": [
          "scheduling_serving",
          "multimodal_embedding",
          "build_platform_ops"
        ]
      },
      {
        "number": 13485,
        "kind": "issue",
        "state": "open",
        "title": "LoRA training example",
        "labels": [
          "research 🔬"
        ],
        "created_at": "2025-05-12T18:52:04Z",
        "updated_at": "2025-09-06T01:07:50Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/13485",
        "categories": [
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 13523,
        "kind": "issue",
        "state": "open",
        "title": "tutorials : list for llama.cpp",
        "labels": [
          "help wanted",
          "good first issue",
          "roadmap"
        ],
        "created_at": "2025-05-14T05:00:53Z",
        "updated_at": "2026-07-17T01:26:35Z",
        "closed_at": null,
        "comments": 23,
        "url": "https://github.com/ggml-org/llama.cpp/issues/13523",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "api_ui_docs"
        ]
      },
      {
        "number": 13574,
        "kind": "pr",
        "state": "open",
        "title": "Update python verions",
        "labels": [
          "examples",
          "python",
          "server"
        ],
        "created_at": "2025-05-15T20:00:54Z",
        "updated_at": "2025-05-23T17:18:17Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/pull/13574",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 13600,
        "kind": "pr",
        "state": "open",
        "title": "ggml : fix race-condition in ggml-rpc",
        "labels": [
          "ggml"
        ],
        "created_at": "2025-05-17T09:07:05Z",
        "updated_at": "2025-05-25T11:38:09Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/13600",
        "categories": [
          "scheduling_serving"
        ]
      },
      {
        "number": 13601,
        "kind": "pr",
        "state": "open",
        "title": "ggml : add memset_tensor for rpc",
        "labels": [
          "ggml"
        ],
        "created_at": "2025-05-17T09:14:38Z",
        "updated_at": "2026-07-19T10:57:51Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/13601",
        "categories": [
          "models_architectures",
          "scheduling_serving"
        ]
      },
      {
        "number": 13615,
        "kind": "pr",
        "state": "open",
        "title": "scripts: update pyproject.toml - deprecated poetry config + support uv",
        "labels": [],
        "created_at": "2025-05-18T16:42:25Z",
        "updated_at": "2025-07-03T17:41:34Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/13615",
        "categories": [
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 13710,
        "kind": "pr",
        "state": "open",
        "title": "common/llama: align structures for reduce cacheline size on 64bit platforms",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2025-05-22T15:02:52Z",
        "updated_at": "2025-05-22T16:24:14Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/pull/13710",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 13731,
        "kind": "pr",
        "state": "open",
        "title": "Move page cache via mbind to prevent cross-NUMA access",
        "labels": [
          "build"
        ],
        "created_at": "2025-05-23T19:03:33Z",
        "updated_at": "2025-06-30T04:14:02Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/13731",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 13741,
        "kind": "pr",
        "state": "open",
        "title": "cmake : set `RPATH` to `$ORIGIN` on Linux (#13740)",
        "labels": [
          "build"
        ],
        "created_at": "2025-05-24T08:28:02Z",
        "updated_at": "2025-06-25T15:26:08Z",
        "closed_at": null,
        "comments": 12,
        "url": "https://github.com/ggml-org/llama.cpp/pull/13741",
        "categories": [
          "build_platform_ops"
        ]
      },
      {
        "number": 13764,
        "kind": "pr",
        "state": "open",
        "title": "Add comprehensive test for llama_batch/sbatch/ubatch concepts",
        "labels": [
          "testing"
        ],
        "created_at": "2025-05-24T23:32:53Z",
        "updated_at": "2025-05-24T23:33:06Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/13764",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 13799,
        "kind": "pr",
        "state": "open",
        "title": "Add OPT model support - Add OPT architecture support in C++ code - Im…",
        "labels": [
          "python"
        ],
        "created_at": "2025-05-26T11:36:55Z",
        "updated_at": "2025-11-10T13:16:50Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/13799",
        "categories": [
          "quantization",
          "models_architectures",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 13821,
        "kind": "pr",
        "state": "open",
        "title": "Tokenize logging",
        "labels": [
          "examples"
        ],
        "created_at": "2025-05-27T07:46:52Z",
        "updated_at": "2025-05-27T08:09:25Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/13821",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 13828,
        "kind": "pr",
        "state": "open",
        "title": "examples : support MiniCPM-V-2",
        "labels": [
          "examples",
          "python"
        ],
        "created_at": "2025-05-27T14:20:04Z",
        "updated_at": "2026-06-29T09:53:34Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/13828",
        "categories": [
          "models_architectures",
          "multimodal_embedding",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 13987,
        "kind": "pr",
        "state": "open",
        "title": "chore(server): split context-server to its own file",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2025-06-03T09:53:53Z",
        "updated_at": "2025-06-03T13:33:41Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/13987",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "api_ui_docs"
        ]
      },
      {
        "number": 14032,
        "kind": "pr",
        "state": "open",
        "title": "cpu: Update RISC-V condition to require GCC version 14 or higher",
        "labels": [
          "ggml"
        ],
        "created_at": "2025-06-05T14:25:55Z",
        "updated_at": "2025-08-16T06:04:11Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/14032",
        "categories": [
          "scheduling_serving"
        ]
      },
      {
        "number": 14117,
        "kind": "pr",
        "state": "open",
        "title": "ggml: aarch64: Implement SVE Kernels for Int 8 Quantization",
        "labels": [
          "ggml"
        ],
        "created_at": "2025-06-11T04:28:05Z",
        "updated_at": "2025-08-18T05:54:00Z",
        "closed_at": null,
        "comments": 11,
        "url": "https://github.com/ggml-org/llama.cpp/pull/14117",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 14139,
        "kind": "pr",
        "state": "open",
        "title": "tests : add test-model-random",
        "labels": [
          "help wanted",
          "testing"
        ],
        "created_at": "2025-06-12T04:58:44Z",
        "updated_at": "2025-07-08T22:24:19Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/14139",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 14148,
        "kind": "pr",
        "state": "open",
        "title": "models/templates: add mistralai/Mistral-Small-3.1-24B-Instruct-2503 template with tool calling support",
        "labels": [],
        "created_at": "2025-06-12T11:05:14Z",
        "updated_at": "2025-06-13T09:39:23Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/14148",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 14229,
        "kind": "pr",
        "state": "open",
        "title": "logit_bias: apply configurable escalating EOG bias at low n_remain",
        "labels": [
          "testing",
          "examples",
          "server"
        ],
        "created_at": "2025-06-16T23:32:16Z",
        "updated_at": "2025-10-23T04:54:47Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/14229",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 14232,
        "kind": "pr",
        "state": "open",
        "title": "ggml: introduce GGML_NUMA_MIGRATE to optimize cross NUMA op computation",
        "labels": [
          "examples",
          "ggml"
        ],
        "created_at": "2025-06-17T06:31:57Z",
        "updated_at": "2026-04-19T06:45:20Z",
        "closed_at": null,
        "comments": 16,
        "url": "https://github.com/ggml-org/llama.cpp/pull/14232",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 14358,
        "kind": "pr",
        "state": "open",
        "title": "llama : expose C API to get layer device type",
        "labels": [],
        "created_at": "2025-06-24T09:01:56Z",
        "updated_at": "2025-09-07T19:52:41Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/14358",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 14380,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cpu: Build variant targeting Neoverse-V2",
        "labels": [
          "ggml"
        ],
        "created_at": "2025-06-25T19:14:19Z",
        "updated_at": "2026-07-13T09:51:09Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/14380",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 14413,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Unexpected empty grammar stack after accepting piece: <unused32>",
        "labels": [
          "bug"
        ],
        "created_at": "2025-06-27T07:35:28Z",
        "updated_at": "2026-02-16T06:18:40Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/issues/14413",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 14463,
        "kind": "pr",
        "state": "open",
        "title": "Chore: batch prompts, extract tensors specific layer",
        "labels": [
          "examples"
        ],
        "created_at": "2025-06-30T11:38:37Z",
        "updated_at": "2025-09-07T09:06:28Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/14463",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 14464,
        "kind": "issue",
        "state": "open",
        "title": "Compile bug: zero-size array ‘gemm_gemv_kernels’ / invalid feature modifier ‘sme’",
        "labels": [
          "bug"
        ],
        "created_at": "2025-06-30T13:57:00Z",
        "updated_at": "2025-11-21T09:00:10Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/14464",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 14484,
        "kind": "pr",
        "state": "open",
        "title": "Compute buffer and KV-cache aware layer distribution for multi-GPU inference",
        "labels": [],
        "created_at": "2025-07-01T16:38:24Z",
        "updated_at": "2026-02-03T21:52:37Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/14484",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 14493,
        "kind": "pr",
        "state": "open",
        "title": "Allow truncation when embedding",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2025-07-02T04:30:59Z",
        "updated_at": "2025-07-18T11:29:02Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/14493",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "api_ui_docs"
        ]
      },
      {
        "number": 14522,
        "kind": "pr",
        "state": "open",
        "title": "train: add simple loading already tokenized data from parquet dataset",
        "labels": [
          "build",
          "examples"
        ],
        "created_at": "2025-07-03T20:17:48Z",
        "updated_at": "2026-08-13T23:38:53Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/pull/14522",
        "categories": [
          "scheduling_serving",
          "loader_conversion",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 14532,
        "kind": "pr",
        "state": "open",
        "title": "common: detect and prefer big cores on AArch64 hybrid CPU on linux",
        "labels": [],
        "created_at": "2025-07-04T12:42:38Z",
        "updated_at": "2025-08-18T00:23:16Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/14532",
        "categories": [
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 14570,
        "kind": "pr",
        "state": "open",
        "title": "metal : reuse graphs",
        "labels": [
          "demo",
          "ggml",
          "Apple Metal"
        ],
        "created_at": "2025-07-07T18:45:00Z",
        "updated_at": "2025-07-17T16:08:38Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/14570",
        "categories": [
          "native_backends",
          "models_architectures"
        ]
      },
      {
        "number": 14622,
        "kind": "pr",
        "state": "open",
        "title": "tool: add convertation of text/parquet to custom format",
        "labels": [
          "build",
          "examples"
        ],
        "created_at": "2025-07-10T17:31:55Z",
        "updated_at": "2026-08-13T23:39:17Z",
        "closed_at": null,
        "comments": 20,
        "url": "https://github.com/ggml-org/llama.cpp/pull/14622",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 14635,
        "kind": "pr",
        "state": "open",
        "title": "OpenCL: add `mul_mat_f16_f32_image` kernel ",
        "labels": [
          "ggml",
          "OpenCL"
        ],
        "created_at": "2025-07-11T11:07:07Z",
        "updated_at": "2025-07-15T11:37:27Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/14635",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 14639,
        "kind": "pr",
        "state": "open",
        "title": "Add CUDA non-contiguous Unary Ops support",
        "labels": [
          "documentation",
          "build",
          "testing",
          "Nvidia GPU",
          "ggml"
        ],
        "created_at": "2025-07-11T23:33:14Z",
        "updated_at": "2025-07-31T12:41:48Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/14639",
        "categories": [
          "native_backends",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 14702,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: OpenAI API v1/responses llama-server",
        "labels": [
          "enhancement",
          "roadmap"
        ],
        "created_at": "2025-07-15T21:10:07Z",
        "updated_at": "2026-07-16T19:17:36Z",
        "closed_at": null,
        "comments": 41,
        "url": "https://github.com/ggml-org/llama.cpp/issues/14702",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 14728,
        "kind": "pr",
        "state": "open",
        "title": "server: add prompt processing progress streaming for /completion endpoint #14685",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2025-07-16T19:59:03Z",
        "updated_at": "2025-07-16T21:36:06Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/14728",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "api_ui_docs"
        ]
      },
      {
        "number": 14739,
        "kind": "pr",
        "state": "open",
        "title": "examples : predicted output for text generation",
        "labels": [
          "examples"
        ],
        "created_at": "2025-07-17T15:32:14Z",
        "updated_at": "2025-07-24T10:51:21Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/14739",
        "categories": [
          "models_architectures",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 14765,
        "kind": "pr",
        "state": "open",
        "title": "feat: Add extended sampling API with candidate token lists #14612",
        "labels": [],
        "created_at": "2025-07-19T00:26:15Z",
        "updated_at": "2025-07-19T00:26:15Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/14765",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 14766,
        "kind": "pr",
        "state": "open",
        "title": "docs : mention apt installation method",
        "labels": [
          "documentation"
        ],
        "created_at": "2025-07-19T05:45:47Z",
        "updated_at": "2025-08-23T09:27:55Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/14766",
        "categories": [
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 14811,
        "kind": "pr",
        "state": "open",
        "title": "feat(batched): Add functionality to upload benchmark test results",
        "labels": [
          "examples"
        ],
        "created_at": "2025-07-22T09:17:56Z",
        "updated_at": "2025-07-22T09:19:00Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/14811",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 14833,
        "kind": "pr",
        "state": "open",
        "title": "test-backend-ops: enables perf/eval testing of composite ops",
        "labels": [
          "testing",
          "ggml"
        ],
        "created_at": "2025-07-23T13:03:01Z",
        "updated_at": "2025-07-24T15:38:31Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/14833",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 14891,
        "kind": "pr",
        "state": "open",
        "title": "imatrix: calculate activation-based statistics for new format (GGUF) imatrices",
        "labels": [
          "documentation",
          "examples"
        ],
        "created_at": "2025-07-26T16:47:29Z",
        "updated_at": "2026-08-23T11:51:05Z",
        "closed_at": null,
        "comments": 30,
        "url": "https://github.com/ggml-org/llama.cpp/pull/14891",
        "categories": [
          "quantization",
          "models_architectures",
          "structured_generation_tools",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 14908,
        "kind": "pr",
        "state": "open",
        "title": "opencl: fixed a typo",
        "labels": [
          "ggml",
          "OpenCL"
        ],
        "created_at": "2025-07-27T21:14:18Z",
        "updated_at": "2025-07-27T21:14:28Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/14908",
        "categories": [
          "native_backends"
        ]
      },
      {
        "number": 14909,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Implement missing ops from backends",
        "labels": [
          "enhancement",
          "good first issue"
        ],
        "created_at": "2025-07-28T03:39:11Z",
        "updated_at": "2026-08-17T02:38:01Z",
        "closed_at": null,
        "comments": 51,
        "url": "https://github.com/ggml-org/llama.cpp/issues/14909",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 14918,
        "kind": "pr",
        "state": "open",
        "title": "repack : optimize mul_mat_id path",
        "labels": [
          "ggml"
        ],
        "created_at": "2025-07-28T12:29:53Z",
        "updated_at": "2025-09-29T09:57:24Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/14918",
        "categories": [
          "native_backends",
          "models_architectures"
        ]
      },
      {
        "number": 14970,
        "kind": "pr",
        "state": "open",
        "title": "Optimize l2_norm_f32 op with SIMD",
        "labels": [
          "ggml"
        ],
        "created_at": "2025-07-30T15:41:48Z",
        "updated_at": "2025-08-05T18:21:33Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/14970",
        "categories": [
          "native_backends",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 15060,
        "kind": "pr",
        "state": "open",
        "title": "quantize : configurable neutral imatrix prior",
        "labels": [
          "generation quality",
          "research 🔬",
          "need feedback",
          "examples"
        ],
        "created_at": "2025-08-03T21:49:06Z",
        "updated_at": "2026-03-12T17:25:45Z",
        "closed_at": null,
        "comments": 11,
        "url": "https://github.com/ggml-org/llama.cpp/pull/15060",
        "categories": [
          "quantization",
          "models_architectures",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 15104,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: seq_rm does not support seq_id < 0, documentation claiming the value is supported",
        "labels": [
          "bug"
        ],
        "created_at": "2025-08-05T22:14:30Z",
        "updated_at": "2025-08-11T23:41:50Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/15104",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 15123,
        "kind": "pr",
        "state": "open",
        "title": "Add T5Gemma support #14940",
        "labels": [
          "python"
        ],
        "created_at": "2025-08-06T13:52:58Z",
        "updated_at": "2025-12-26T03:36:33Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/15123",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 15165,
        "kind": "pr",
        "state": "open",
        "title": "MoE Expert manipulation args",
        "labels": [
          "demo"
        ],
        "created_at": "2025-08-08T04:23:50Z",
        "updated_at": "2025-08-11T08:25:16Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/15165",
        "categories": [
          "native_backends",
          "models_architectures",
          "build_platform_ops"
        ]
      },
      {
        "number": 15189,
        "kind": "pr",
        "state": "open",
        "title": "server: implementation of v1/completions echo logprobs support",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2025-08-09T02:04:25Z",
        "updated_at": "2026-03-01T12:16:02Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/pull/15189",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "api_ui_docs"
        ]
      },
      {
        "number": 15217,
        "kind": "pr",
        "state": "open",
        "title": "Adding Resume for curl downloads",
        "labels": [
          "testing"
        ],
        "created_at": "2025-08-10T18:37:54Z",
        "updated_at": "2025-08-10T18:38:08Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/15217",
        "categories": [
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 15248,
        "kind": "pr",
        "state": "open",
        "title": "Fixes #15247 | Update chat.cpp to support (at least) qwen3 reasoning + tool_choice = required",
        "labels": [
          "testing",
          "examples",
          "python",
          "server"
        ],
        "created_at": "2025-08-11T16:25:21Z",
        "updated_at": "2025-11-04T16:20:20Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/pull/15248",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 15269,
        "kind": "issue",
        "state": "open",
        "title": "Research: Intel oneAPI Compiler",
        "labels": [
          "research 🔬"
        ],
        "created_at": "2025-08-12T15:55:49Z",
        "updated_at": "2025-08-13T01:35:02Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/15269",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 15298,
        "kind": "pr",
        "state": "open",
        "title": "64 bit CUDA copy routines via GGML_CUDA_ALLOW_LARGE_TENSORS",
        "labels": [
          "Nvidia GPU",
          "ggml"
        ],
        "created_at": "2025-08-13T18:24:53Z",
        "updated_at": "2025-11-02T14:21:20Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/15298",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 15364,
        "kind": "issue",
        "state": "open",
        "title": "Research: RekaQuant, better quantization with existing quant primitives though better error reduction than imatrix",
        "labels": [
          "research 🔬"
        ],
        "created_at": "2025-08-17T07:44:04Z",
        "updated_at": "2025-08-17T07:44:04Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/15364",
        "categories": [
          "quantization",
          "models_architectures",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 15499,
        "kind": "pr",
        "state": "open",
        "title": "llama : add ggml version and commit functions",
        "labels": [
          "ggml"
        ],
        "created_at": "2025-08-22T06:57:12Z",
        "updated_at": "2025-08-28T12:48:54Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/15499",
        "categories": [
          "models_architectures"
        ]
      },
      {
        "number": 15515,
        "kind": "pr",
        "state": "open",
        "title": "Added updated check for multi modal projector and vision projector in…",
        "labels": [
          "python"
        ],
        "created_at": "2025-08-22T23:39:46Z",
        "updated_at": "2025-08-23T07:18:00Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/15515",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "multimodal_embedding",
          "loader_conversion",
          "api_ui_docs"
        ]
      },
      {
        "number": 15517,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: No generation with follow up on high token responses on GPT-OSS 120B",
        "labels": [
          "bug",
          "Nvidia GPU",
          "AMD GPU"
        ],
        "created_at": "2025-08-23T01:31:19Z",
        "updated_at": "2025-11-11T10:32:59Z",
        "closed_at": null,
        "comments": 20,
        "url": "https://github.com/ggml-org/llama.cpp/issues/15517",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 15550,
        "kind": "pr",
        "state": "open",
        "title": "quantize: add option to automatically choose optimal quant types to reach a file/bpw target size at lowest error",
        "labels": [
          "documentation",
          "examples"
        ],
        "created_at": "2025-08-24T21:44:58Z",
        "updated_at": "2026-08-23T07:48:05Z",
        "closed_at": null,
        "comments": 14,
        "url": "https://github.com/ggml-org/llama.cpp/pull/15550",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 15582,
        "kind": "pr",
        "state": "open",
        "title": "fix(ggml-sycl): add synchronization before exiting argsort kernel",
        "labels": [
          "ggml",
          "SYCL"
        ],
        "created_at": "2025-08-26T04:21:51Z",
        "updated_at": "2026-05-20T06:13:26Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/pull/15582",
        "categories": [
          "native_backends"
        ]
      },
      {
        "number": 15601,
        "kind": "pr",
        "state": "open",
        "title": "Partial code documentation",
        "labels": [
          "documentation"
        ],
        "created_at": "2025-08-26T20:59:55Z",
        "updated_at": "2025-09-27T05:24:04Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/15601",
        "categories": [
          "models_architectures",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 15602,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Repeated Unecessary Activation Quantization Ops",
        "labels": [
          "enhancement",
          "research 🔬"
        ],
        "created_at": "2025-08-26T21:20:58Z",
        "updated_at": "2025-11-07T14:20:56Z",
        "closed_at": null,
        "comments": 15,
        "url": "https://github.com/ggml-org/llama.cpp/issues/15602",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 15626,
        "kind": "pr",
        "state": "open",
        "title": "Possible fix: use ne0..ne3 (dst dims) instead of ne00..ne03 in ggml_compute_forward_dup_f16",
        "labels": [
          "ggml"
        ],
        "created_at": "2025-08-28T00:10:19Z",
        "updated_at": "2025-08-28T07:53:13Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/15626",
        "categories": [
          "quantization",
          "memory_kv_context"
        ]
      },
      {
        "number": 15636,
        "kind": "pr",
        "state": "open",
        "title": "batch : add `pad_equal` [RFC]",
        "labels": [],
        "created_at": "2025-08-28T11:36:17Z",
        "updated_at": "2025-08-28T11:44:39Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/15636",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 15643,
        "kind": "pr",
        "state": "open",
        "title": "tools: update llama-bench to include TTFT, E2E, ITL metrics",
        "labels": [
          "examples"
        ],
        "created_at": "2025-08-28T17:51:15Z",
        "updated_at": "2025-09-04T23:20:04Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/15643",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "api_ui_docs"
        ]
      },
      {
        "number": 15648,
        "kind": "pr",
        "state": "open",
        "title": "gguf-py: reduce peak RAM during convert by streaming dtype casts",
        "labels": [
          "python"
        ],
        "created_at": "2025-08-28T21:20:08Z",
        "updated_at": "2025-09-03T07:43:56Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/pull/15648",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 15719,
        "kind": "pr",
        "state": "open",
        "title": "ggml : block repack support for Q4_K quanti for AArch64 architecture",
        "labels": [
          "ggml"
        ],
        "created_at": "2025-09-01T16:46:40Z",
        "updated_at": "2026-01-21T17:12:34Z",
        "closed_at": null,
        "comments": 9,
        "url": "https://github.com/ggml-org/llama.cpp/pull/15719",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 15727,
        "kind": "pr",
        "state": "open",
        "title": "convert : use reflinks for faster conversion",
        "labels": [
          "demo",
          "python",
          "ggml"
        ],
        "created_at": "2025-09-02T02:51:52Z",
        "updated_at": "2025-11-07T04:03:30Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/15727",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 15738,
        "kind": "pr",
        "state": "open",
        "title": "Add scale_diag_mask_inf_softmax operation for transformer attention",
        "labels": [
          "testing",
          "ggml"
        ],
        "created_at": "2025-09-02T10:31:20Z",
        "updated_at": "2025-09-19T11:54:12Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/pull/15738",
        "categories": [
          "models_architectures",
          "correctness_testing"
        ]
      },
      {
        "number": 15747,
        "kind": "pr",
        "state": "open",
        "title": "nix: Added missing packages and options for ROCm build",
        "labels": [
          "nix",
          "devops",
          "ggml"
        ],
        "created_at": "2025-09-02T15:34:58Z",
        "updated_at": "2025-10-02T14:58:30Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/15747",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 15749,
        "kind": "issue",
        "state": "open",
        "title": "Consolidation of tensor copies to backend to reduce API overhead",
        "labels": [
          "enhancement",
          "research 🔬"
        ],
        "created_at": "2025-09-02T16:02:57Z",
        "updated_at": "2025-11-20T17:03:24Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/15749",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 15774,
        "kind": "pr",
        "state": "open",
        "title": "example: add prediction-next-token command line argument handling Example for show probability of next token",
        "labels": [
          "examples"
        ],
        "created_at": "2025-09-03T17:35:38Z",
        "updated_at": "2025-09-04T06:49:50Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/15774",
        "categories": [
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 15800,
        "kind": "pr",
        "state": "open",
        "title": "vulkan: add mul_mat variant for embedded gpus ",
        "labels": [
          "Vulkan",
          "ggml"
        ],
        "created_at": "2025-09-04T16:25:36Z",
        "updated_at": "2026-02-05T19:04:07Z",
        "closed_at": null,
        "comments": 27,
        "url": "https://github.com/ggml-org/llama.cpp/pull/15800",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing"
        ]
      },
      {
        "number": 15805,
        "kind": "pr",
        "state": "open",
        "title": "cuda : Add conv2d Implicit GEMM",
        "labels": [
          "testing",
          "Nvidia GPU",
          "ggml"
        ],
        "created_at": "2025-09-04T20:16:52Z",
        "updated_at": "2026-03-11T11:58:50Z",
        "closed_at": null,
        "comments": 64,
        "url": "https://github.com/ggml-org/llama.cpp/pull/15805",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 15822,
        "kind": "pr",
        "state": "open",
        "title": "Improve user-facing warnings for chat template and context length",
        "labels": [
          "examples"
        ],
        "created_at": "2025-09-05T19:37:56Z",
        "updated_at": "2025-09-05T20:09:35Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/15822",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 15837,
        "kind": "pr",
        "state": "open",
        "title": "llguidance : use attrs to determine special tokens",
        "labels": [],
        "created_at": "2025-09-06T13:34:58Z",
        "updated_at": "2025-09-10T00:41:44Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/15837",
        "categories": [
          "build_platform_ops"
        ]
      },
      {
        "number": 15845,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: HIP gfx908 (MI100) cublass error when prompt is too long.",
        "labels": [
          "bug",
          "AMD GPU"
        ],
        "created_at": "2025-09-06T23:04:05Z",
        "updated_at": "2025-12-14T04:17:28Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/15845",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 15885,
        "kind": "pr",
        "state": "open",
        "title": "Update Mirostat sampler function parameters",
        "labels": [],
        "created_at": "2025-09-09T01:10:19Z",
        "updated_at": "2025-09-09T01:10:41Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/15885",
        "categories": [
          "structured_generation_tools"
        ]
      },
      {
        "number": 15996,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Vulkan backend hangs forever with NVIDIA GPU on FreeBSD",
        "labels": [
          "bug",
          "Vulkan"
        ],
        "created_at": "2025-09-14T20:01:44Z",
        "updated_at": "2025-11-21T04:45:51Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/issues/15996",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 15998,
        "kind": "pr",
        "state": "open",
        "title": "server: added support for systemd socket activation and readiness reporting",
        "labels": [
          "build",
          "examples",
          "server"
        ],
        "created_at": "2025-09-14T21:42:52Z",
        "updated_at": "2026-05-04T19:55:06Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/15998",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 16000,
        "kind": "pr",
        "state": "open",
        "title": "`--numa mirror`: mirror model weights to every Numa node in the system",
        "labels": [
          "testing",
          "Nvidia GPU",
          "Vulkan",
          "examples",
          "python",
          "devops",
          "ggml",
          "SYCL",
          "Apple Metal",
          "Ascend NPU",
          "OpenCL",
          "IBM zDNN"
        ],
        "created_at": "2025-09-15T06:13:51Z",
        "updated_at": "2026-08-19T08:16:27Z",
        "closed_at": null,
        "comments": 91,
        "url": "https://github.com/ggml-org/llama.cpp/pull/16000",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 16004,
        "kind": "pr",
        "state": "open",
        "title": "ggml : add CPU backend reference implementation (wip)",
        "labels": [
          "testing",
          "ggml"
        ],
        "created_at": "2025-09-15T09:08:39Z",
        "updated_at": "2026-03-02T06:32:04Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/16004",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 16014,
        "kind": "pr",
        "state": "open",
        "title": "Guard ThreadPowerThrottling for non-MSVC builds",
        "labels": [
          "ggml"
        ],
        "created_at": "2025-09-15T17:14:41Z",
        "updated_at": "2025-10-22T20:35:19Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/pull/16014",
        "categories": [
          "models_architectures",
          "build_platform_ops"
        ]
      },
      {
        "number": 16016,
        "kind": "pr",
        "state": "open",
        "title": "Deterministic inference mode (CUDA): RMSNorm, MatMul, Attention, KV-cache",
        "labels": [
          "documentation",
          "script",
          "testing",
          "Nvidia GPU",
          "ggml"
        ],
        "created_at": "2025-09-15T18:27:38Z",
        "updated_at": "2025-09-15T19:12:51Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/16016",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 16020,
        "kind": "pr",
        "state": "open",
        "title": "[WIP] Rpc split row",
        "labels": [
          "examples",
          "ggml",
          "Apple Metal"
        ],
        "created_at": "2025-09-16T03:22:50Z",
        "updated_at": "2025-11-10T18:24:24Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/pull/16020",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 16045,
        "kind": "pr",
        "state": "open",
        "title": "docker : added explicit arm64 build (#11888 and #13891)",
        "labels": [
          "devops"
        ],
        "created_at": "2025-09-17T09:05:28Z",
        "updated_at": "2025-11-22T03:58:20Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/16045",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "build_platform_ops"
        ]
      },
      {
        "number": 16088,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cuda: Vulkan direct conv 2D ported to CUDA",
        "labels": [
          "Nvidia GPU",
          "ggml"
        ],
        "created_at": "2025-09-18T15:12:52Z",
        "updated_at": "2025-10-28T20:53:20Z",
        "closed_at": null,
        "comments": 38,
        "url": "https://github.com/ggml-org/llama.cpp/pull/16088",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 16134,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Vulkan backend shows negative scaling at low batch sizes with MOE models",
        "labels": [
          "research 🔬",
          "Vulkan"
        ],
        "created_at": "2025-09-20T19:23:28Z",
        "updated_at": "2025-11-09T08:58:34Z",
        "closed_at": null,
        "comments": 13,
        "url": "https://github.com/ggml-org/llama.cpp/issues/16134",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 16143,
        "kind": "pr",
        "state": "open",
        "title": "[metal] Add fused RMS_NORM + MUL + SWIGLU for Qwen3Next",
        "labels": [
          "testing",
          "ggml",
          "Apple Metal"
        ],
        "created_at": "2025-09-21T10:06:28Z",
        "updated_at": "2025-09-27T13:58:19Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/16143",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 16147,
        "kind": "pr",
        "state": "open",
        "title": "split out non-HTTP server parts",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2025-09-21T14:49:40Z",
        "updated_at": "2025-09-30T12:52:57Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/16147",
        "categories": [
          "scheduling_serving",
          "api_ui_docs"
        ]
      },
      {
        "number": 16234,
        "kind": "pr",
        "state": "open",
        "title": "Extend CI for i8mm kernels as well",
        "labels": [
          "testing"
        ],
        "created_at": "2025-09-24T19:57:04Z",
        "updated_at": "2026-08-26T08:42:47Z",
        "closed_at": null,
        "comments": 10,
        "url": "https://github.com/ggml-org/llama.cpp/pull/16234",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 16260,
        "kind": "pr",
        "state": "open",
        "title": "Correct XTC threshold args documentation",
        "labels": [],
        "created_at": "2025-09-25T15:55:32Z",
        "updated_at": "2025-09-25T15:55:33Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/16260",
        "categories": [
          "models_architectures",
          "api_ui_docs"
        ]
      },
      {
        "number": 16270,
        "kind": "pr",
        "state": "open",
        "title": "Support FP16 as intermediate results in graph computation",
        "labels": [
          "ggml"
        ],
        "created_at": "2025-09-26T06:14:43Z",
        "updated_at": "2025-09-26T06:23:03Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/16270",
        "categories": [
          "native_backends",
          "models_architectures",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 16280,
        "kind": "pr",
        "state": "open",
        "title": "Update convert_hf_to_gguf_update.py",
        "labels": [
          "python"
        ],
        "created_at": "2025-09-26T14:04:29Z",
        "updated_at": "2025-09-27T08:31:11Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/16280",
        "categories": [
          "quantization",
          "models_architectures",
          "loader_conversion",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 16301,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Critical Crash and Performance Degradation in Vulkan Build since Release B6524",
        "labels": [
          "bug",
          "Vulkan"
        ],
        "created_at": "2025-09-28T10:52:56Z",
        "updated_at": "2026-06-21T10:31:39Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/issues/16301",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 16393,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: tool to list and delete cached models",
        "labels": [
          "enhancement"
        ],
        "created_at": "2025-10-02T21:00:20Z",
        "updated_at": "2026-08-05T01:11:09Z",
        "closed_at": null,
        "comments": 26,
        "url": "https://github.com/ggml-org/llama.cpp/issues/16393",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 16404,
        "kind": "pr",
        "state": "open",
        "title": "contrib : add fish completions via --completion-fish",
        "labels": [],
        "created_at": "2025-10-03T06:46:04Z",
        "updated_at": "2026-06-30T23:49:35Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/16404",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 16457,
        "kind": "pr",
        "state": "open",
        "title": "Add hipblasLt implementation for batched gemm to improve performance for CDNA3 only",
        "labels": [
          "Nvidia GPU",
          "ggml"
        ],
        "created_at": "2025-10-07T07:27:19Z",
        "updated_at": "2025-10-08T18:24:05Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/16457",
        "categories": [
          "native_backends",
          "scheduling_serving"
        ]
      },
      {
        "number": 16458,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: batched endpoints for tokenization",
        "labels": [
          "enhancement",
          "roadmap"
        ],
        "created_at": "2025-10-07T08:55:12Z",
        "updated_at": "2025-10-09T07:38:22Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/16458",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 16475,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: [Mac] LoRA inference crashes with GGML_ASSERT graph size error (works on Windows)",
        "labels": [
          "bug"
        ],
        "created_at": "2025-10-08T19:14:07Z",
        "updated_at": "2026-01-24T14:42:47Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/16475",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 16530,
        "kind": "pr",
        "state": "open",
        "title": "metal: add support for LOG op (f32, f16)",
        "labels": [
          "ggml",
          "Apple Metal"
        ],
        "created_at": "2025-10-12T09:33:50Z",
        "updated_at": "2026-07-17T08:07:55Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/16530",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 16534,
        "kind": "pr",
        "state": "open",
        "title": "server: add /slots/status endpoint for secure monitoring",
        "labels": [
          "examples",
          "python",
          "server"
        ],
        "created_at": "2025-10-12T15:59:17Z",
        "updated_at": "2025-10-12T20:46:56Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/16534",
        "categories": [
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 16535,
        "kind": "pr",
        "state": "open",
        "title": "Update close-issue.yml",
        "labels": [
          "devops"
        ],
        "created_at": "2025-10-12T16:31:45Z",
        "updated_at": "2025-10-12T16:31:56Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/16535",
        "categories": [
          "models_architectures",
          "build_platform_ops"
        ]
      },
      {
        "number": 16544,
        "kind": "pr",
        "state": "open",
        "title": "tests: increase NMSE threshold for q5_1 MUL_MAT tests",
        "labels": [
          "testing"
        ],
        "created_at": "2025-10-12T20:16:16Z",
        "updated_at": "2025-10-12T20:27:48Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/16544",
        "categories": [
          "native_backends",
          "quantization",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 16548,
        "kind": "pr",
        "state": "open",
        "title": "Implement and use cuda graph plans",
        "labels": [
          "Nvidia GPU",
          "ggml"
        ],
        "created_at": "2025-10-13T01:26:43Z",
        "updated_at": "2025-12-02T07:18:09Z",
        "closed_at": null,
        "comments": 21,
        "url": "https://github.com/ggml-org/llama.cpp/pull/16548",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 16561,
        "kind": "pr",
        "state": "open",
        "title": "embeddings: Fix --log-disable should not suppress embedding outputs",
        "labels": [
          "examples"
        ],
        "created_at": "2025-10-13T11:39:21Z",
        "updated_at": "2025-10-13T20:31:45Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/pull/16561",
        "categories": [
          "models_architectures",
          "multimodal_embedding",
          "api_ui_docs"
        ]
      },
      {
        "number": 16574,
        "kind": "pr",
        "state": "open",
        "title": "mtmd: Add JinaCLIP v2 vision projector + GGUF support for jina-bert-v3 (merged-LoRA or adapter)",
        "labels": [
          "examples",
          "python",
          "ggml"
        ],
        "created_at": "2025-10-14T09:04:34Z",
        "updated_at": "2026-02-09T10:55:04Z",
        "closed_at": null,
        "comments": 11,
        "url": "https://github.com/ggml-org/llama.cpp/pull/16574",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 16650,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cpu: optimise rms_norm op",
        "labels": [
          "testing",
          "ggml"
        ],
        "created_at": "2025-10-18T15:38:10Z",
        "updated_at": "2025-11-11T16:18:52Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/16650",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 16669,
        "kind": "pr",
        "state": "open",
        "title": "metal: add ops DIAG_MASK_INF, IM2COL_3D, fix op PAD",
        "labels": [
          "ggml",
          "Apple Metal"
        ],
        "created_at": "2025-10-19T21:04:53Z",
        "updated_at": "2026-08-25T02:45:24Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/16669",
        "categories": [
          "native_backends",
          "memory_kv_context",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing"
        ]
      },
      {
        "number": 16692,
        "kind": "pr",
        "state": "open",
        "title": "llama-context: fix build fails with `-Werror=missing-braces`",
        "labels": [],
        "created_at": "2025-10-21T00:06:44Z",
        "updated_at": "2025-10-21T00:06:45Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/16692",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 16698,
        "kind": "pr",
        "state": "open",
        "title": "fix[readme]: Update docs/build.md to match the new GPU_TARGETS",
        "labels": [
          "documentation"
        ],
        "created_at": "2025-10-21T09:14:49Z",
        "updated_at": "2025-10-21T09:47:11Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/16698",
        "categories": [
          "native_backends",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 16743,
        "kind": "pr",
        "state": "open",
        "title": "get_rows & dequantize function implementation for repacked weights of type q6_K (q6_Kx8) ",
        "labels": [
          "ggml"
        ],
        "created_at": "2025-10-23T14:40:14Z",
        "updated_at": "2026-03-17T13:20:05Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/pull/16743",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 16753,
        "kind": "pr",
        "state": "open",
        "title": "rpc: use `XXHash64` instead of `FNV-1a` for hashing tensors",
        "labels": [
          "ggml"
        ],
        "created_at": "2025-10-24T10:18:31Z",
        "updated_at": "2026-02-23T21:07:12Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/16753",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 16773,
        "kind": "pr",
        "state": "open",
        "title": "Adding CUDA release for Ubuntu",
        "labels": [
          "devops"
        ],
        "created_at": "2025-10-26T05:23:11Z",
        "updated_at": "2025-10-26T07:02:32Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/16773",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 16817,
        "kind": "pr",
        "state": "open",
        "title": "Implement SparseK Attention mechanism — new GGML operator with CPU backend (GPU planned next)",
        "labels": [
          "testing",
          "python",
          "ggml"
        ],
        "created_at": "2025-10-28T13:16:40Z",
        "updated_at": "2025-11-24T09:31:25Z",
        "closed_at": null,
        "comments": 41,
        "url": "https://github.com/ggml-org/llama.cpp/pull/16817",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 16844,
        "kind": "pr",
        "state": "open",
        "title": "Enable CUDA graphs for embed gemma 300m",
        "labels": [
          "Nvidia GPU",
          "ggml"
        ],
        "created_at": "2025-10-29T14:04:34Z",
        "updated_at": "2025-10-31T14:05:23Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/16844",
        "categories": [
          "native_backends",
          "models_architectures",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 16874,
        "kind": "pr",
        "state": "open",
        "title": "server: add support for local image path loading for server",
        "labels": [
          "examples",
          "python",
          "server"
        ],
        "created_at": "2025-10-30T20:33:29Z",
        "updated_at": "2025-11-06T22:19:49Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/16874",
        "categories": [
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 16882,
        "kind": "pr",
        "state": "open",
        "title": "Disable NUMA-specific chunking for high-core-count HPC systems",
        "labels": [
          "ggml"
        ],
        "created_at": "2025-10-31T04:45:46Z",
        "updated_at": "2026-05-13T13:40:42Z",
        "closed_at": null,
        "comments": 10,
        "url": "https://github.com/ggml-org/llama.cpp/pull/16882",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 16915,
        "kind": "pr",
        "state": "open",
        "title": "add TheRock HIP backend build instructions",
        "labels": [
          "documentation"
        ],
        "created_at": "2025-11-01T07:06:05Z",
        "updated_at": "2026-02-07T05:44:14Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/16915",
        "categories": [
          "native_backends",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 16926,
        "kind": "pr",
        "state": "open",
        "title": "Add initial devcontainer configuration",
        "labels": [],
        "created_at": "2025-11-01T21:33:24Z",
        "updated_at": "2025-11-01T21:33:24Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/16926",
        "categories": [
          "models_architectures"
        ]
      },
      {
        "number": 16939,
        "kind": "pr",
        "state": "open",
        "title": "doc: Windows + clang/ninja build guide format cleanup",
        "labels": [
          "documentation"
        ],
        "created_at": "2025-11-02T14:36:29Z",
        "updated_at": "2026-05-23T03:25:02Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/16939",
        "categories": [
          "multimodal_embedding",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 16940,
        "kind": "pr",
        "state": "open",
        "title": "Add e2e tests for embedding raw flag",
        "labels": [
          "testing",
          "examples",
          "python",
          "devops"
        ],
        "created_at": "2025-11-02T16:00:13Z",
        "updated_at": "2026-08-21T17:42:31Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/16940",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 16948,
        "kind": "pr",
        "state": "open",
        "title": "CUDA: add implicit conv3d ",
        "labels": [
          "testing",
          "Nvidia GPU",
          "ggml"
        ],
        "created_at": "2025-11-02T17:38:04Z",
        "updated_at": "2025-12-13T18:21:46Z",
        "closed_at": null,
        "comments": 31,
        "url": "https://github.com/ggml-org/llama.cpp/pull/16948",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 16965,
        "kind": "pr",
        "state": "open",
        "title": "Refactor llm_chat_template_from_str to avoid throwing exceptions",
        "labels": [],
        "created_at": "2025-11-03T11:49:02Z",
        "updated_at": "2025-11-03T11:49:03Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/16965",
        "categories": [
          "other"
        ]
      },
      {
        "number": 16969,
        "kind": "pr",
        "state": "open",
        "title": "sycl: flash-attention implementation",
        "labels": [
          "testing",
          "ggml",
          "SYCL"
        ],
        "created_at": "2025-11-03T13:11:27Z",
        "updated_at": "2026-05-20T06:12:24Z",
        "closed_at": null,
        "comments": 13,
        "url": "https://github.com/ggml-org/llama.cpp/pull/16969",
        "categories": [
          "native_backends",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 17034,
        "kind": "pr",
        "state": "open",
        "title": "common: \"Profile Guided Speculative Decoding\"",
        "labels": [],
        "created_at": "2025-11-05T18:46:52Z",
        "updated_at": "2025-11-05T19:39:16Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/17034",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 17044,
        "kind": "pr",
        "state": "open",
        "title": "Add MoE dynamic routing with expert caching",
        "labels": [
          "documentation",
          "build",
          "examples"
        ],
        "created_at": "2025-11-06T05:11:48Z",
        "updated_at": "2025-11-06T05:12:13Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/17044",
        "categories": [
          "models_architectures",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 17068,
        "kind": "pr",
        "state": "open",
        "title": "Fix NetBSD compilation error",
        "labels": [],
        "created_at": "2025-11-07T02:01:56Z",
        "updated_at": "2025-11-07T02:01:57Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/17068",
        "categories": [
          "build_platform_ops"
        ]
      },
      {
        "number": 17096,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Crash at second prompt",
        "labels": [
          "bug"
        ],
        "created_at": "2025-11-08T05:05:55Z",
        "updated_at": "2025-12-15T15:02:19Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/17096",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 17098,
        "kind": "pr",
        "state": "open",
        "title": "Add Metal-4 Tensor API test harness for iOS",
        "labels": [
          "examples"
        ],
        "created_at": "2025-11-08T08:13:49Z",
        "updated_at": "2025-11-08T08:44:10Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/17098",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 17113,
        "kind": "pr",
        "state": "open",
        "title": "CPU SIMD and pipeline optimizations across vec/mmq/ops/kv-cache/repack ",
        "labels": [
          "ggml"
        ],
        "created_at": "2025-11-08T21:27:22Z",
        "updated_at": "2025-11-10T20:40:28Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/17113",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 17141,
        "kind": "pr",
        "state": "open",
        "title": "Add complete Megrez-MoE support: GGUF conversion + inference.",
        "labels": [
          "model",
          "python"
        ],
        "created_at": "2025-11-10T10:14:50Z",
        "updated_at": "2026-01-18T13:47:22Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/17141",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 17147,
        "kind": "pr",
        "state": "open",
        "title": "vulkan: add q2_K implementation in mul_mmq with ACC_TYPE_VEC2",
        "labels": [
          "Vulkan",
          "ggml"
        ],
        "created_at": "2025-11-10T14:33:08Z",
        "updated_at": "2026-03-25T11:15:54Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/17147",
        "categories": [
          "native_backends",
          "quantization"
        ]
      },
      {
        "number": 17255,
        "kind": "pr",
        "state": "open",
        "title": "CUDA: add CONV_3D operator support",
        "labels": [
          "documentation",
          "Nvidia GPU",
          "ggml"
        ],
        "created_at": "2025-11-14T01:04:35Z",
        "updated_at": "2025-11-16T12:24:15Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/17255",
        "categories": [
          "native_backends",
          "quantization",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 17257,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cpu : prevent kleidiai build/install targets",
        "labels": [
          "ggml"
        ],
        "created_at": "2025-11-14T05:39:33Z",
        "updated_at": "2025-12-22T10:59:34Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/17257",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 17330,
        "kind": "pr",
        "state": "open",
        "title": "Fix grammar README: Update non-terminal naming rules",
        "labels": [],
        "created_at": "2025-11-17T21:26:47Z",
        "updated_at": "2026-04-10T03:47:03Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/17330",
        "categories": [
          "structured_generation_tools",
          "api_ui_docs"
        ]
      },
      {
        "number": 17342,
        "kind": "pr",
        "state": "open",
        "title": "Throughput improvement for small batch sizes",
        "labels": [
          "ggml"
        ],
        "created_at": "2025-11-18T01:09:15Z",
        "updated_at": "2026-03-24T17:05:52Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/pull/17342",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 17374,
        "kind": "pr",
        "state": "open",
        "title": "vulkan: Change default subgroup/block size on selected pipelines for Intel Xe2 and newer GPU",
        "labels": [
          "testing",
          "Vulkan",
          "ggml"
        ],
        "created_at": "2025-11-19T01:06:28Z",
        "updated_at": "2026-07-15T07:39:35Z",
        "closed_at": null,
        "comments": 27,
        "url": "https://github.com/ggml-org/llama.cpp/pull/17374",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 17383,
        "kind": "pr",
        "state": "open",
        "title": "ggml : enhance rel-pos and window ops with CUDA support",
        "labels": [
          "testing",
          "Nvidia GPU",
          "ggml"
        ],
        "created_at": "2025-11-19T14:16:00Z",
        "updated_at": "2025-12-03T19:28:22Z",
        "closed_at": null,
        "comments": 9,
        "url": "https://github.com/ggml-org/llama.cpp/pull/17383",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing"
        ]
      },
      {
        "number": 17387,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: llama server response hangs for /slots/0?action=erase",
        "labels": [
          "bug",
          "help wanted",
          "server/api",
          "medium severity"
        ],
        "created_at": "2025-11-19T16:46:39Z",
        "updated_at": "2025-11-26T07:34:24Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/issues/17387",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 17398,
        "kind": "pr",
        "state": "open",
        "title": "models : add Nougat OCR support with mBART and Swin Transformer",
        "labels": [
          "model",
          "examples",
          "python"
        ],
        "created_at": "2025-11-20T08:43:47Z",
        "updated_at": "2025-11-20T08:45:41Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/17398",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 17417,
        "kind": "pr",
        "state": "open",
        "title": "Vulkan: Add `GGML_OP_GET_REL_POS`",
        "labels": [
          "Vulkan",
          "ggml"
        ],
        "created_at": "2025-11-20T23:13:19Z",
        "updated_at": "2025-12-02T00:52:00Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/17417",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 17422,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: llama-server crashes during 'encoding image slice...'",
        "labels": [
          "bug",
          "model"
        ],
        "created_at": "2025-11-21T06:44:29Z",
        "updated_at": "2026-02-06T12:59:15Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/issues/17422",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 17432,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: AMD driver 25.11.1 crashes with llamacpp built with Vulkan SDK 1.4.328.1",
        "labels": [
          "bug",
          "Vulkan"
        ],
        "created_at": "2025-11-21T16:03:03Z",
        "updated_at": "2026-01-19T00:53:47Z",
        "closed_at": null,
        "comments": 10,
        "url": "https://github.com/ggml-org/llama.cpp/issues/17432",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 17436,
        "kind": "pr",
        "state": "open",
        "title": "docs: vulkan add GGML_VK_ALLOW_SYSMEM_FALLBACK=1 docs",
        "labels": [
          "documentation"
        ],
        "created_at": "2025-11-22T05:56:49Z",
        "updated_at": "2025-11-23T12:19:35Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/17436",
        "categories": [
          "native_backends",
          "memory_kv_context",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 17438,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Vulkan issues starting with 439342ea0be347ff279ec204719794df3b3108f6",
        "labels": [
          "bug",
          "testing",
          "Vulkan"
        ],
        "created_at": "2025-11-22T11:44:33Z",
        "updated_at": "2026-08-06T11:40:21Z",
        "closed_at": null,
        "comments": 13,
        "url": "https://github.com/ggml-org/llama.cpp/issues/17438",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 17454,
        "kind": "pr",
        "state": "open",
        "title": "model : add LLADA 2.0 diffusion support ",
        "labels": [
          "model",
          "examples",
          "python"
        ],
        "created_at": "2025-11-23T19:21:19Z",
        "updated_at": "2026-07-22T18:11:14Z",
        "closed_at": null,
        "comments": 16,
        "url": "https://github.com/ggml-org/llama.cpp/pull/17454",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 17459,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Can't use longer context than model via RoPE due to server-imposed restrictions",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2025-11-24T00:12:16Z",
        "updated_at": "2026-08-07T06:59:17Z",
        "closed_at": null,
        "comments": 24,
        "url": "https://github.com/ggml-org/llama.cpp/issues/17459",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "training_adapters"
        ]
      },
      {
        "number": 17460,
        "kind": "pr",
        "state": "open",
        "title": "Typo in json.gbnf",
        "labels": [],
        "created_at": "2025-11-24T03:05:42Z",
        "updated_at": "2025-11-24T18:09:03Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/pull/17460",
        "categories": [
          "other"
        ]
      },
      {
        "number": 17483,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cpu:add RISC-V Zvfh implementation for ggml_vec_dot_f16_unroll",
        "labels": [
          "ggml"
        ],
        "created_at": "2025-11-25T03:19:11Z",
        "updated_at": "2025-11-27T13:32:53Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/17483",
        "categories": [
          "quantization"
        ]
      },
      {
        "number": 17485,
        "kind": "pr",
        "state": "open",
        "title": "vulkan : add dynamic VRAM heuristic for low-VRAM GPUs",
        "labels": [
          "documentation",
          "testing",
          "Vulkan",
          "examples",
          "ggml"
        ],
        "created_at": "2025-11-25T05:17:44Z",
        "updated_at": "2025-12-13T07:50:36Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/17485",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 17488,
        "kind": "issue",
        "state": "open",
        "title": "GGUF convert support for Vibevoice",
        "labels": [
          "enhancement",
          "help wanted",
          "good first issue",
          "model",
          "3rd party",
          "tts"
        ],
        "created_at": "2025-11-25T07:29:13Z",
        "updated_at": "2026-07-24T05:11:10Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/issues/17488",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion"
        ]
      },
      {
        "number": 17495,
        "kind": "pr",
        "state": "open",
        "title": "HIP: Add RDNA3 WMMA support to MMF",
        "labels": [
          "Nvidia GPU",
          "ggml"
        ],
        "created_at": "2025-11-25T14:35:21Z",
        "updated_at": "2025-12-03T03:07:15Z",
        "closed_at": null,
        "comments": 22,
        "url": "https://github.com/ggml-org/llama.cpp/pull/17495",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 17510,
        "kind": "pr",
        "state": "open",
        "title": "docs: build llama.cpp using MSVC - Visual Studio 17 2022",
        "labels": [
          "documentation"
        ],
        "created_at": "2025-11-26T00:45:42Z",
        "updated_at": "2025-11-29T08:11:38Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/17510",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 17511,
        "kind": "pr",
        "state": "open",
        "title": "llama: remove init f16 tables",
        "labels": [],
        "created_at": "2025-11-26T02:34:19Z",
        "updated_at": "2025-11-26T02:42:34Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/17511",
        "categories": [
          "quantization",
          "models_architectures"
        ]
      },
      {
        "number": 17529,
        "kind": "pr",
        "state": "open",
        "title": "llama.cpp with sentencepiece",
        "labels": [
          "testing"
        ],
        "created_at": "2025-11-26T18:40:32Z",
        "updated_at": "2025-11-26T19:08:39Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/17529",
        "categories": [
          "models_architectures",
          "correctness_testing"
        ]
      },
      {
        "number": 17534,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Qwen 2.5 VL 3B Instruct infinite ? output for certain image dimensions",
        "labels": [
          "bug",
          "model",
          "medium severity",
          "mtmd"
        ],
        "created_at": "2025-11-26T23:07:52Z",
        "updated_at": "2025-12-25T15:19:08Z",
        "closed_at": null,
        "comments": 17,
        "url": "https://github.com/ggml-org/llama.cpp/issues/17534",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 17546,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: model run with `cpu-moe` fail",
        "labels": [
          "bug",
          "Vulkan"
        ],
        "created_at": "2025-11-27T12:04:48Z",
        "updated_at": "2025-12-04T16:46:40Z",
        "closed_at": null,
        "comments": 13,
        "url": "https://github.com/ggml-org/llama.cpp/issues/17546",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 17583,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Running llama-server only possible with single AMD GPU, running multiple always causes Segmentation fault regardless of model size",
        "labels": [
          "bug",
          "AMD GPU",
          "RoCM"
        ],
        "created_at": "2025-11-28T22:07:58Z",
        "updated_at": "2025-12-14T00:32:26Z",
        "closed_at": null,
        "comments": 16,
        "url": "https://github.com/ggml-org/llama.cpp/issues/17583",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 17611,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: wrong default threads in llama-bench",
        "labels": [
          "good first issue",
          "performance",
          "low severity"
        ],
        "created_at": "2025-11-30T07:12:34Z",
        "updated_at": "2026-06-22T15:08:35Z",
        "closed_at": null,
        "comments": 11,
        "url": "https://github.com/ggml-org/llama.cpp/issues/17611",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing"
        ]
      },
      {
        "number": 17647,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: rpc-server crashes with a gpf",
        "labels": [
          "bug",
          "RPC"
        ],
        "created_at": "2025-12-01T10:57:45Z",
        "updated_at": "2026-01-04T02:50:55Z",
        "closed_at": null,
        "comments": 13,
        "url": "https://github.com/ggml-org/llama.cpp/issues/17647",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 17654,
        "kind": "issue",
        "state": "open",
        "title": "Refactor: Deprecate DIAG_MASK_* operators, migrate kernels to TRI",
        "labels": [
          "breaking change",
          "refactoring"
        ],
        "created_at": "2025-12-01T11:48:28Z",
        "updated_at": "2026-08-26T12:11:11Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/17654",
        "categories": [
          "native_backends"
        ]
      },
      {
        "number": 17687,
        "kind": "pr",
        "state": "open",
        "title": "Add Support for Microsoft Phi-3.5 Vision Instruct Models",
        "labels": [
          "documentation",
          "build",
          "script",
          "testing",
          "Nvidia GPU",
          "Vulkan",
          "examples",
          "python",
          "devops",
          "server",
          "ggml",
          "Apple Metal",
          "Ascend NPU"
        ],
        "created_at": "2025-12-02T11:15:03Z",
        "updated_at": "2025-12-04T08:04:52Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/17687",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 17706,
        "kind": "pr",
        "state": "open",
        "title": "CANN: Support fusion operator that supports mul and add",
        "labels": [
          "testing",
          "ggml",
          "Ascend NPU"
        ],
        "created_at": "2025-12-03T01:55:00Z",
        "updated_at": "2025-12-03T03:41:56Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/17706",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 17709,
        "kind": "pr",
        "state": "open",
        "title": "build: for GGML_BACKEND_DL, ggml need not depend on backend",
        "labels": [
          "ggml"
        ],
        "created_at": "2025-12-03T02:30:39Z",
        "updated_at": "2025-12-04T15:23:56Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/17709",
        "categories": [
          "native_backends",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 17745,
        "kind": "pr",
        "state": "open",
        "title": "Remove deprecated trigger_words support from grammar sampler",
        "labels": [],
        "created_at": "2025-12-03T19:11:18Z",
        "updated_at": "2025-12-03T19:11:18Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/17745",
        "categories": [
          "models_architectures",
          "structured_generation_tools",
          "api_ui_docs"
        ]
      },
      {
        "number": 17759,
        "kind": "pr",
        "state": "open",
        "title": "Fix too stringent check on CUDA \"fast copy\" (can_be_transposed) condition and extend with one more case",
        "labels": [
          "testing",
          "Nvidia GPU",
          "ggml"
        ],
        "created_at": "2025-12-04T13:52:46Z",
        "updated_at": "2026-04-12T16:01:10Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/17759",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing"
        ]
      },
      {
        "number": 17776,
        "kind": "pr",
        "state": "open",
        "title": "Add link to AshkanYarmoradi/go-llama.cpp",
        "labels": [],
        "created_at": "2025-12-05T00:51:01Z",
        "updated_at": "2025-12-05T00:51:02Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/17776",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 17791,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cpu: add repack GEMM and GEMV for floating-point",
        "labels": [
          "ggml"
        ],
        "created_at": "2025-12-05T11:22:55Z",
        "updated_at": "2026-07-29T12:26:29Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/pull/17791",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 17798,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Add support for Multilple Responses in WebUI",
        "labels": [
          "enhancement"
        ],
        "created_at": "2025-12-05T16:57:05Z",
        "updated_at": "2026-08-21T18:28:22Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/issues/17798",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 17802,
        "kind": "pr",
        "state": "open",
        "title": "(CUDA-only) Efficient inference using llama-mtmd-cli for high resolution images with reduced GPU VRAM usage (#17801)",
        "labels": [
          "examples"
        ],
        "created_at": "2025-12-05T19:27:21Z",
        "updated_at": "2026-01-21T07:53:56Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/17802",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "multimodal_embedding",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 17822,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Qwen 3 Next CUDA poor performance",
        "labels": [
          "bug",
          "performance",
          "CUDA"
        ],
        "created_at": "2025-12-06T12:55:18Z",
        "updated_at": "2026-03-23T22:19:38Z",
        "closed_at": null,
        "comments": 23,
        "url": "https://github.com/ggml-org/llama.cpp/issues/17822",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 17832,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: visual <think> tag issues with Ring-mini-2.0 GGUF",
        "labels": [
          "bug",
          "chat parser"
        ],
        "created_at": "2025-12-06T20:56:53Z",
        "updated_at": "2025-12-12T23:31:23Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/issues/17832",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 17837,
        "kind": "pr",
        "state": "open",
        "title": "debug:Adding CPU-side visual trace for hexagon",
        "labels": [
          "script",
          "ggml"
        ],
        "created_at": "2025-12-07T02:38:21Z",
        "updated_at": "2025-12-12T12:00:00Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/17837",
        "categories": [
          "native_backends",
          "correctness_testing"
        ]
      },
      {
        "number": 17840,
        "kind": "pr",
        "state": "open",
        "title": "Add support for R-4B multimodal model",
        "labels": [
          "examples",
          "python"
        ],
        "created_at": "2025-12-07T05:28:37Z",
        "updated_at": "2025-12-07T05:28:49Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/17840",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 17861,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: support the  multi-modal Bagel model",
        "labels": [
          "enhancement",
          "help wanted",
          "model",
          "Review Complexity : High",
          "mtmd"
        ],
        "created_at": "2025-12-08T10:33:25Z",
        "updated_at": "2025-12-08T16:31:56Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/17861",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 17899,
        "kind": "pr",
        "state": "open",
        "title": "ggml: correct inaccurate comments for GGML_OP_MUL_MAT backward pass [no ci]",
        "labels": [
          "ggml"
        ],
        "created_at": "2025-12-10T02:35:31Z",
        "updated_at": "2025-12-10T02:40:26Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/17899",
        "categories": [
          "build_platform_ops"
        ]
      },
      {
        "number": 17907,
        "kind": "pr",
        "state": "open",
        "title": "[WIP]gml-hexagon: Q4_0 mm opt",
        "labels": [
          "ggml"
        ],
        "created_at": "2025-12-10T13:31:05Z",
        "updated_at": "2026-01-07T13:38:36Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/pull/17907",
        "categories": [
          "native_backends",
          "quantization",
          "scheduling_serving",
          "loader_conversion"
        ]
      },
      {
        "number": 17956,
        "kind": "pr",
        "state": "open",
        "title": "server: add encoder-decoder model support (T5, BART, MADLAD)",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2025-12-12T11:05:31Z",
        "updated_at": "2026-02-24T16:52:03Z",
        "closed_at": null,
        "comments": 9,
        "url": "https://github.com/ggml-org/llama.cpp/pull/17956",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "api_ui_docs"
        ]
      },
      {
        "number": 17995,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Since b7101 model output by bin-win-vulkan-x64 with quantized model and cache-type-k and cache-type-v is flawed",
        "labels": [
          "bug",
          "Vulkan",
          "regression"
        ],
        "created_at": "2025-12-13T13:37:48Z",
        "updated_at": "2025-12-15T10:57:51Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/17995",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 18001,
        "kind": "pr",
        "state": "open",
        "title": "server: add /v1/metrics endpoint",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2025-12-13T16:48:35Z",
        "updated_at": "2025-12-14T11:44:57Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18001",
        "categories": [
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 18015,
        "kind": "pr",
        "state": "open",
        "title": "Add message for pre-RDNA AMD GPU support via opencl",
        "labels": [
          "documentation"
        ],
        "created_at": "2025-12-14T04:48:33Z",
        "updated_at": "2026-01-30T16:46:00Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18015",
        "categories": [
          "native_backends",
          "models_architectures",
          "api_ui_docs"
        ]
      },
      {
        "number": 18029,
        "kind": "pr",
        "state": "open",
        "title": "feat: add --moe-n-expert flag for MoE expert count override",
        "labels": [
          "model"
        ],
        "created_at": "2025-12-14T12:42:33Z",
        "updated_at": "2026-04-27T11:53:11Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18029",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 18044,
        "kind": "pr",
        "state": "open",
        "title": "chat-parser: handle whitespace around JSON in tool call parsing",
        "labels": [
          "testing"
        ],
        "created_at": "2025-12-15T01:13:30Z",
        "updated_at": "2025-12-15T01:14:25Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18044",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing"
        ]
      },
      {
        "number": 18046,
        "kind": "issue",
        "state": "open",
        "title": "Compile bug: BUILD_TOOLS now requires BUILD_SERVER",
        "labels": [
          "bug",
          "build"
        ],
        "created_at": "2025-12-15T01:51:58Z",
        "updated_at": "2025-12-15T10:49:51Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/18046",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 18053,
        "kind": "pr",
        "state": "open",
        "title": "CLI: llama-cli and llama-completion cosmetics",
        "labels": [
          "documentation",
          "script",
          "python",
          "devops",
          "SYCL"
        ],
        "created_at": "2025-12-15T10:27:36Z",
        "updated_at": "2026-05-09T01:16:56Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18053",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 18068,
        "kind": "pr",
        "state": "open",
        "title": "webui: add responsive chat width option to webui (#18067)",
        "labels": [
          "server/webui",
          "examples",
          "server"
        ],
        "created_at": "2025-12-15T21:20:59Z",
        "updated_at": "2026-05-12T17:09:58Z",
        "closed_at": null,
        "comments": 35,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18068",
        "categories": [
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 18102,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cuda: Delta-Net linear attention for Qwen3-Next",
        "labels": [
          "model",
          "testing",
          "Nvidia GPU",
          "ggml"
        ],
        "created_at": "2025-12-16T15:41:58Z",
        "updated_at": "2026-04-18T13:05:37Z",
        "closed_at": null,
        "comments": 16,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18102",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing"
        ]
      },
      {
        "number": 18113,
        "kind": "pr",
        "state": "open",
        "title": "ggml-hexagon: Add lightweight atomic synchronization support to htp_ops_context for inter-task coordination",
        "labels": [
          "ggml"
        ],
        "created_at": "2025-12-16T21:37:38Z",
        "updated_at": "2026-01-05T06:31:49Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18113",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 18123,
        "kind": "pr",
        "state": "open",
        "title": "server: validate n_batch == n_ubatch for embeddings (#6263)",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2025-12-17T04:27:26Z",
        "updated_at": "2025-12-19T20:50:56Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18123",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 18126,
        "kind": "pr",
        "state": "open",
        "title": "[WIP] Enable cooperative matrix support for Intel Arrow Lake H GPUs",
        "labels": [
          "Vulkan",
          "ggml"
        ],
        "created_at": "2025-12-17T07:12:21Z",
        "updated_at": "2026-01-15T22:14:23Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18126",
        "categories": [
          "native_backends",
          "build_platform_ops"
        ]
      },
      {
        "number": 18133,
        "kind": "issue",
        "state": "open",
        "title": "Error running on 910b",
        "labels": [
          "bug",
          "Ascend NPU"
        ],
        "created_at": "2025-12-17T10:23:06Z",
        "updated_at": "2026-07-13T02:38:06Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/18133",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 18138,
        "kind": "pr",
        "state": "open",
        "title": "[WIP] Reduce the number of fa rows for Intel",
        "labels": [
          "Vulkan",
          "ggml"
        ],
        "created_at": "2025-12-17T12:43:25Z",
        "updated_at": "2025-12-18T15:05:41Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18138",
        "categories": [
          "native_backends"
        ]
      },
      {
        "number": 18150,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cpu: fix todo comment #15953 and SIMD-like calculate 4 elems",
        "labels": [
          "ggml"
        ],
        "created_at": "2025-12-17T22:41:20Z",
        "updated_at": "2025-12-18T04:47:40Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18150",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 18170,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Llama-server.exe in router mode errors and bails on processing image task, no error when in non-router mode",
        "labels": [
          "bug",
          "server/api",
          "server",
          "mtmd"
        ],
        "created_at": "2025-12-18T13:06:55Z",
        "updated_at": "2026-06-11T19:57:45Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/issues/18170",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 18183,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Hermes 2 Pro parser accepts unwrapped JSON as tool call",
        "labels": [
          "bug",
          "chat parser"
        ],
        "created_at": "2025-12-18T19:56:40Z",
        "updated_at": "2025-12-19T12:51:42Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/issues/18183",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 18191,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: graph reuse consistently fails with GLM-4.6V-Flash-Q4_K_M.gguf",
        "labels": [
          "bug"
        ],
        "created_at": "2025-12-19T05:29:04Z",
        "updated_at": "2025-12-19T08:09:47Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/18191",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 18200,
        "kind": "pr",
        "state": "open",
        "title": "CLI: implemented non interactive mode",
        "labels": [
          "examples"
        ],
        "created_at": "2025-12-19T12:05:58Z",
        "updated_at": "2025-12-22T08:25:00Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18200",
        "categories": [
          "api_ui_docs"
        ]
      },
      {
        "number": 18275,
        "kind": "pr",
        "state": "open",
        "title": "KYLIN: fix compile error for cuda backend",
        "labels": [
          "Nvidia GPU",
          "ggml"
        ],
        "created_at": "2025-12-22T03:06:24Z",
        "updated_at": "2025-12-22T07:54:07Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18275",
        "categories": [
          "native_backends",
          "models_architectures",
          "build_platform_ops"
        ]
      },
      {
        "number": 18307,
        "kind": "pr",
        "state": "open",
        "title": "utils: beging using log.h in tokenize.cpp",
        "labels": [
          "examples"
        ],
        "created_at": "2025-12-22T23:59:52Z",
        "updated_at": "2025-12-23T00:53:21Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18307",
        "categories": [
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 18322,
        "kind": "pr",
        "state": "open",
        "title": "server: (preset) add `unsafe-allow-api-override`",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2025-12-23T11:05:33Z",
        "updated_at": "2026-06-18T14:26:58Z",
        "closed_at": null,
        "comments": 17,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18322",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 18348,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cpu : add riscv vec dot kernel dispatch based on vlen",
        "labels": [
          "ggml"
        ],
        "created_at": "2025-12-24T15:46:39Z",
        "updated_at": "2026-01-19T08:06:08Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18348",
        "categories": [
          "native_backends"
        ]
      },
      {
        "number": 18353,
        "kind": "pr",
        "state": "open",
        "title": "[WIP] tool-call: experimental migration of all parsers to peg-parser infra (w/ better test coverage)",
        "labels": [
          "documentation",
          "script",
          "testing",
          "examples",
          "python",
          "server"
        ],
        "created_at": "2025-12-24T18:30:32Z",
        "updated_at": "2026-02-13T00:53:04Z",
        "closed_at": null,
        "comments": 11,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18353",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 18359,
        "kind": "pr",
        "state": "open",
        "title": "NLLB-600 language translation implementation",
        "labels": [
          "model",
          "python"
        ],
        "created_at": "2025-12-25T02:40:43Z",
        "updated_at": "2025-12-30T22:14:39Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18359",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 18373,
        "kind": "pr",
        "state": "open",
        "title": "feat: Add model pinning feature to protect critical models from LRU eviction",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2025-12-25T19:43:33Z",
        "updated_at": "2025-12-26T16:41:54Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18373",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "api_ui_docs"
        ]
      },
      {
        "number": 18401,
        "kind": "pr",
        "state": "open",
        "title": "ggml: add ggml_rope_comp",
        "labels": [
          "testing",
          "examples",
          "server",
          "ggml",
          "Apple Metal"
        ],
        "created_at": "2025-12-26T23:37:57Z",
        "updated_at": "2025-12-27T14:10:52Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18401",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 18432,
        "kind": "pr",
        "state": "open",
        "title": "android: fix infinite generation in shift_context()",
        "labels": [
          "android",
          "examples"
        ],
        "created_at": "2025-12-28T18:16:16Z",
        "updated_at": "2026-04-10T03:46:40Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18432",
        "categories": [
          "memory_kv_context",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 18472,
        "kind": "pr",
        "state": "open",
        "title": "CUDA: fix 0.0f/0.0f for FA fixup",
        "labels": [
          "Nvidia GPU",
          "ggml"
        ],
        "created_at": "2025-12-29T23:26:03Z",
        "updated_at": "2025-12-30T22:48:21Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18472",
        "categories": [
          "native_backends",
          "models_architectures",
          "build_platform_ops"
        ]
      },
      {
        "number": 18476,
        "kind": "pr",
        "state": "open",
        "title": "android: Fix infinite generation by updating stop_generation_position and implement attention sinks",
        "labels": [
          "android",
          "examples"
        ],
        "created_at": "2025-12-30T03:33:24Z",
        "updated_at": "2025-12-30T08:21:54Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18476",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 18480,
        "kind": "pr",
        "state": "open",
        "title": "Add initial devcontainer configuration",
        "labels": [],
        "created_at": "2025-12-30T06:51:21Z",
        "updated_at": "2025-12-30T06:51:21Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18480",
        "categories": [
          "models_architectures"
        ]
      },
      {
        "number": 18495,
        "kind": "pr",
        "state": "open",
        "title": "ggml: optimized runtime for x86 cpu backend and Q4_K quantized weights paired with Q8_K activations ",
        "labels": [
          "ggml"
        ],
        "created_at": "2025-12-30T16:48:29Z",
        "updated_at": "2026-01-12T12:59:46Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18495",
        "categories": [
          "native_backends",
          "quantization",
          "correctness_testing"
        ]
      },
      {
        "number": 18499,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: llama-finetune won't work even with 17M parameters arch:llama",
        "labels": [
          "bug",
          "help wanted",
          "Review Complexity : High"
        ],
        "created_at": "2025-12-30T23:03:49Z",
        "updated_at": "2026-02-16T22:49:14Z",
        "closed_at": null,
        "comments": 11,
        "url": "https://github.com/ggml-org/llama.cpp/issues/18499",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 18517,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Support IQuest-Coder-V1-40B",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-01-01T07:43:57Z",
        "updated_at": "2026-08-07T07:57:39Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/issues/18517",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 18528,
        "kind": "pr",
        "state": "open",
        "title": "llama-graph: fix UB warning from cross-attn mask loop indices",
        "labels": [],
        "created_at": "2026-01-01T21:09:04Z",
        "updated_at": "2026-01-05T18:38:10Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18528",
        "categories": [
          "models_architectures",
          "build_platform_ops"
        ]
      },
      {
        "number": 18538,
        "kind": "pr",
        "state": "open",
        "title": "CUDA: cache intermediate tensors",
        "labels": [
          "Nvidia GPU",
          "ggml"
        ],
        "created_at": "2026-01-02T07:19:03Z",
        "updated_at": "2026-01-07T07:59:57Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18538",
        "categories": [
          "native_backends",
          "quantization",
          "memory_kv_context"
        ]
      },
      {
        "number": 18574,
        "kind": "pr",
        "state": "open",
        "title": "GGML RPC - Add support for Unix Domain Sockets",
        "labels": [
          "examples",
          "ggml"
        ],
        "created_at": "2026-01-03T22:44:25Z",
        "updated_at": "2026-02-21T17:38:38Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18574",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 18602,
        "kind": "pr",
        "state": "open",
        "title": "common: build as shared library when BUILD_SHARED_LIBS is ON",
        "labels": [],
        "created_at": "2026-01-05T00:11:24Z",
        "updated_at": "2026-01-05T00:11:25Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18602",
        "categories": [
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 18612,
        "kind": "pr",
        "state": "open",
        "title": "sampling: add tail-free (TFS) sampling",
        "labels": [],
        "created_at": "2026-01-05T06:45:04Z",
        "updated_at": "2026-01-06T20:30:13Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18612",
        "categories": [
          "native_backends",
          "quantization",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing"
        ]
      },
      {
        "number": 18633,
        "kind": "pr",
        "state": "open",
        "title": "Added note for compiling on integrated GPUs",
        "labels": [
          "documentation"
        ],
        "created_at": "2026-01-06T04:58:39Z",
        "updated_at": "2026-01-07T23:32:48Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18633",
        "categories": [
          "native_backends",
          "models_architectures",
          "api_ui_docs"
        ]
      },
      {
        "number": 18641,
        "kind": "pr",
        "state": "open",
        "title": "[Do Not Merge] model : LFM2.5-Audio-1.5B",
        "labels": [
          "documentation",
          "model",
          "testing",
          "examples",
          "python",
          "server",
          "ggml",
          "mtmd",
          "conversion"
        ],
        "created_at": "2026-01-06T14:25:07Z",
        "updated_at": "2026-07-16T08:01:10Z",
        "closed_at": null,
        "comments": 41,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18641",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 18650,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: HipVMM is broken on rocm 7.1",
        "labels": [
          "bug",
          "AMD GPU"
        ],
        "created_at": "2026-01-06T22:13:12Z",
        "updated_at": "2026-01-06T22:13:33Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/18650",
        "categories": [
          "native_backends",
          "correctness_testing"
        ]
      },
      {
        "number": 18698,
        "kind": "pr",
        "state": "open",
        "title": "Improving inference speed for the repack buffer type on NUMA architectures",
        "labels": [
          "examples",
          "ggml"
        ],
        "created_at": "2026-01-08T15:01:37Z",
        "updated_at": "2026-01-18T07:11:19Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18698",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 18701,
        "kind": "pr",
        "state": "open",
        "title": "[WIP] ggml-opencl: op args init refactoring",
        "labels": [
          "python",
          "ggml",
          "OpenCL"
        ],
        "created_at": "2026-01-08T16:49:36Z",
        "updated_at": "2026-01-31T06:07:06Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18701",
        "categories": [
          "native_backends",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 18711,
        "kind": "pr",
        "state": "open",
        "title": "Support parsing JSON into grammar for schemas with no type and no properties",
        "labels": [],
        "created_at": "2026-01-09T07:37:47Z",
        "updated_at": "2026-01-09T08:06:05Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18711",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "api_ui_docs"
        ]
      },
      {
        "number": 18719,
        "kind": "pr",
        "state": "open",
        "title": "model: Add VAETKI support",
        "labels": [
          "model",
          "examples",
          "python"
        ],
        "created_at": "2026-01-09T14:42:46Z",
        "updated_at": "2026-01-20T12:59:06Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18719",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 18731,
        "kind": "pr",
        "state": "open",
        "title": "feat: add support for WeDLM architecture",
        "labels": [
          "python"
        ],
        "created_at": "2026-01-10T02:07:45Z",
        "updated_at": "2026-01-10T03:51:52Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18731",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 18735,
        "kind": "pr",
        "state": "open",
        "title": "llama: add canaries to Markdown files",
        "labels": [],
        "created_at": "2026-01-10T11:03:04Z",
        "updated_at": "2026-02-13T16:31:45Z",
        "closed_at": null,
        "comments": 12,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18735",
        "categories": [
          "models_architectures",
          "build_platform_ops"
        ]
      },
      {
        "number": 18742,
        "kind": "pr",
        "state": "open",
        "title": "server: add missing rerank and chat presets (#10932)",
        "labels": [],
        "created_at": "2026-01-10T17:02:16Z",
        "updated_at": "2026-01-10T17:02:16Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18742",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding"
        ]
      },
      {
        "number": 18745,
        "kind": "pr",
        "state": "open",
        "title": "fix: use actual tensor embedding dimension instead of model parameter",
        "labels": [],
        "created_at": "2026-01-10T22:15:57Z",
        "updated_at": "2026-02-11T18:16:50Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18745",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 18747,
        "kind": "pr",
        "state": "open",
        "title": "ggml, llama : add KV cache size limiting and block tracking infrastructure",
        "labels": [
          "model",
          "testing",
          "examples",
          "ggml"
        ],
        "created_at": "2026-01-11T00:46:10Z",
        "updated_at": "2026-04-27T11:53:00Z",
        "closed_at": null,
        "comments": 10,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18747",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 18790,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: CPU/GPU split not working anymore",
        "labels": [
          "bug",
          "regression"
        ],
        "created_at": "2026-01-12T18:02:54Z",
        "updated_at": "2026-01-19T22:45:07Z",
        "closed_at": null,
        "comments": 10,
        "url": "https://github.com/ggml-org/llama.cpp/issues/18790",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 18808,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Totally broken for agentic use on Intel dGPUs",
        "labels": [
          "bug",
          "performance",
          "regression"
        ],
        "created_at": "2026-01-13T14:13:46Z",
        "updated_at": "2026-03-22T16:20:24Z",
        "closed_at": null,
        "comments": 38,
        "url": "https://github.com/ggml-org/llama.cpp/issues/18808",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 18816,
        "kind": "pr",
        "state": "open",
        "title": "HIP: tune mmq/rocblas switching for RDNA4",
        "labels": [
          "Nvidia GPU",
          "ggml"
        ],
        "created_at": "2026-01-13T16:19:38Z",
        "updated_at": "2026-02-08T06:54:15Z",
        "closed_at": null,
        "comments": 9,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18816",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 18817,
        "kind": "pr",
        "state": "open",
        "title": "ggml-backend: Separate dynamic lib install and search paths, add relative search",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-01-13T20:01:29Z",
        "updated_at": "2026-05-17T11:58:52Z",
        "closed_at": null,
        "comments": 16,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18817",
        "categories": [
          "native_backends",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 18858,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cpu: add q4_0 repack support for wasm",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-01-15T09:59:01Z",
        "updated_at": "2026-04-01T18:08:12Z",
        "closed_at": null,
        "comments": 10,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18858",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing"
        ]
      },
      {
        "number": 18861,
        "kind": "pr",
        "state": "open",
        "title": "wasm, tests: fix ctests with emscripten",
        "labels": [
          "build",
          "testing",
          "ggml"
        ],
        "created_at": "2026-01-15T12:24:23Z",
        "updated_at": "2026-03-02T10:59:50Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18861",
        "categories": [
          "native_backends",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 18864,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: generalize MMQ CUDA kernel for floating-point data",
        "labels": [
          "enhancement",
          "stale"
        ],
        "created_at": "2026-01-15T13:23:03Z",
        "updated_at": "2026-08-21T02:13:35Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/issues/18864",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 18865,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Tune CUDA kernel selection logic for datacenter GPUs",
        "labels": [
          "enhancement",
          "research 🔬"
        ],
        "created_at": "2026-01-15T13:42:31Z",
        "updated_at": "2026-04-16T07:17:18Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/18865",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 18872,
        "kind": "pr",
        "state": "open",
        "title": "Try fixing non-ASCII parameters in llama-cli on Windows",
        "labels": [
          "examples"
        ],
        "created_at": "2026-01-16T00:20:02Z",
        "updated_at": "2026-01-17T11:59:24Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18872",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 18878,
        "kind": "pr",
        "state": "open",
        "title": "Metal : Supplement floor operator",
        "labels": [
          "ggml",
          "Apple Metal"
        ],
        "created_at": "2026-01-16T15:40:12Z",
        "updated_at": "2026-01-24T09:31:36Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18878",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 18879,
        "kind": "pr",
        "state": "open",
        "title": "llama-bench: add global --seed and reduce per-token synchronization",
        "labels": [
          "examples"
        ],
        "created_at": "2026-01-16T17:21:34Z",
        "updated_at": "2026-01-16T18:42:15Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18879",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 18892,
        "kind": "pr",
        "state": "open",
        "title": "examples: llama evaluation tool for mmlu, aime, gsm8k",
        "labels": [
          "examples",
          "python"
        ],
        "created_at": "2026-01-17T02:52:18Z",
        "updated_at": "2026-01-24T14:51:07Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18892",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 18923,
        "kind": "pr",
        "state": "open",
        "title": "common: add reranking server configuration preset ",
        "labels": [],
        "created_at": "2026-01-18T19:39:11Z",
        "updated_at": "2026-01-18T19:39:12Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18923",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding"
        ]
      },
      {
        "number": 18933,
        "kind": "pr",
        "state": "open",
        "title": "ggml-rpc: Add graceful error handling for graph compute operations",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-01-19T16:12:27Z",
        "updated_at": "2026-01-20T13:55:39Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18933",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 18963,
        "kind": "pr",
        "state": "open",
        "title": "server: fix json_schema response_format handling",
        "labels": [
          "examples",
          "python",
          "server"
        ],
        "created_at": "2026-01-20T15:22:35Z",
        "updated_at": "2026-07-04T17:23:20Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18963",
        "categories": [
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 18988,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Stack overflow in GBNF grammar via nested repetition",
        "labels": [
          "bug-unconfirmed",
          "security"
        ],
        "created_at": "2026-01-21T08:32:12Z",
        "updated_at": "2026-04-14T06:05:49Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/18988",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 18993,
        "kind": "pr",
        "state": "open",
        "title": "Fix: Prevent Stack Overflow in GBNF Grammar",
        "labels": [
          "testing"
        ],
        "created_at": "2026-01-21T12:29:42Z",
        "updated_at": "2026-04-17T06:05:26Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/18993",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 18996,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: GPT-OSS inference error",
        "labels": [
          "bug",
          "regression",
          "CUDA"
        ],
        "created_at": "2026-01-21T14:00:40Z",
        "updated_at": "2026-02-02T16:08:44Z",
        "closed_at": null,
        "comments": 13,
        "url": "https://github.com/ggml-org/llama.cpp/issues/18996",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 19034,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: completion differs across two identical requests, both with `cache_prompt = True`",
        "labels": [
          "bug",
          "server/api",
          "server"
        ],
        "created_at": "2026-01-22T21:50:22Z",
        "updated_at": "2026-01-23T23:45:01Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/19034",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 19085,
        "kind": "pr",
        "state": "open",
        "title": "fix(jinja): enforce recursion limit to prevent stack overflow",
        "labels": [
          "testing",
          "jinja parser"
        ],
        "created_at": "2026-01-25T10:16:58Z",
        "updated_at": "2026-01-29T12:44:31Z",
        "closed_at": null,
        "comments": 16,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19085",
        "categories": [
          "quantization",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing"
        ]
      },
      {
        "number": 19086,
        "kind": "pr",
        "state": "open",
        "title": "fix(grammar): enforce recursion limit to prevent stack overflow",
        "labels": [
          "testing"
        ],
        "created_at": "2026-01-25T10:59:24Z",
        "updated_at": "2026-01-25T19:53:25Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19086",
        "categories": [
          "quantization",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing"
        ]
      },
      {
        "number": 19098,
        "kind": "pr",
        "state": "open",
        "title": "CUDA: fix FA rowsum + fp16 softmax offset",
        "labels": [
          "Nvidia GPU",
          "ggml"
        ],
        "created_at": "2026-01-26T00:12:28Z",
        "updated_at": "2026-01-27T14:45:52Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19098",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 19101,
        "kind": "pr",
        "state": "open",
        "title": "Check if ctx or model is null before calling sampler",
        "labels": [
          "examples"
        ],
        "created_at": "2026-01-26T03:36:29Z",
        "updated_at": "2026-01-29T04:45:31Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19101",
        "categories": [
          "models_architectures",
          "structured_generation_tools",
          "api_ui_docs"
        ]
      },
      {
        "number": 19110,
        "kind": "issue",
        "state": "open",
        "title": "`--threads -1` double counts via `std::thread::hardware_concurrency()` due to hyper-threading",
        "labels": [
          "bug",
          "performance"
        ],
        "created_at": "2026-01-26T11:03:40Z",
        "updated_at": "2026-01-26T11:31:22Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/19110",
        "categories": [
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 19138,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Support OpenAI Responses API (/v1/responses) in llama.cpp server",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-01-27T14:16:29Z",
        "updated_at": "2026-08-13T20:33:45Z",
        "closed_at": null,
        "comments": 18,
        "url": "https://github.com/ggml-org/llama.cpp/issues/19138",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 19148,
        "kind": "pr",
        "state": "open",
        "title": "common: fix 32-bit overflow for avoiding year 2038 problem",
        "labels": [],
        "created_at": "2026-01-27T22:52:22Z",
        "updated_at": "2026-01-27T22:52:23Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19148",
        "categories": [
          "other"
        ]
      },
      {
        "number": 19166,
        "kind": "pr",
        "state": "open",
        "title": "Update server-models.cpp",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2026-01-28T20:20:33Z",
        "updated_at": "2026-01-28T20:29:36Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19166",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "api_ui_docs"
        ]
      },
      {
        "number": 19167,
        "kind": "pr",
        "state": "open",
        "title": "model : support LongCat-Flash-Lite (ngram embeddings)",
        "labels": [
          "model",
          "testing",
          "conversion"
        ],
        "created_at": "2026-01-28T22:17:29Z",
        "updated_at": "2026-08-05T22:20:00Z",
        "closed_at": null,
        "comments": 10,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19167",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "speculative_generation",
          "multimodal_embedding",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 19171,
        "kind": "pr",
        "state": "open",
        "title": "ggml: optimize ggml_vec_dot_mxfp4_q8_0 dot product on ARM SVE",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-01-29T06:48:18Z",
        "updated_at": "2026-02-16T08:44:11Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19171",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 19173,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Responses API (/v1/responses) can`t cancel a stream to stop generation",
        "labels": [
          "bug",
          "server/api",
          "server"
        ],
        "created_at": "2026-01-29T08:24:15Z",
        "updated_at": "2026-07-18T05:22:24Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/19173",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 19180,
        "kind": "pr",
        "state": "open",
        "title": "Hybrid model loading with DirectIO and MMAP",
        "labels": [],
        "created_at": "2026-01-29T14:59:51Z",
        "updated_at": "2026-01-29T15:00:03Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19180",
        "categories": [
          "native_backends",
          "models_architectures",
          "loader_conversion",
          "build_platform_ops"
        ]
      },
      {
        "number": 19182,
        "kind": "pr",
        "state": "open",
        "title": "model: support Longcat-Flash (need testing)",
        "labels": [
          "help wanted",
          "model",
          "testing",
          "python",
          "conversion"
        ],
        "created_at": "2026-01-29T17:34:59Z",
        "updated_at": "2026-08-14T10:07:52Z",
        "closed_at": null,
        "comments": 15,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19182",
        "categories": [
          "quantization",
          "models_architectures",
          "speculative_generation",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 19196,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cpu: optimize q4_0_q8_0 scales using Zvfhmin",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-01-30T05:24:40Z",
        "updated_at": "2026-01-30T06:16:54Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19196",
        "categories": [
          "quantization"
        ]
      },
      {
        "number": 19210,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: GLM 4.6V crashes llama-server when loading",
        "labels": [
          "bug"
        ],
        "created_at": "2026-01-30T15:42:20Z",
        "updated_at": "2026-01-31T11:57:38Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/19210",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 19234,
        "kind": "pr",
        "state": "open",
        "title": "chat: add thinking/reasoning support for Kimi K2",
        "labels": [],
        "created_at": "2026-01-31T18:52:59Z",
        "updated_at": "2026-03-02T06:14:34Z",
        "closed_at": null,
        "comments": 13,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19234",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing"
        ]
      },
      {
        "number": 19238,
        "kind": "pr",
        "state": "open",
        "title": "Feat: Adding token healing support for auto complete",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2026-02-01T04:03:31Z",
        "updated_at": "2026-02-01T13:52:23Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19238",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 19254,
        "kind": "pr",
        "state": "open",
        "title": "Add test for vk_buffer from host memory",
        "labels": [
          "Vulkan",
          "ggml"
        ],
        "created_at": "2026-02-01T20:50:07Z",
        "updated_at": "2026-02-24T06:29:47Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19254",
        "categories": [
          "native_backends",
          "memory_kv_context",
          "correctness_testing"
        ]
      },
      {
        "number": 19260,
        "kind": "pr",
        "state": "open",
        "title": "fix: use physical cores for --threads auto-detect (#19110)",
        "labels": [],
        "created_at": "2026-02-02T04:04:12Z",
        "updated_at": "2026-02-04T04:24:29Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19260",
        "categories": [
          "models_architectures",
          "build_platform_ops"
        ]
      },
      {
        "number": 19284,
        "kind": "pr",
        "state": "open",
        "title": "CANN: Multi-stream support",
        "labels": [
          "ggml",
          "Ascend NPU"
        ],
        "created_at": "2026-02-03T06:10:24Z",
        "updated_at": "2026-02-10T06:58:26Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19284",
        "categories": [
          "native_backends",
          "models_architectures"
        ]
      },
      {
        "number": 19285,
        "kind": "issue",
        "state": "open",
        "title": "Question: Metal4 tensor api support int8， will llama.cpp support it？",
        "labels": [
          "research 🔬"
        ],
        "created_at": "2026-02-03T07:47:03Z",
        "updated_at": "2026-02-06T03:45:57Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/19285",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "api_ui_docs"
        ]
      },
      {
        "number": 19294,
        "kind": "pr",
        "state": "open",
        "title": "spec: ngram-mod, score-based pruning",
        "labels": [],
        "created_at": "2026-02-03T13:59:07Z",
        "updated_at": "2026-02-03T14:14:00Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19294",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 19339,
        "kind": "pr",
        "state": "open",
        "title": "ui: Notebook Page",
        "labels": [
          "server/webui",
          "examples",
          "server",
          "server/ui"
        ],
        "created_at": "2026-02-04T18:21:35Z",
        "updated_at": "2026-08-17T11:04:42Z",
        "closed_at": null,
        "comments": 17,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19339",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 19341,
        "kind": "pr",
        "state": "open",
        "title": "Fix link failures in s390x",
        "labels": [
          "invalid",
          "ggml"
        ],
        "created_at": "2026-02-04T18:54:22Z",
        "updated_at": "2026-04-04T03:25:38Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19341",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 19349,
        "kind": "pr",
        "state": "open",
        "title": "fix: fail closed when grammar parsing fails (#19051)",
        "labels": [],
        "created_at": "2026-02-05T03:27:07Z",
        "updated_at": "2026-02-05T03:32:45Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19349",
        "categories": [
          "scheduling_serving",
          "structured_generation_tools",
          "training_adapters"
        ]
      },
      {
        "number": 19357,
        "kind": "pr",
        "state": "open",
        "title": "ggml : add int2 quantization and kleidiai sme2 gemm/gemv kernels",
        "labels": [
          "documentation",
          "testing",
          "examples",
          "ggml"
        ],
        "created_at": "2026-02-05T11:39:55Z",
        "updated_at": "2026-07-28T11:15:24Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19357",
        "categories": [
          "native_backends",
          "quantization",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 19396,
        "kind": "issue",
        "state": "open",
        "title": "Research: Qwen3 Next 80b - 8bit on Ryzen AI 9 HX370 (96GB System RAM)",
        "labels": [
          "research 🔬"
        ],
        "created_at": "2026-02-06T17:06:28Z",
        "updated_at": "2026-02-11T18:12:39Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/19396",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 19409,
        "kind": "pr",
        "state": "open",
        "title": "sampling : blue noise rng",
        "labels": [
          "testing",
          "examples",
          "server"
        ],
        "created_at": "2026-02-07T07:42:09Z",
        "updated_at": "2026-03-29T17:04:19Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19409",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 19410,
        "kind": "issue",
        "state": "open",
        "title": "Research: a GPU based bpe",
        "labels": [
          "research 🔬"
        ],
        "created_at": "2026-02-07T08:00:29Z",
        "updated_at": "2026-02-08T06:11:50Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/19410",
        "categories": [
          "native_backends",
          "models_architectures",
          "api_ui_docs"
        ]
      },
      {
        "number": 19413,
        "kind": "pr",
        "state": "open",
        "title": "Add compiler flags for UWP",
        "labels": [],
        "created_at": "2026-02-07T11:17:19Z",
        "updated_at": "2026-02-07T13:49:25Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19413",
        "categories": [
          "models_architectures",
          "build_platform_ops"
        ]
      },
      {
        "number": 19434,
        "kind": "pr",
        "state": "open",
        "title": "tools: add quant-bench for profiling raw kernel performance",
        "labels": [
          "examples"
        ],
        "created_at": "2026-02-08T17:27:59Z",
        "updated_at": "2026-02-09T17:44:55Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19434",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 19455,
        "kind": "pr",
        "state": "open",
        "title": "python: Use NumPy 2.0+",
        "labels": [
          "examples",
          "python",
          "server"
        ],
        "created_at": "2026-02-09T13:01:58Z",
        "updated_at": "2026-02-09T17:15:27Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19455",
        "categories": [
          "scheduling_serving",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 19458,
        "kind": "issue",
        "state": "open",
        "title": "Research: binary token input and output",
        "labels": [
          "research 🔬"
        ],
        "created_at": "2026-02-09T14:16:58Z",
        "updated_at": "2026-02-09T14:16:58Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/19458",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "loader_conversion",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 19459,
        "kind": "issue",
        "state": "open",
        "title": "Research: GPU accelerated utf8 to utf32",
        "labels": [
          "research 🔬"
        ],
        "created_at": "2026-02-09T14:32:04Z",
        "updated_at": "2026-02-09T14:32:04Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/19459",
        "categories": [
          "native_backends",
          "api_ui_docs"
        ]
      },
      {
        "number": 19466,
        "kind": "issue",
        "state": "open",
        "title": "Saving KV cache (using /slots/3?action=save API endpoint) does not work for vision-enabled models",
        "labels": [
          "enhancement",
          "server"
        ],
        "created_at": "2026-02-09T16:26:43Z",
        "updated_at": "2026-08-13T23:32:14Z",
        "closed_at": null,
        "comments": 38,
        "url": "https://github.com/ggml-org/llama.cpp/issues/19466",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 19472,
        "kind": "pr",
        "state": "open",
        "title": "common : fix filename validation for subfolders",
        "labels": [
          "testing",
          "examples",
          "server"
        ],
        "created_at": "2026-02-10T01:36:11Z",
        "updated_at": "2026-02-12T09:13:25Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19472",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 19482,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: larger model loading on ROCm hangs",
        "labels": [
          "bug",
          "RoCM"
        ],
        "created_at": "2026-02-10T08:58:27Z",
        "updated_at": "2026-08-26T16:53:54Z",
        "closed_at": null,
        "comments": 14,
        "url": "https://github.com/ggml-org/llama.cpp/issues/19482",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 19496,
        "kind": "pr",
        "state": "open",
        "title": "yet another android build & java bindings",
        "labels": [],
        "created_at": "2026-02-11T01:02:05Z",
        "updated_at": "2026-02-11T01:02:45Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19496",
        "categories": [
          "models_architectures",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 19503,
        "kind": "pr",
        "state": "open",
        "title": "grammar : fix grammar trigger crash when token extends beyond trigger pattern",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2026-02-11T06:07:48Z",
        "updated_at": "2026-02-15T04:22:14Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19503",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 19514,
        "kind": "pr",
        "state": "open",
        "title": "Fix OS support for x86 simd architectures",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-02-11T11:41:20Z",
        "updated_at": "2026-02-11T14:55:16Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19514",
        "categories": [
          "models_architectures"
        ]
      },
      {
        "number": 19527,
        "kind": "pr",
        "state": "open",
        "title": "fix: Metal backend performance on AMD discrete GPUs",
        "labels": [
          "ggml",
          "Apple Metal",
          "vibe-coded"
        ],
        "created_at": "2026-02-11T21:29:59Z",
        "updated_at": "2026-08-15T19:06:29Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19527",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 19553,
        "kind": "pr",
        "state": "open",
        "title": "templates : fix double-escaping in gpt-oss tool call arguments and responses",
        "labels": [],
        "created_at": "2026-02-12T11:07:21Z",
        "updated_at": "2026-02-13T12:31:25Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19553",
        "categories": [
          "models_architectures",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 19590,
        "kind": "pr",
        "state": "open",
        "title": "WASM Relaxed SIMD Enhancement",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-02-13T13:08:46Z",
        "updated_at": "2026-02-13T13:13:00Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19590",
        "categories": [
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 19600,
        "kind": "pr",
        "state": "open",
        "title": "metal: use mul_mv_ext for large n on non-simdgroup_mm GPUs",
        "labels": [
          "ggml",
          "Apple Metal"
        ],
        "created_at": "2026-02-13T17:27:07Z",
        "updated_at": "2026-02-13T17:27:34Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19600",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 19662,
        "kind": "issue",
        "state": "open",
        "title": "CUDA compilation fails on Blackwell (sm_120) with MXFP4: \"Instruction 'mma with block scale' not supported\"",
        "labels": [
          "bug",
          "build",
          "CUDA"
        ],
        "created_at": "2026-02-16T09:02:06Z",
        "updated_at": "2026-05-04T11:23:43Z",
        "closed_at": null,
        "comments": 15,
        "url": "https://github.com/ggml-org/llama.cpp/issues/19662",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 19670,
        "kind": "pr",
        "state": "open",
        "title": "Allow partial success of seq_rm for hybrid memory",
        "labels": [],
        "created_at": "2026-02-16T17:28:27Z",
        "updated_at": "2026-03-12T15:23:52Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19670",
        "categories": [
          "models_architectures",
          "memory_kv_context"
        ]
      },
      {
        "number": 19671,
        "kind": "pr",
        "state": "open",
        "title": "Add Pylint workflow for Python code analysis",
        "labels": [
          "devops",
          "server"
        ],
        "created_at": "2026-02-16T18:46:32Z",
        "updated_at": "2026-08-26T00:35:31Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19671",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 19691,
        "kind": "pr",
        "state": "open",
        "title": "test(server): add multi-image and no-image vision API tests",
        "labels": [
          "examples",
          "python",
          "server"
        ],
        "created_at": "2026-02-17T15:24:46Z",
        "updated_at": "2026-02-17T15:57:05Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19691",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 19694,
        "kind": "pr",
        "state": "open",
        "title": "server : fix V-L embedding model support",
        "labels": [],
        "created_at": "2026-02-17T16:24:03Z",
        "updated_at": "2026-02-23T06:00:38Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19694",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing"
        ]
      },
      {
        "number": 19706,
        "kind": "pr",
        "state": "open",
        "title": "Q6_K - Block Interleaving Implementation for x86 SIMD (AVX512/AVX2)",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-02-18T07:11:02Z",
        "updated_at": "2026-07-08T04:59:08Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19706",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 19707,
        "kind": "pr",
        "state": "open",
        "title": "Q5_K - Block Interleaving Implementation for x86 SIMD (AVX512/AVX2)",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-02-18T07:11:19Z",
        "updated_at": "2026-02-19T10:43:01Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19707",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 19720,
        "kind": "pr",
        "state": "open",
        "title": "server: add OpenAI Responses API compliance",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2026-02-19T02:45:13Z",
        "updated_at": "2026-03-25T21:41:35Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19720",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 19725,
        "kind": "pr",
        "state": "open",
        "title": "llama: add BertForTokenClassification support",
        "labels": [
          "examples",
          "python",
          "server"
        ],
        "created_at": "2026-02-19T06:24:06Z",
        "updated_at": "2026-02-26T18:39:22Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19725",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "api_ui_docs"
        ]
      },
      {
        "number": 19743,
        "kind": "pr",
        "state": "open",
        "title": "ggml-vulkan: Add support for TQ1_0 and TQ2_0 types for MUL_MAT",
        "labels": [
          "testing",
          "Vulkan",
          "ggml"
        ],
        "created_at": "2026-02-19T20:47:12Z",
        "updated_at": "2026-02-26T00:41:55Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19743",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 19755,
        "kind": "pr",
        "state": "open",
        "title": "cmake : respect GGML_LIB_INSTALL_DIR and LLAMA_LIB_INSTALL_DIR",
        "labels": [
          "build",
          "ggml"
        ],
        "created_at": "2026-02-20T08:05:08Z",
        "updated_at": "2026-04-07T01:47:13Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19755",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 19812,
        "kind": "pr",
        "state": "open",
        "title": "implemented max pooling for embeddings",
        "labels": [
          "examples",
          "python",
          "server"
        ],
        "created_at": "2026-02-22T19:03:47Z",
        "updated_at": "2026-02-27T23:26:15Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19812",
        "categories": [
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 19832,
        "kind": "pr",
        "state": "open",
        "title": "Add Aya 101 multi-lingual translation support to llama.cpp",
        "labels": [
          "examples",
          "python",
          "server"
        ],
        "created_at": "2026-02-23T13:23:58Z",
        "updated_at": "2026-03-06T16:00:01Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19832",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 19841,
        "kind": "pr",
        "state": "open",
        "title": "server : add chat truncation to keep chat going",
        "labels": [
          "testing",
          "examples",
          "python",
          "server"
        ],
        "created_at": "2026-02-23T21:36:34Z",
        "updated_at": "2026-04-05T07:52:06Z",
        "closed_at": null,
        "comments": 19,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19841",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 19855,
        "kind": "pr",
        "state": "open",
        "title": "server : add default-model preset and fallback logic",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2026-02-24T16:30:31Z",
        "updated_at": "2026-08-21T21:06:11Z",
        "closed_at": null,
        "comments": 15,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19855",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 19936,
        "kind": "pr",
        "state": "open",
        "title": "fix dots.ocr: correct RoPE sections and FFN tensor mapping",
        "labels": [
          "examples",
          "python"
        ],
        "created_at": "2026-02-26T16:27:36Z",
        "updated_at": "2026-02-26T16:39:36Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19936",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 19938,
        "kind": "pr",
        "state": "open",
        "title": "scripts: ini_to_opencode.py",
        "labels": [
          "script",
          "python"
        ],
        "created_at": "2026-02-26T18:13:01Z",
        "updated_at": "2026-02-28T00:32:29Z",
        "closed_at": null,
        "comments": 14,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19938",
        "categories": [
          "quantization",
          "models_architectures",
          "multimodal_embedding",
          "loader_conversion",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 19941,
        "kind": "pr",
        "state": "open",
        "title": "Quantization laboratory: new per-tensor codebook quants, quantization recipe generator",
        "labels": [
          "documentation",
          "script",
          "testing",
          "Nvidia GPU",
          "Vulkan",
          "examples",
          "python",
          "ggml",
          "OpenCL",
          "CUDA",
          "OpenVINO",
          "conversion"
        ],
        "created_at": "2026-02-26T22:15:54Z",
        "updated_at": "2026-08-18T15:40:08Z",
        "closed_at": null,
        "comments": 33,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19941",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 19971,
        "kind": "pr",
        "state": "open",
        "title": "feat: add --threads-all option to llama-bench",
        "labels": [
          "examples"
        ],
        "created_at": "2026-02-28T05:57:26Z",
        "updated_at": "2026-02-28T05:57:41Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19971",
        "categories": [
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 19974,
        "kind": "pr",
        "state": "open",
        "title": "Add file existence and type checks for imatrix",
        "labels": [
          "examples"
        ],
        "created_at": "2026-02-28T08:04:27Z",
        "updated_at": "2026-02-28T08:04:42Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19974",
        "categories": [
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 19990,
        "kind": "pr",
        "state": "open",
        "title": "build: fix various compiler warnings on Windows MinGW",
        "labels": [
          "testing",
          "examples"
        ],
        "created_at": "2026-02-28T15:58:42Z",
        "updated_at": "2026-02-28T16:11:30Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19990",
        "categories": [
          "models_architectures",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 19992,
        "kind": "pr",
        "state": "open",
        "title": "common : fix common_chat_peg_parse for incomplete utf-8 sequence tail",
        "labels": [
          "testing"
        ],
        "created_at": "2026-02-28T17:36:09Z",
        "updated_at": "2026-03-02T06:43:15Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/19992",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing"
        ]
      },
      {
        "number": 20009,
        "kind": "pr",
        "state": "open",
        "title": "server: add Qwen3-Reranker instruction support",
        "labels": [
          "examples",
          "python",
          "server"
        ],
        "created_at": "2026-03-01T14:15:41Z",
        "updated_at": "2026-04-13T02:38:56Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20009",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 20017,
        "kind": "pr",
        "state": "open",
        "title": "vulkan: add sparse OOM fallback for large UMA allocations and chunked staging fallback",
        "labels": [
          "testing",
          "Vulkan",
          "ggml"
        ],
        "created_at": "2026-03-01T20:02:34Z",
        "updated_at": "2026-03-02T03:35:18Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20017",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 20029,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: b8143 produces garbage Mac x86 Vulkan with AMD GPU",
        "labels": [
          "macos",
          "AMD GPU",
          "bug-unconfirmed",
          "Vulkan",
          "stale"
        ],
        "created_at": "2026-03-02T00:35:11Z",
        "updated_at": "2026-08-24T01:07:52Z",
        "closed_at": null,
        "comments": 32,
        "url": "https://github.com/ggml-org/llama.cpp/issues/20029",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 20050,
        "kind": "pr",
        "state": "open",
        "title": "server: fix infinite retry loop when KV cache is full",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2026-03-02T23:02:10Z",
        "updated_at": "2026-04-13T19:38:48Z",
        "closed_at": null,
        "comments": 10,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20050",
        "categories": [
          "memory_kv_context",
          "scheduling_serving",
          "api_ui_docs"
        ]
      },
      {
        "number": 20064,
        "kind": "pr",
        "state": "open",
        "title": "cmake: fix ARM feature detection hang on platforms without SVE/SME",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-03-03T10:36:50Z",
        "updated_at": "2026-03-04T12:03:40Z",
        "closed_at": null,
        "comments": 10,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20064",
        "categories": [
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 20076,
        "kind": "pr",
        "state": "open",
        "title": "fix: correct EXAONE3 FFN_DOWN tensor mapping prefix",
        "labels": [
          "python"
        ],
        "created_at": "2026-03-03T15:47:55Z",
        "updated_at": "2026-03-03T16:47:28Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20076",
        "categories": [
          "models_architectures",
          "loader_conversion",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 20086,
        "kind": "pr",
        "state": "open",
        "title": "llama : add attention weights extraction API [EXPERIMENTAL]",
        "labels": [
          "examples",
          "python"
        ],
        "created_at": "2026-03-03T17:12:19Z",
        "updated_at": "2026-03-31T20:14:30Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20086",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 20114,
        "kind": "pr",
        "state": "open",
        "title": "add elixir bindings in readme",
        "labels": [],
        "created_at": "2026-03-04T20:09:58Z",
        "updated_at": "2026-06-21T01:12:03Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20114",
        "categories": [
          "api_ui_docs"
        ]
      },
      {
        "number": 20147,
        "kind": "pr",
        "state": "open",
        "title": "[DRAFT]: add c++ wrapper backend.",
        "labels": [
          "build",
          "ggml",
          "Apple Metal"
        ],
        "created_at": "2026-03-05T21:23:49Z",
        "updated_at": "2026-03-15T23:51:48Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20147",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 20155,
        "kind": "pr",
        "state": "open",
        "title": "server: add optional security audit logging for compliance and monitoring",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2026-03-06T04:11:03Z",
        "updated_at": "2026-03-07T08:42:12Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20155",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 20176,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Qwen 3.5 Loading checkpoints causes a crash",
        "labels": [
          "bug",
          "AMD GPU"
        ],
        "created_at": "2026-03-06T22:13:53Z",
        "updated_at": "2026-07-28T09:24:19Z",
        "closed_at": null,
        "comments": 17,
        "url": "https://github.com/ggml-org/llama.cpp/issues/20176",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing"
        ]
      },
      {
        "number": 20242,
        "kind": "pr",
        "state": "open",
        "title": "gguf-py: validate metadata values against declared types",
        "labels": [
          "python"
        ],
        "created_at": "2026-03-08T16:30:24Z",
        "updated_at": "2026-03-08T19:53:11Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20242",
        "categories": [
          "quantization",
          "loader_conversion",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 20260,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: unsloth/Qwen3.5-35B-A3B-GGUF `peg-native` chat format parser fails when model outputs text before `<tool_call>` (thinking model + tool calling)",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-03-09T02:46:35Z",
        "updated_at": "2026-07-28T09:44:45Z",
        "closed_at": null,
        "comments": 33,
        "url": "https://github.com/ggml-org/llama.cpp/issues/20260",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "training_adapters"
        ]
      },
      {
        "number": 20265,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Model output directed to Reasoning block instead of standard response.",
        "labels": [
          "bug",
          "regression",
          "chat parser"
        ],
        "created_at": "2026-03-09T04:33:18Z",
        "updated_at": "2026-03-31T14:53:38Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/issues/20265",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 20375,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cuda : add flash attention support for head size 88 (Llama 4 Vision)",
        "labels": [
          "Nvidia GPU",
          "python",
          "ggml"
        ],
        "created_at": "2026-03-11T02:19:07Z",
        "updated_at": "2026-03-12T18:46:32Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20375",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 20376,
        "kind": "pr",
        "state": "open",
        "title": "vulkan: f16 mixed-precision state for GATED_DELTA_NET",
        "labels": [
          "testing",
          "Vulkan",
          "ggml"
        ],
        "created_at": "2026-03-11T03:19:23Z",
        "updated_at": "2026-03-12T16:42:34Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20376",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 20377,
        "kind": "pr",
        "state": "open",
        "title": "vulkan: chunked parallel kernel for GATED_DELTA_NET",
        "labels": [
          "documentation",
          "model",
          "script",
          "testing",
          "Nvidia GPU",
          "Vulkan",
          "ggml"
        ],
        "created_at": "2026-03-11T03:19:32Z",
        "updated_at": "2026-07-05T08:54:43Z",
        "closed_at": null,
        "comments": 17,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20377",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 20388,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cpu: fix x86 SIMD detection — validate OS YMM/ZMM/AMX state via XCR0",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-03-11T08:35:27Z",
        "updated_at": "2026-03-12T01:33:59Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20388",
        "categories": [
          "native_backends",
          "quantization",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 20433,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: #17795 introduces subtle correctness errors",
        "labels": [
          "bug",
          "AMD GPU",
          "critical severity",
          "CUDA"
        ],
        "created_at": "2026-03-11T22:24:01Z",
        "updated_at": "2026-03-13T13:51:27Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/issues/20433",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 20451,
        "kind": "pr",
        "state": "open",
        "title": "vulkan: Slang flash attention shader",
        "labels": [
          "Vulkan",
          "ggml"
        ],
        "created_at": "2026-03-12T11:04:53Z",
        "updated_at": "2026-03-13T15:59:38Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20451",
        "categories": [
          "native_backends",
          "quantization",
          "memory_kv_context",
          "build_platform_ops"
        ]
      },
      {
        "number": 20454,
        "kind": "pr",
        "state": "open",
        "title": "MoE expert profiling and REAP-based pruning tools",
        "labels": [
          "examples",
          "python"
        ],
        "created_at": "2026-03-12T12:01:09Z",
        "updated_at": "2026-05-04T05:54:25Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20454",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 20456,
        "kind": "pr",
        "state": "open",
        "title": "metal : add NVFP4 quantization support",
        "labels": [
          "ggml",
          "Apple Metal"
        ],
        "created_at": "2026-03-12T13:44:27Z",
        "updated_at": "2026-06-23T18:47:04Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20456",
        "categories": [
          "native_backends",
          "quantization",
          "correctness_testing"
        ]
      },
      {
        "number": 20470,
        "kind": "pr",
        "state": "open",
        "title": "common : fix SIGSEGV in _visit_pattern when quantifier follows empty seq",
        "labels": [],
        "created_at": "2026-03-12T21:37:56Z",
        "updated_at": "2026-03-12T21:37:56Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20470",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 20475,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Webui seems not providing api key when using `Use llama-server proxy` for MCP",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-03-13T02:52:58Z",
        "updated_at": "2026-08-01T01:15:02Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/issues/20475",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 20479,
        "kind": "pr",
        "state": "open",
        "title": "server : support reasoning parameters from OAI, OpenRouter and Claude",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2026-03-13T04:55:31Z",
        "updated_at": "2026-06-26T07:00:54Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20479",
        "categories": [
          "scheduling_serving",
          "api_ui_docs"
        ]
      },
      {
        "number": 20487,
        "kind": "pr",
        "state": "open",
        "title": "tests : add test save and restore state",
        "labels": [
          "testing"
        ],
        "created_at": "2026-03-13T08:00:51Z",
        "updated_at": "2026-03-13T08:12:40Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20487",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 20520,
        "kind": "pr",
        "state": "open",
        "title": "CUDA: add new op template rms_norm_f32_vec4 for float4 vectorized load/store",
        "labels": [
          "testing",
          "Nvidia GPU",
          "ggml"
        ],
        "created_at": "2026-03-13T16:48:49Z",
        "updated_at": "2026-04-13T19:23:42Z",
        "closed_at": null,
        "comments": 9,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20520",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 20535,
        "kind": "pr",
        "state": "open",
        "title": "chat: fix thinking support for Kimi K2 Thinking and K2.5",
        "labels": [],
        "created_at": "2026-03-14T04:36:07Z",
        "updated_at": "2026-04-09T16:59:04Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20535",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "structured_generation_tools",
          "correctness_testing"
        ]
      },
      {
        "number": 20566,
        "kind": "pr",
        "state": "open",
        "title": "fix(openvino): replace hardcoded TBB archive path with find_package(TBB)",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-03-15T00:18:29Z",
        "updated_at": "2026-03-15T05:41:31Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20566",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 20596,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cpu: improve `--n-cpu-moe` TG performance",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-03-15T15:51:29Z",
        "updated_at": "2026-03-28T13:20:54Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20596",
        "categories": [
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 20609,
        "kind": "pr",
        "state": "open",
        "title": "ggml: MXFP flash attention with SoA layout (CPU scalar reference)",
        "labels": [
          "testing",
          "examples",
          "ggml",
          "SYCL",
          "Apple Metal",
          "OpenCL"
        ],
        "created_at": "2026-03-15T21:33:51Z",
        "updated_at": "2026-03-23T09:12:38Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20609",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 20623,
        "kind": "pr",
        "state": "open",
        "title": "Speed up claude-code by prevent adding system message that starts with 'x-anthropic-'",
        "labels": [
          "server"
        ],
        "created_at": "2026-03-16T06:12:16Z",
        "updated_at": "2026-03-26T06:57:20Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20623",
        "categories": [
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 20624,
        "kind": "pr",
        "state": "open",
        "title": "MiroThinker tool call parser",
        "labels": [
          "testing"
        ],
        "created_at": "2026-03-16T06:21:51Z",
        "updated_at": "2026-04-01T15:24:38Z",
        "closed_at": null,
        "comments": 13,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20624",
        "categories": [
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing"
        ]
      },
      {
        "number": 20628,
        "kind": "pr",
        "state": "open",
        "title": "model : add QKV weight fusion for LLaMA, Qwen2, and Qwen3",
        "labels": [
          "model",
          "script",
          "python"
        ],
        "created_at": "2026-03-16T07:59:13Z",
        "updated_at": "2026-03-19T02:39:11Z",
        "closed_at": null,
        "comments": 14,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20628",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "loader_conversion",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 20632,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Graceful reasoning budget termination. Avoid mid-sentence cutoff.",
        "labels": [
          "enhancement",
          "help wanted",
          "good first issue"
        ],
        "created_at": "2026-03-16T10:29:32Z",
        "updated_at": "2026-08-23T06:37:06Z",
        "closed_at": null,
        "comments": 10,
        "url": "https://github.com/ggml-org/llama.cpp/issues/20632",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 20650,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Getting 500 - Failed to parse input at pos x when tool calling",
        "labels": [
          "bug"
        ],
        "created_at": "2026-03-16T19:17:34Z",
        "updated_at": "2026-05-08T16:10:35Z",
        "closed_at": null,
        "comments": 36,
        "url": "https://github.com/ggml-org/llama.cpp/issues/20650",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 20670,
        "kind": "pr",
        "state": "open",
        "title": "Mention ANV_SYS_MEM_LIMIT in Vulkan/Linux section of build.md",
        "labels": [
          "documentation"
        ],
        "created_at": "2026-03-17T07:45:39Z",
        "updated_at": "2026-03-17T07:45:55Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20670",
        "categories": [
          "native_backends",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 20674,
        "kind": "pr",
        "state": "open",
        "title": "resolve missing <chrono> and pointer type issues for Windows/Clang build",
        "labels": [
          "examples"
        ],
        "created_at": "2026-03-17T10:53:58Z",
        "updated_at": "2026-03-17T10:54:14Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20674",
        "categories": [
          "quantization",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 20675,
        "kind": "issue",
        "state": "open",
        "title": "Refactor: MCP client on CLI",
        "labels": [],
        "created_at": "2026-03-17T11:05:07Z",
        "updated_at": "2026-08-05T01:11:04Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/20675",
        "categories": [
          "models_architectures",
          "scheduling_serving"
        ]
      },
      {
        "number": 20678,
        "kind": "issue",
        "state": "open",
        "title": "Refactor: support skills",
        "labels": [
          "stale"
        ],
        "created_at": "2026-03-17T11:07:40Z",
        "updated_at": "2026-08-17T01:07:47Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/20678",
        "categories": [
          "other"
        ]
      },
      {
        "number": 20697,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Disk-based context checkpoint offloading (`--cache-disk`)",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-03-17T19:49:31Z",
        "updated_at": "2026-08-01T18:28:38Z",
        "closed_at": null,
        "comments": 19,
        "url": "https://github.com/ggml-org/llama.cpp/issues/20697",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 20716,
        "kind": "pr",
        "state": "open",
        "title": "gguf : fix division by zero",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-03-18T10:40:54Z",
        "updated_at": "2026-07-15T17:10:26Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20716",
        "categories": [
          "quantization",
          "multimodal_embedding",
          "correctness_testing"
        ]
      },
      {
        "number": 20723,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cpu: extend RVV repack GEMM and GEMV to other VLENs",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-03-18T16:44:04Z",
        "updated_at": "2026-07-29T12:26:45Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20723",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 20729,
        "kind": "pr",
        "state": "open",
        "title": "server: workaround new chat parser regression",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2026-03-18T19:55:05Z",
        "updated_at": "2026-03-20T05:40:41Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20729",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 20759,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cpu: fix cross-compilation to Windows from Linux",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-03-19T14:29:30Z",
        "updated_at": "2026-03-19T14:45:59Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20759",
        "categories": [
          "build_platform_ops"
        ]
      },
      {
        "number": 20769,
        "kind": "issue",
        "state": "open",
        "title": "Refactor: allow using mcp.json file, expose them via /tools API",
        "labels": [
          "refactoring"
        ],
        "created_at": "2026-03-19T18:44:18Z",
        "updated_at": "2026-07-16T20:19:04Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/20769",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "api_ui_docs"
        ]
      },
      {
        "number": 20778,
        "kind": "pr",
        "state": "open",
        "title": "sampling : handle grammar prefill crash for Functionary v3.2",
        "labels": [
          "testing"
        ],
        "created_at": "2026-03-20T01:50:55Z",
        "updated_at": "2026-03-20T14:04:56Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20778",
        "categories": [
          "models_architectures",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 20792,
        "kind": "pr",
        "state": "open",
        "title": "fix: create_tensor returns non-null for missing TENSOR_NOT_REQUIRED tensors when using llama_model_init_from_user",
        "labels": [],
        "created_at": "2026-03-20T10:04:15Z",
        "updated_at": "2026-03-20T10:06:18Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20792",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 20811,
        "kind": "pr",
        "state": "open",
        "title": "fix(opencl): free host buffers after MoE MXFP4 router table reorder",
        "labels": [
          "ggml",
          "OpenCL"
        ],
        "created_at": "2026-03-20T16:52:26Z",
        "updated_at": "2026-03-27T21:19:00Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20811",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing"
        ]
      },
      {
        "number": 20819,
        "kind": "pr",
        "state": "open",
        "title": "server: persist context checkpoints across slot save/restore",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2026-03-20T20:16:51Z",
        "updated_at": "2026-06-07T18:45:11Z",
        "closed_at": null,
        "comments": 9,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20819",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 20822,
        "kind": "pr",
        "state": "open",
        "title": "server: auto-save/restore slot state in router mode",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2026-03-20T21:36:27Z",
        "updated_at": "2026-03-20T21:40:07Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20822",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 20831,
        "kind": "pr",
        "state": "open",
        "title": "cuda : dynamic MMVQ nwarps for narrow matrices",
        "labels": [
          "Nvidia GPU",
          "ggml"
        ],
        "created_at": "2026-03-21T02:32:22Z",
        "updated_at": "2026-08-18T19:08:01Z",
        "closed_at": null,
        "comments": 16,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20831",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing"
        ]
      },
      {
        "number": 20837,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Qwen3.5 9B often prints tool calls in XML and stops when thinking is enabled - tool calls inside thinking block",
        "labels": [
          "bug",
          "chat parser"
        ],
        "created_at": "2026-03-21T11:21:18Z",
        "updated_at": "2026-08-10T13:57:13Z",
        "closed_at": null,
        "comments": 59,
        "url": "https://github.com/ggml-org/llama.cpp/issues/20837",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 20858,
        "kind": "pr",
        "state": "open",
        "title": "fix(server): Print free memory as a signed integer to indicate VRAM spillover into host memory",
        "labels": [],
        "created_at": "2026-03-22T06:58:09Z",
        "updated_at": "2026-03-22T06:58:10Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20858",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 20862,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: With Vulkan, `-sm row` and `-sm layer` are not equivalent",
        "labels": [
          "bug",
          "Vulkan"
        ],
        "created_at": "2026-03-22T13:15:23Z",
        "updated_at": "2026-05-08T04:43:36Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/20862",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 20872,
        "kind": "pr",
        "state": "open",
        "title": "server : add SSE headers to fix streaming behind reverse proxy",
        "labels": [
          "documentation",
          "model",
          "testing",
          "examples",
          "python",
          "server",
          "conversion"
        ],
        "created_at": "2026-03-22T20:58:44Z",
        "updated_at": "2026-08-19T08:33:07Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20872",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 20940,
        "kind": "pr",
        "state": "open",
        "title": "fix: gate AVX/AVX-512/AMX on OS XSAVE support via XGETBV",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-03-24T08:59:23Z",
        "updated_at": "2026-03-24T11:15:16Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20940",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 20956,
        "kind": "pr",
        "state": "open",
        "title": "fix(common): fix fail to run TranslateGemma",
        "labels": [],
        "created_at": "2026-03-24T16:31:44Z",
        "updated_at": "2026-07-05T11:58:42Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20956",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 20958,
        "kind": "pr",
        "state": "open",
        "title": "Update description for llama.cpp to clarify quantization in README.md",
        "labels": [],
        "created_at": "2026-03-24T17:12:21Z",
        "updated_at": "2026-05-25T01:48:59Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20958",
        "categories": [
          "quantization",
          "models_architectures",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 20961,
        "kind": "pr",
        "state": "open",
        "title": "grammar: optimize big grammar matching by using left-factorization / optimize grammar cloning",
        "labels": [
          "testing"
        ],
        "created_at": "2026-03-24T19:34:01Z",
        "updated_at": "2026-03-25T09:19:02Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20961",
        "categories": [
          "models_architectures",
          "structured_generation_tools",
          "correctness_testing"
        ]
      },
      {
        "number": 20966,
        "kind": "pr",
        "state": "open",
        "title": "Install libraries into LLAMA_LIB_INSTALL_DIR",
        "labels": [
          "build",
          "examples",
          "ggml"
        ],
        "created_at": "2026-03-25T00:49:34Z",
        "updated_at": "2026-04-01T23:51:00Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20966",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "multimodal_embedding",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 20983,
        "kind": "pr",
        "state": "open",
        "title": "ggml : skip null devices in backend registry",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-03-25T10:46:10Z",
        "updated_at": "2026-08-24T22:53:32Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20983",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 20991,
        "kind": "pr",
        "state": "open",
        "title": "ci: add riscv64 to release binaries",
        "labels": [
          "devops",
          "ggml"
        ],
        "created_at": "2026-03-25T14:19:57Z",
        "updated_at": "2026-05-07T13:24:31Z",
        "closed_at": null,
        "comments": 11,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20991",
        "categories": [
          "models_architectures",
          "build_platform_ops"
        ]
      },
      {
        "number": 20992,
        "kind": "pr",
        "state": "open",
        "title": "Update Android link in README.md",
        "labels": [],
        "created_at": "2026-03-25T14:33:25Z",
        "updated_at": "2026-03-25T14:35:44Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/20992",
        "categories": [
          "models_architectures",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 21002,
        "kind": "pr",
        "state": "open",
        "title": "ci: introduce audits for self-hosted runners",
        "labels": [
          "devops"
        ],
        "created_at": "2026-03-25T18:05:25Z",
        "updated_at": "2026-03-26T14:13:26Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21002",
        "categories": [
          "native_backends",
          "models_architectures",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 21003,
        "kind": "pr",
        "state": "open",
        "title": "grammar: increase MAX_REPETITION_THRESHOLD + make it configurable via envvar",
        "labels": [
          "testing"
        ],
        "created_at": "2026-03-25T18:08:10Z",
        "updated_at": "2026-05-10T14:17:34Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21003",
        "categories": [
          "native_backends",
          "models_architectures",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 21004,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Error in CI on coopmat2 in Vulkan",
        "labels": [
          "bug",
          "Vulkan",
          "CI / packaging"
        ],
        "created_at": "2026-03-25T18:13:12Z",
        "updated_at": "2026-03-25T19:28:15Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/21004",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 21024,
        "kind": "pr",
        "state": "open",
        "title": "Vulkan Repack PoC",
        "labels": [
          "Vulkan",
          "ggml"
        ],
        "created_at": "2026-03-26T12:03:17Z",
        "updated_at": "2026-06-10T14:00:31Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21024",
        "categories": [
          "native_backends",
          "memory_kv_context"
        ]
      },
      {
        "number": 21032,
        "kind": "pr",
        "state": "open",
        "title": "model: add support for nvidia/gpt-oss-puzzle-88B",
        "labels": [
          "model",
          "python"
        ],
        "created_at": "2026-03-26T15:38:43Z",
        "updated_at": "2026-06-29T23:38:51Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21032",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 21045,
        "kind": "pr",
        "state": "open",
        "title": "model: add Falcon OCR support",
        "labels": [
          "model",
          "examples",
          "python",
          "ggml",
          "Apple Metal"
        ],
        "created_at": "2026-03-26T22:07:50Z",
        "updated_at": "2026-04-24T13:01:19Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21045",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 21067,
        "kind": "pr",
        "state": "open",
        "title": "ggml: allow prefetching tensor overrides",
        "labels": [
          "Nvidia GPU",
          "Vulkan",
          "examples",
          "ggml",
          "SYCL",
          "Apple Metal",
          "Ascend NPU",
          "OpenCL",
          "IBM zDNN",
          "Hexagon",
          "CUDA",
          "AMD ZenDNN",
          "OpenVINO",
          "WebGPU"
        ],
        "created_at": "2026-03-27T14:02:20Z",
        "updated_at": "2026-08-15T04:22:14Z",
        "closed_at": null,
        "comments": 22,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21067",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 21070,
        "kind": "pr",
        "state": "open",
        "title": "Add quantization recipes from custom recipe files",
        "labels": [
          "testing",
          "examples"
        ],
        "created_at": "2026-03-27T14:52:15Z",
        "updated_at": "2026-03-27T14:56:14Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21070",
        "categories": [
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 21119,
        "kind": "pr",
        "state": "open",
        "title": "metal: add opt-in V skip for negligible attention weights",
        "labels": [
          "ggml",
          "Apple Metal"
        ],
        "created_at": "2026-03-28T13:09:46Z",
        "updated_at": "2026-04-05T01:15:45Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21119",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 21125,
        "kind": "issue",
        "state": "open",
        "title": "Compile bug: convert_lora_to_gguf.py fails for Qwen3.5 LoRA at _reorder_v_heads -> LoraTorchTensor.reshape() with NotImplementedError",
        "labels": [
          "bug"
        ],
        "created_at": "2026-03-28T16:03:10Z",
        "updated_at": "2026-05-05T11:36:03Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/21125",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 21133,
        "kind": "issue",
        "state": "open",
        "title": "server: slot save/restore and checkpoints blocked for text-only conversations when mmproj is loaded",
        "labels": [
          "stale"
        ],
        "created_at": "2026-03-28T23:10:09Z",
        "updated_at": "2026-08-16T01:07:56Z",
        "closed_at": null,
        "comments": 10,
        "url": "https://github.com/ggml-org/llama.cpp/issues/21133",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 21139,
        "kind": "pr",
        "state": "open",
        "title": "grammar: make MAX_REPETITION_THRESHOLD configurable via env var",
        "labels": [],
        "created_at": "2026-03-29T01:12:18Z",
        "updated_at": "2026-03-29T01:40:04Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21139",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 21141,
        "kind": "pr",
        "state": "open",
        "title": "common: add two-phase graceful reasoning budget termination ...",
        "labels": [
          "testing"
        ],
        "created_at": "2026-03-29T03:03:37Z",
        "updated_at": "2026-07-22T19:18:26Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21141",
        "categories": [
          "models_architectures",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 21160,
        "kind": "pr",
        "state": "open",
        "title": "Cross-backend profiler",
        "labels": [
          "documentation",
          "testing",
          "Nvidia GPU",
          "Vulkan",
          "examples",
          "python",
          "server",
          "ggml",
          "SYCL",
          "Apple Metal",
          "Ascend NPU",
          "OpenCL",
          "IBM zDNN",
          "Hexagon",
          "CUDA",
          "AMD ZenDNN",
          "OpenVINO",
          "WebGPU"
        ],
        "created_at": "2026-03-29T19:53:14Z",
        "updated_at": "2026-08-23T14:13:27Z",
        "closed_at": null,
        "comments": 15,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21160",
        "categories": [
          "native_backends",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 21161,
        "kind": "pr",
        "state": "open",
        "title": "cpp: Adding new arch RUGPT3XL",
        "labels": [
          "model",
          "python",
          "conversion"
        ],
        "created_at": "2026-03-29T19:55:57Z",
        "updated_at": "2026-07-25T21:46:02Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21161",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 21170,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cuda: fix ROCm multi-GPU illegal memory access in recurrent state restore",
        "labels": [
          "Nvidia GPU",
          "ggml"
        ],
        "created_at": "2026-03-30T03:43:03Z",
        "updated_at": "2026-05-10T09:15:25Z",
        "closed_at": null,
        "comments": 13,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21170",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 21174,
        "kind": "pr",
        "state": "open",
        "title": "server: improve Responses API compliance and Codex CLI compatibility",
        "labels": [
          "documentation",
          "testing",
          "examples",
          "python",
          "server"
        ],
        "created_at": "2026-03-30T07:54:24Z",
        "updated_at": "2026-07-15T11:02:40Z",
        "closed_at": null,
        "comments": 56,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21174",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 21187,
        "kind": "pr",
        "state": "open",
        "title": "llama-server: translating structured generation request parameters from responses API format to completions API format",
        "labels": [
          "examples",
          "python",
          "server"
        ],
        "created_at": "2026-03-30T17:02:40Z",
        "updated_at": "2026-04-09T20:32:57Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21187",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "loader_conversion",
          "api_ui_docs"
        ]
      },
      {
        "number": 21218,
        "kind": "pr",
        "state": "open",
        "title": "CANN: split RoPE cache init into host and device phases",
        "labels": [
          "ggml",
          "Ascend NPU"
        ],
        "created_at": "2026-03-31T08:43:11Z",
        "updated_at": "2026-03-31T09:34:42Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21218",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "build_platform_ops"
        ]
      },
      {
        "number": 21221,
        "kind": "pr",
        "state": "open",
        "title": "kv-cache: add LLAMA_STATE_SEQ_FLAGS_APPEND for incremental state restore",
        "labels": [],
        "created_at": "2026-03-31T10:55:34Z",
        "updated_at": "2026-03-31T17:25:50Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21221",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "build_platform_ops"
        ]
      },
      {
        "number": 21231,
        "kind": "pr",
        "state": "open",
        "title": "server: add router device memory margin parameter for dynamic unloading",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2026-03-31T14:40:58Z",
        "updated_at": "2026-08-22T16:32:37Z",
        "closed_at": null,
        "comments": 27,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21231",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "api_ui_docs"
        ]
      },
      {
        "number": 21244,
        "kind": "pr",
        "state": "open",
        "title": "gguf-py: add type validation to GGUFWriter.add_key_value",
        "labels": [
          "python"
        ],
        "created_at": "2026-04-01T03:01:43Z",
        "updated_at": "2026-04-01T08:29:59Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21244",
        "categories": [
          "quantization",
          "models_architectures",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 21265,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Memory leak on RPC CUDA backend",
        "labels": [
          "bug"
        ],
        "created_at": "2026-04-01T14:10:09Z",
        "updated_at": "2026-07-15T10:42:18Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/21265",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 21266,
        "kind": "issue",
        "state": "open",
        "title": "server : disaggregated prefill/decode support",
        "labels": [
          "enhancement",
          "server",
          "roadmap"
        ],
        "created_at": "2026-04-01T14:28:16Z",
        "updated_at": "2026-08-17T20:28:52Z",
        "closed_at": null,
        "comments": 33,
        "url": "https://github.com/ggml-org/llama.cpp/issues/21266",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 21282,
        "kind": "pr",
        "state": "open",
        "title": "llama-bench: Fix to reduce very high ± variability",
        "labels": [
          "examples"
        ],
        "created_at": "2026-04-01T23:55:08Z",
        "updated_at": "2026-04-03T08:36:38Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21282",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 21284,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Inefficient defaults for gfx1151 cost substantial performance for prefill (ROCm)",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-04-02T00:24:28Z",
        "updated_at": "2026-08-19T02:07:51Z",
        "closed_at": null,
        "comments": 24,
        "url": "https://github.com/ggml-org/llama.cpp/issues/21284",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 21308,
        "kind": "pr",
        "state": "open",
        "title": "Integration tests for multi-turn tool use",
        "labels": [
          "examples",
          "python",
          "server"
        ],
        "created_at": "2026-04-02T14:32:56Z",
        "updated_at": "2026-05-21T13:02:06Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21308",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 21397,
        "kind": "pr",
        "state": "open",
        "title": "fix tensor stride for quantized types in create_tensor",
        "labels": [],
        "created_at": "2026-04-03T22:40:15Z",
        "updated_at": "2026-04-06T13:20:14Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21397",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "multimodal_embedding",
          "correctness_testing"
        ]
      },
      {
        "number": 21407,
        "kind": "pr",
        "state": "open",
        "title": "cmake: add flag to use system httplib",
        "labels": [
          "build",
          "vendor"
        ],
        "created_at": "2026-04-04T03:21:39Z",
        "updated_at": "2026-08-20T08:05:19Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21407",
        "categories": [
          "models_architectures",
          "build_platform_ops"
        ]
      },
      {
        "number": 21412,
        "kind": "pr",
        "state": "open",
        "title": "model: add Zamba2 architecture support",
        "labels": [
          "model",
          "python",
          "conversion"
        ],
        "created_at": "2026-04-04T04:58:02Z",
        "updated_at": "2026-08-24T00:30:33Z",
        "closed_at": null,
        "comments": 23,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21412",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 21433,
        "kind": "pr",
        "state": "open",
        "title": "fix(gemma4): handle nullable type arrays",
        "labels": [
          "testing"
        ],
        "created_at": "2026-04-04T13:55:00Z",
        "updated_at": "2026-04-05T17:43:46Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21433",
        "categories": [
          "native_backends",
          "models_architectures",
          "structured_generation_tools",
          "correctness_testing"
        ]
      },
      {
        "number": 21443,
        "kind": "pr",
        "state": "open",
        "title": "metal: speed up Qwen3-VL image encoding on large images by ~11%",
        "labels": [
          "ggml",
          "Apple Metal"
        ],
        "created_at": "2026-04-04T19:12:46Z",
        "updated_at": "2026-06-25T02:22:15Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21443",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 21452,
        "kind": "pr",
        "state": "open",
        "title": "metal : add GATED_LINEAR_ATTN op",
        "labels": [
          "documentation",
          "testing",
          "ggml",
          "Apple Metal"
        ],
        "created_at": "2026-04-05T00:58:12Z",
        "updated_at": "2026-07-02T13:09:53Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21452",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 21458,
        "kind": "pr",
        "state": "open",
        "title": "ggml : add GGML_OP_GATHER for DeepSeek Sparse Attention (DSA) #21149",
        "labels": [
          "testing",
          "ggml"
        ],
        "created_at": "2026-04-05T03:55:36Z",
        "updated_at": "2026-04-05T20:36:31Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21458",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 21467,
        "kind": "pr",
        "state": "open",
        "title": "ggml : fix repeat_back assert with non-contiguous gradients",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-04-05T11:25:20Z",
        "updated_at": "2026-04-07T18:51:40Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21467",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "training_adapters"
        ]
      },
      {
        "number": 21475,
        "kind": "pr",
        "state": "open",
        "title": "llama-quant: use LLM_KV constants instead of hardcoded strings",
        "labels": [],
        "created_at": "2026-04-05T15:28:38Z",
        "updated_at": "2026-04-05T15:32:17Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21475",
        "categories": [
          "quantization",
          "models_architectures"
        ]
      },
      {
        "number": 21476,
        "kind": "pr",
        "state": "open",
        "title": "gguf-py: Fix lazy tensor handling for keyword arguments",
        "labels": [
          "python"
        ],
        "created_at": "2026-04-05T16:21:33Z",
        "updated_at": "2026-05-04T07:02:31Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21476",
        "categories": [
          "quantization",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 21477,
        "kind": "pr",
        "state": "open",
        "title": "server: add null check for context to prevent segfault on init failure",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2026-04-05T16:54:46Z",
        "updated_at": "2026-04-19T19:39:58Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21477",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 21507,
        "kind": "pr",
        "state": "open",
        "title": "llama-quant : overlap compute and write with double buffering",
        "labels": [],
        "created_at": "2026-04-06T09:04:41Z",
        "updated_at": "2026-05-04T23:58:58Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21507",
        "categories": [
          "quantization",
          "models_architectures",
          "correctness_testing"
        ]
      },
      {
        "number": 21537,
        "kind": "pr",
        "state": "open",
        "title": "server : fix json_schema response_format ignored by some chat templates",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2026-04-07T00:27:13Z",
        "updated_at": "2026-06-17T14:40:47Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21537",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 21545,
        "kind": "issue",
        "state": "open",
        "title": "server: Add model capabilities information (instruction, embedding etc.)",
        "labels": [
          "server/webui"
        ],
        "created_at": "2026-04-07T08:22:39Z",
        "updated_at": "2026-08-05T01:11:02Z",
        "closed_at": null,
        "comments": 11,
        "url": "https://github.com/ggml-org/llama.cpp/issues/21545",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "api_ui_docs"
        ]
      },
      {
        "number": 21547,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Gemma4 fails with NotImplemented: map: filter-mapping not implemented",
        "labels": [
          "enhancement",
          "good first issue"
        ],
        "created_at": "2026-04-07T08:44:26Z",
        "updated_at": "2026-05-24T10:10:28Z",
        "closed_at": null,
        "comments": 9,
        "url": "https://github.com/ggml-org/llama.cpp/issues/21547",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing"
        ]
      },
      {
        "number": 21551,
        "kind": "pr",
        "state": "open",
        "title": "common : only warn when using unsupported kv cache types",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-04-07T10:09:53Z",
        "updated_at": "2026-07-15T14:03:17Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21551",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 21556,
        "kind": "pr",
        "state": "open",
        "title": "Add GCN architecture detection fix for MoltenVK (AMD eGPU on macOS)",
        "labels": [
          "Vulkan",
          "ggml"
        ],
        "created_at": "2026-04-07T11:42:26Z",
        "updated_at": "2026-04-08T05:15:49Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21556",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing"
        ]
      },
      {
        "number": 21560,
        "kind": "pr",
        "state": "open",
        "title": "feat/verifiable inference",
        "labels": [
          "examples",
          "python",
          "server"
        ],
        "created_at": "2026-04-07T14:13:13Z",
        "updated_at": "2026-04-07T18:15:32Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21560",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 21563,
        "kind": "pr",
        "state": "open",
        "title": "optionally enable ccache for use in Dockerfiles",
        "labels": [
          "devops"
        ],
        "created_at": "2026-04-07T14:23:36Z",
        "updated_at": "2026-06-30T23:12:13Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21563",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 21575,
        "kind": "issue",
        "state": "open",
        "title": "Research: Tokenizer with faster string search",
        "labels": [
          "research 🔬"
        ],
        "created_at": "2026-04-07T18:38:32Z",
        "updated_at": "2026-07-24T18:43:16Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/21575",
        "categories": [
          "native_backends",
          "models_architectures",
          "loader_conversion",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 21576,
        "kind": "pr",
        "state": "open",
        "title": " kv-cache : fix crash in state save/restore",
        "labels": [],
        "created_at": "2026-04-07T19:25:13Z",
        "updated_at": "2026-07-12T20:36:25Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21576",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 21587,
        "kind": "pr",
        "state": "open",
        "title": "vocab : fix Gemma 4 BPE tokenizer SIGSEGV on long prompts without newlines",
        "labels": [
          "testing",
          "examples",
          "server"
        ],
        "created_at": "2026-04-07T21:15:08Z",
        "updated_at": "2026-04-08T00:50:46Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21587",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 21652,
        "kind": "pr",
        "state": "open",
        "title": "Prevent the sum of the dequantized activation in q8_1 from overflowing",
        "labels": [
          "Nvidia GPU",
          "ggml"
        ],
        "created_at": "2026-04-09T00:51:57Z",
        "updated_at": "2026-05-01T21:44:03Z",
        "closed_at": null,
        "comments": 28,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21652",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 21659,
        "kind": "pr",
        "state": "open",
        "title": "common: load parser in common_chat_parser_params constructor",
        "labels": [],
        "created_at": "2026-04-09T05:03:02Z",
        "updated_at": "2026-04-09T05:17:21Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21659",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing"
        ]
      },
      {
        "number": 21686,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: tensor parallelism failing with -nkvo (Gemma 31B)",
        "labels": [
          "bug",
          "CUDA"
        ],
        "created_at": "2026-04-09T15:38:03Z",
        "updated_at": "2026-04-10T09:24:21Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/21686",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 21691,
        "kind": "pr",
        "state": "open",
        "title": "debug: functionality to dump full tensors and compare",
        "labels": [
          "examples",
          "python"
        ],
        "created_at": "2026-04-09T18:38:55Z",
        "updated_at": "2026-04-15T18:48:34Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21691",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 21695,
        "kind": "pr",
        "state": "open",
        "title": "server: fix SWA prompt reuse boundary condition",
        "labels": [
          "Nvidia GPU",
          "examples",
          "server",
          "ggml"
        ],
        "created_at": "2026-04-09T20:58:56Z",
        "updated_at": "2026-04-15T11:43:36Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21695",
        "categories": [
          "native_backends",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 21698,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cuda: better vram to lds loading pipeline in load_tiles_q8_0",
        "labels": [
          "Nvidia GPU",
          "ggml"
        ],
        "created_at": "2026-04-09T22:13:01Z",
        "updated_at": "2026-04-14T09:00:10Z",
        "closed_at": null,
        "comments": 12,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21698",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "loader_conversion",
          "correctness_testing"
        ]
      },
      {
        "number": 21710,
        "kind": "pr",
        "state": "open",
        "title": "tests: Fix unit test failing when built with ICX on Windows",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-04-10T06:00:08Z",
        "updated_at": "2026-07-06T08:34:25Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21710",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 21717,
        "kind": "pr",
        "state": "open",
        "title": "TP: fix arbitrary -ot",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-04-10T09:23:29Z",
        "updated_at": "2026-05-26T19:32:09Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21717",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving"
        ]
      },
      {
        "number": 21723,
        "kind": "issue",
        "state": "open",
        "title": "Refactor: Split up ggml-vulkan.cpp",
        "labels": [
          "enhancement",
          "Vulkan"
        ],
        "created_at": "2026-04-10T12:55:51Z",
        "updated_at": "2026-08-05T15:47:40Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/issues/21723",
        "categories": [
          "native_backends",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 21725,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: XDNA backend",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-04-10T13:25:58Z",
        "updated_at": "2026-08-18T21:45:27Z",
        "closed_at": null,
        "comments": 24,
        "url": "https://github.com/ggml-org/llama.cpp/issues/21725",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 21729,
        "kind": "pr",
        "state": "open",
        "title": "fix rerank  value error  (Add token_type_ids input for rerank model with type-embedding)",
        "labels": [
          "model",
          "examples",
          "python",
          "server"
        ],
        "created_at": "2026-04-10T14:21:57Z",
        "updated_at": "2026-04-15T02:58:08Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21729",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 21752,
        "kind": "pr",
        "state": "open",
        "title": "common: honor HTTP_PROXY/HTTPS_PROXY env vars in http client",
        "labels": [],
        "created_at": "2026-04-11T01:06:06Z",
        "updated_at": "2026-06-18T15:41:13Z",
        "closed_at": null,
        "comments": 17,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21752",
        "categories": [
          "models_architectures",
          "build_platform_ops"
        ]
      },
      {
        "number": 21757,
        "kind": "pr",
        "state": "open",
        "title": "kv : add dynamic KV cache resize (--kv-dynamic)",
        "labels": [],
        "created_at": "2026-04-11T05:48:28Z",
        "updated_at": "2026-04-24T15:02:17Z",
        "closed_at": null,
        "comments": 12,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21757",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 21761,
        "kind": "pr",
        "state": "open",
        "title": "CUDA: only init NCCL for setups with multi GPU",
        "labels": [
          "Nvidia GPU",
          "ggml"
        ],
        "created_at": "2026-04-11T07:24:18Z",
        "updated_at": "2026-04-15T09:54:22Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21761",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing"
        ]
      },
      {
        "number": 21771,
        "kind": "issue",
        "state": "open",
        "title": "Qwen3 TAG_WITH_TAGGED tool format: p.json() fails on array<object> parameter values; partial tool_call leaked to client poisons multi-turn history",
        "labels": [
          "bug",
          "regression",
          "chat parser"
        ],
        "created_at": "2026-04-11T16:20:05Z",
        "updated_at": "2026-04-19T10:30:22Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/issues/21771",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 21779,
        "kind": "issue",
        "state": "open",
        "title": "(Planning) add model management API to server",
        "labels": [],
        "created_at": "2026-04-11T22:09:48Z",
        "updated_at": "2026-08-01T01:14:58Z",
        "closed_at": null,
        "comments": 9,
        "url": "https://github.com/ggml-org/llama.cpp/issues/21779",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "loader_conversion",
          "api_ui_docs"
        ]
      },
      {
        "number": 21782,
        "kind": "pr",
        "state": "open",
        "title": "ggml-metal: add Metal kernel for ggml_roll",
        "labels": [
          "ggml",
          "Apple Metal"
        ],
        "created_at": "2026-04-11T23:49:11Z",
        "updated_at": "2026-04-12T12:21:14Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21782",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 21792,
        "kind": "pr",
        "state": "open",
        "title": "kv: Add optional mmap kv cache",
        "labels": [
          "testing",
          "examples"
        ],
        "created_at": "2026-04-12T05:34:30Z",
        "updated_at": "2026-04-13T02:35:15Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21792",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 21794,
        "kind": "pr",
        "state": "open",
        "title": "llama-bench: fix accumulated load_time in perf timings",
        "labels": [
          "examples"
        ],
        "created_at": "2026-04-12T06:03:27Z",
        "updated_at": "2026-04-12T15:50:50Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21794",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "api_ui_docs"
        ]
      },
      {
        "number": 21815,
        "kind": "pr",
        "state": "open",
        "title": "server : reinit speculative ngram state after context shift to fix GGML_ABORT",
        "labels": [
          "testing",
          "examples",
          "server"
        ],
        "created_at": "2026-04-12T15:51:51Z",
        "updated_at": "2026-04-12T18:26:37Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21815",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 21819,
        "kind": "pr",
        "state": "open",
        "title": "fix(poetry): update python version",
        "labels": [],
        "created_at": "2026-04-12T17:40:46Z",
        "updated_at": "2026-04-12T17:53:23Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21819",
        "categories": [
          "models_architectures",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 21828,
        "kind": "pr",
        "state": "open",
        "title": "ggml-rpc: fix 32-bit ARM (ILP32) serialization bugs",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-04-12T23:25:31Z",
        "updated_at": "2026-04-12T23:29:38Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21828",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 21831,
        "kind": "issue",
        "state": "open",
        "title": "Server forces full prompt re-processing on subsequent requests (SWA/recurrent memory error)",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-04-13T00:08:36Z",
        "updated_at": "2026-08-15T20:28:39Z",
        "closed_at": null,
        "comments": 52,
        "url": "https://github.com/ggml-org/llama.cpp/issues/21831",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 21849,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cuda: make mul_mat_q tile selection tunable per-type and per-arch",
        "labels": [
          "testing",
          "Nvidia GPU",
          "ggml"
        ],
        "created_at": "2026-04-13T10:47:10Z",
        "updated_at": "2026-04-16T08:41:29Z",
        "closed_at": null,
        "comments": 10,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21849",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 21890,
        "kind": "pr",
        "state": "open",
        "title": "server: add --parallel-tool-calling flag to enable by default",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2026-04-14T11:36:42Z",
        "updated_at": "2026-04-22T00:32:46Z",
        "closed_at": null,
        "comments": 9,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21890",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 21897,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cuda: enable concurrent streams for linear attention",
        "labels": [
          "Nvidia GPU",
          "ggml"
        ],
        "created_at": "2026-04-14T13:11:37Z",
        "updated_at": "2026-05-11T08:02:41Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21897",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 21904,
        "kind": "pr",
        "state": "open",
        "title": "feat: add MPNet model architecture",
        "labels": [
          "model",
          "python"
        ],
        "created_at": "2026-04-14T14:04:10Z",
        "updated_at": "2026-05-06T07:05:30Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21904",
        "categories": [
          "models_architectures",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 21923,
        "kind": "pr",
        "state": "open",
        "title": "common : add --hf-prune-old-files (-hfp) parameter to automatically delete outdated HF files",
        "labels": [],
        "created_at": "2026-04-14T21:28:12Z",
        "updated_at": "2026-05-29T17:17:13Z",
        "closed_at": null,
        "comments": 12,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21923",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 21924,
        "kind": "pr",
        "state": "open",
        "title": "fix: llama-finetune backward pass crashes",
        "labels": [
          "examples",
          "ggml"
        ],
        "created_at": "2026-04-14T21:35:46Z",
        "updated_at": "2026-04-15T14:55:45Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21924",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 21928,
        "kind": "pr",
        "state": "open",
        "title": "nix: support unified apple-sdk",
        "labels": [
          "nix",
          "devops"
        ],
        "created_at": "2026-04-14T23:30:21Z",
        "updated_at": "2026-04-18T19:32:10Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21928",
        "categories": [
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 21931,
        "kind": "pr",
        "state": "open",
        "title": "gguf-py: add type and range validation to GGUFWriter.add_key_value",
        "labels": [
          "python"
        ],
        "created_at": "2026-04-15T03:21:49Z",
        "updated_at": "2026-04-15T03:22:07Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21931",
        "categories": [
          "quantization",
          "models_architectures",
          "loader_conversion",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 21956,
        "kind": "issue",
        "state": "open",
        "title": "(Planning) Support audio output in mtmd",
        "labels": [],
        "created_at": "2026-04-15T15:08:26Z",
        "updated_at": "2026-07-28T23:22:44Z",
        "closed_at": null,
        "comments": 25,
        "url": "https://github.com/ggml-org/llama.cpp/issues/21956",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "api_ui_docs"
        ]
      },
      {
        "number": 21958,
        "kind": "pr",
        "state": "open",
        "title": "Add SVE for simd-gemm.h",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-04-15T15:40:58Z",
        "updated_at": "2026-06-06T01:29:39Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21958",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "structured_generation_tools"
        ]
      },
      {
        "number": 21968,
        "kind": "pr",
        "state": "open",
        "title": "feat: CUDA 10.2 / C++14 compatibility for Jetson TX2 (compute 6.2)",
        "labels": [
          "Nvidia GPU",
          "ggml"
        ],
        "created_at": "2026-04-15T20:35:45Z",
        "updated_at": "2026-04-15T21:47:23Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21968",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 21969,
        "kind": "pr",
        "state": "open",
        "title": "Update CMakeLists.txt",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-04-15T22:02:05Z",
        "updated_at": "2026-04-15T23:27:08Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21969",
        "categories": [
          "native_backends",
          "models_architectures",
          "build_platform_ops"
        ]
      },
      {
        "number": 21972,
        "kind": "pr",
        "state": "open",
        "title": "Use/support modern nixpkgs on Darwin (#21381)",
        "labels": [
          "nix",
          "devops"
        ],
        "created_at": "2026-04-15T22:56:00Z",
        "updated_at": "2026-04-26T07:48:30Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21972",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "multimodal_embedding",
          "build_platform_ops"
        ]
      },
      {
        "number": 21977,
        "kind": "pr",
        "state": "open",
        "title": "cmake: fix persistent ARM CPU detection warning on Clang/macOS",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-04-16T02:46:30Z",
        "updated_at": "2026-04-16T06:33:29Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/21977",
        "categories": [
          "models_architectures",
          "loader_conversion",
          "build_platform_ops"
        ]
      },
      {
        "number": 22007,
        "kind": "pr",
        "state": "open",
        "title": "sampling: add segment-level repetition loop detection",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2026-04-16T18:19:08Z",
        "updated_at": "2026-06-04T19:32:03Z",
        "closed_at": null,
        "comments": 13,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22007",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 22022,
        "kind": "pr",
        "state": "open",
        "title": "llama-mmap: add MADV_HUGEPAGE hint for THP on Linux",
        "labels": [],
        "created_at": "2026-04-16T23:37:12Z",
        "updated_at": "2026-04-19T00:16:48Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22022",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 22055,
        "kind": "pr",
        "state": "open",
        "title": "spec : save the dynamic ngram cache file",
        "labels": [],
        "created_at": "2026-04-17T16:58:05Z",
        "updated_at": "2026-08-23T15:08:58Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22055",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "speculative_generation",
          "build_platform_ops"
        ]
      },
      {
        "number": 22059,
        "kind": "pr",
        "state": "open",
        "title": "GGML: Allow static build with dynamic loaded backends",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-04-17T17:40:11Z",
        "updated_at": "2026-04-17T18:25:29Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22059",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 22061,
        "kind": "pr",
        "state": "open",
        "title": "Extend LoRA hotswapping support",
        "labels": [
          "examples",
          "python",
          "server"
        ],
        "created_at": "2026-04-17T18:10:48Z",
        "updated_at": "2026-04-17T18:31:16Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22061",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 22081,
        "kind": "pr",
        "state": "open",
        "title": "server: always include usage in streaming responses",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2026-04-18T12:12:29Z",
        "updated_at": "2026-04-18T12:16:30Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22081",
        "categories": [
          "native_backends",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 22083,
        "kind": "pr",
        "state": "open",
        "title": "server : disable similarity slot selection with --cache-idle-slots and --parallel 1",
        "labels": [
          "examples",
          "python",
          "server"
        ],
        "created_at": "2026-04-18T14:09:38Z",
        "updated_at": "2026-04-20T21:44:08Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22083",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "api_ui_docs"
        ]
      },
      {
        "number": 22086,
        "kind": "pr",
        "state": "open",
        "title": "hexagon: hmx opt phase2",
        "labels": [
          "ggml",
          "Hexagon"
        ],
        "created_at": "2026-04-18T15:42:47Z",
        "updated_at": "2026-05-02T10:38:04Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22086",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 22098,
        "kind": "pr",
        "state": "open",
        "title": "[SYCL] Add Zero-Copy path with Cache Flushing for Intel UMA (Lunar Lake/Meteor Lake)",
        "labels": [
          "ggml",
          "SYCL"
        ],
        "created_at": "2026-04-18T20:14:46Z",
        "updated_at": "2026-05-16T18:49:18Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22098",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 22112,
        "kind": "pr",
        "state": "open",
        "title": "CUDA: PoC for repacking mxfp4",
        "labels": [
          "Nvidia GPU",
          "ggml"
        ],
        "created_at": "2026-04-19T07:47:17Z",
        "updated_at": "2026-04-19T09:48:07Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22112",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 22113,
        "kind": "pr",
        "state": "open",
        "title": "ggml: kleidi, cpu: beginnings of macOS cluster scheduling",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-04-19T08:25:23Z",
        "updated_at": "2026-07-09T20:52:10Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22113",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 22121,
        "kind": "pr",
        "state": "open",
        "title": "Make `llama_token_to_piece` return `INT32_MIN` for invalid tokens",
        "labels": [
          "testing",
          "examples",
          "python"
        ],
        "created_at": "2026-04-19T15:23:51Z",
        "updated_at": "2026-05-26T11:40:03Z",
        "closed_at": null,
        "comments": 15,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22121",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 22126,
        "kind": "pr",
        "state": "open",
        "title": "Gbnf - literal token filtering for non-control tokens.",
        "labels": [
          "testing"
        ],
        "created_at": "2026-04-19T17:35:51Z",
        "updated_at": "2026-04-27T18:37:22Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22126",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 22133,
        "kind": "pr",
        "state": "open",
        "title": "ggml: count new split inputs before appending node (fixes #21730)",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-04-19T21:22:06Z",
        "updated_at": "2026-05-15T19:02:31Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22133",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 22134,
        "kind": "pr",
        "state": "open",
        "title": "ggml : vectorize Q6_K unpack on WASM SIMD128 (strict, deterministic)",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-04-19T21:47:57Z",
        "updated_at": "2026-04-19T21:51:52Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22134",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 22146,
        "kind": "pr",
        "state": "open",
        "title": "server : add missing return after validation errors in /infill endpoint",
        "labels": [
          "examples",
          "python",
          "server"
        ],
        "created_at": "2026-04-20T06:41:44Z",
        "updated_at": "2026-04-20T08:51:53Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22146",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 22178,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Dynamic WebUI System Prompt Content (ex: ability to add dynamic date)",
        "labels": [
          "enhancement",
          "stale"
        ],
        "created_at": "2026-04-20T16:22:34Z",
        "updated_at": "2026-08-22T01:13:41Z",
        "closed_at": null,
        "comments": 9,
        "url": "https://github.com/ggml-org/llama.cpp/issues/22178",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 22181,
        "kind": "pr",
        "state": "open",
        "title": "Optimize reduction stage of dot product of q4_L/q5_K to q8_K on AVX2",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-04-20T19:10:42Z",
        "updated_at": "2026-04-20T21:26:32Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22181",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 22193,
        "kind": "pr",
        "state": "open",
        "title": "cuda: add partial eviction on pool OOM",
        "labels": [
          "Nvidia GPU",
          "ggml"
        ],
        "created_at": "2026-04-21T03:44:48Z",
        "updated_at": "2026-07-13T06:32:26Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22193",
        "categories": [
          "native_backends",
          "quantization",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 22197,
        "kind": "issue",
        "state": "open",
        "title": "ggml-backend-meta: multi buffers are unsupported leading to vulkan segfault",
        "labels": [
          "bug"
        ],
        "created_at": "2026-04-21T04:39:22Z",
        "updated_at": "2026-08-18T18:40:58Z",
        "closed_at": null,
        "comments": 9,
        "url": "https://github.com/ggml-org/llama.cpp/issues/22197",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 22212,
        "kind": "pr",
        "state": "open",
        "title": "ggml: add initial MetaX backend integration",
        "labels": [
          "documentation",
          "Nvidia GPU",
          "python",
          "ggml"
        ],
        "created_at": "2026-04-21T12:41:06Z",
        "updated_at": "2026-08-10T02:52:19Z",
        "closed_at": null,
        "comments": 40,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22212",
        "categories": [
          "native_backends",
          "models_architectures",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 22250,
        "kind": "pr",
        "state": "open",
        "title": "Q5_0 - Block Interleaving Implementation for x86 SIMD (AVX512/AVX2)",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-04-22T13:28:56Z",
        "updated_at": "2026-04-27T11:34:39Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22250",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 22251,
        "kind": "pr",
        "state": "open",
        "title": "ggml-backend: expose multi-buffer accessors and make is_multi_buffer NULL-safe",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-04-22T14:19:20Z",
        "updated_at": "2026-04-30T09:34:59Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22251",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 22270,
        "kind": "pr",
        "state": "open",
        "title": "readme : add D bindings (llama-cpp-d)",
        "labels": [],
        "created_at": "2026-04-23T02:56:03Z",
        "updated_at": "2026-04-23T02:56:03Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22270",
        "categories": [
          "models_architectures",
          "api_ui_docs"
        ]
      },
      {
        "number": 22277,
        "kind": "pr",
        "state": "open",
        "title": "llama : integer type consistency in llama.h #4574",
        "labels": [],
        "created_at": "2026-04-23T09:10:28Z",
        "updated_at": "2026-04-23T09:13:28Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22277",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "structured_generation_tools"
        ]
      },
      {
        "number": 22282,
        "kind": "pr",
        "state": "open",
        "title": "Update CMakeLists.txt",
        "labels": [
          "need more info",
          "ggml"
        ],
        "created_at": "2026-04-23T11:32:41Z",
        "updated_at": "2026-04-23T13:38:23Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22282",
        "categories": [
          "models_architectures",
          "build_platform_ops"
        ]
      },
      {
        "number": 22284,
        "kind": "pr",
        "state": "open",
        "title": "server: router fix model unload reload deadlock",
        "labels": [
          "examples",
          "python",
          "server"
        ],
        "created_at": "2026-04-23T11:42:30Z",
        "updated_at": "2026-04-23T11:42:47Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22284",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 22325,
        "kind": "pr",
        "state": "open",
        "title": "Fix gemma4 prefill parsing",
        "labels": [],
        "created_at": "2026-04-24T15:29:11Z",
        "updated_at": "2026-04-24T16:47:13Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22325",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 22326,
        "kind": "pr",
        "state": "open",
        "title": "Update README.md to add AI Playground to the UI list",
        "labels": [],
        "created_at": "2026-04-24T16:01:35Z",
        "updated_at": "2026-04-24T16:02:29Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22326",
        "categories": [
          "native_backends",
          "models_architectures",
          "multimodal_embedding",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 22329,
        "kind": "pr",
        "state": "open",
        "title": "llama-bench: fix numerical instability in stdev() calculation",
        "labels": [
          "examples"
        ],
        "created_at": "2026-04-24T19:02:22Z",
        "updated_at": "2026-04-24T19:06:59Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22329",
        "categories": [
          "models_architectures",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 22331,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cpu: optimize q8 quantization on x86 SIMD",
        "labels": [
          "testing",
          "ggml"
        ],
        "created_at": "2026-04-24T19:29:30Z",
        "updated_at": "2026-04-24T20:12:11Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22331",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 22336,
        "kind": "pr",
        "state": "open",
        "title": "server: respect per-request enable_thinking toggle via extra_body",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2026-04-24T21:20:31Z",
        "updated_at": "2026-04-26T14:33:01Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22336",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 22348,
        "kind": "pr",
        "state": "open",
        "title": "fix: architecture-aware n_embd read before vocab_only early return for mmproj init",
        "labels": [],
        "created_at": "2026-04-25T08:28:49Z",
        "updated_at": "2026-08-11T07:26:53Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22348",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 22350,
        "kind": "pr",
        "state": "open",
        "title": "rpc: add ipv6 support",
        "labels": [
          "examples",
          "ggml"
        ],
        "created_at": "2026-04-25T10:49:55Z",
        "updated_at": "2026-04-25T11:36:53Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22350",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "api_ui_docs"
        ]
      },
      {
        "number": 22360,
        "kind": "issue",
        "state": "open",
        "title": "Bug: Server Slows Down Significantly Over Time, Requires Frequent Reboots",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-04-25T16:35:10Z",
        "updated_at": "2026-08-17T01:07:42Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/issues/22360",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 22366,
        "kind": "pr",
        "state": "open",
        "title": "Convert argv from UTF-16 on Windows for non-ASCII -p prompts",
        "labels": [
          "examples"
        ],
        "created_at": "2026-04-25T17:55:10Z",
        "updated_at": "2026-04-25T17:56:13Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22366",
        "categories": [
          "models_architectures",
          "loader_conversion",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 22385,
        "kind": "pr",
        "state": "open",
        "title": "Windows: raise stdio limit for loading many GGUF shards",
        "labels": [],
        "created_at": "2026-04-26T07:15:37Z",
        "updated_at": "2026-04-26T07:46:37Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22385",
        "categories": [
          "quantization",
          "models_architectures",
          "loader_conversion",
          "build_platform_ops"
        ]
      },
      {
        "number": 22432,
        "kind": "pr",
        "state": "open",
        "title": "ngram parameters",
        "labels": [
          "testing",
          "examples",
          "server"
        ],
        "created_at": "2026-04-27T11:18:26Z",
        "updated_at": "2026-04-27T11:18:43Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22432",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 22437,
        "kind": "pr",
        "state": "open",
        "title": "server: guard against reranking with embeddings and add regression tests",
        "labels": [
          "examples",
          "python",
          "server"
        ],
        "created_at": "2026-04-27T16:08:31Z",
        "updated_at": "2026-04-27T16:09:48Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22437",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 22440,
        "kind": "pr",
        "state": "open",
        "title": "llama: use int32_t instead of int in llama.h",
        "labels": [],
        "created_at": "2026-04-27T16:42:02Z",
        "updated_at": "2026-05-09T21:27:30Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22440",
        "categories": [
          "models_architectures"
        ]
      },
      {
        "number": 22448,
        "kind": "pr",
        "state": "open",
        "title": "CUDA: use float instead of int64_t for MMQ bounds",
        "labels": [
          "Nvidia GPU",
          "ggml"
        ],
        "created_at": "2026-04-27T21:42:04Z",
        "updated_at": "2026-04-28T07:25:35Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22448",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 22449,
        "kind": "pr",
        "state": "open",
        "title": "[codex] add chat template conformance tests",
        "labels": [
          "testing"
        ],
        "created_at": "2026-04-27T21:49:37Z",
        "updated_at": "2026-07-09T23:34:00Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22449",
        "categories": [
          "models_architectures",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 22460,
        "kind": "pr",
        "state": "open",
        "title": "fix: use cpu_get_num_math() instead of hardware_concurrency() for --threads",
        "labels": [],
        "created_at": "2026-04-28T05:54:38Z",
        "updated_at": "2026-04-29T04:14:58Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22460",
        "categories": [
          "scheduling_serving",
          "structured_generation_tools"
        ]
      },
      {
        "number": 22466,
        "kind": "pr",
        "state": "open",
        "title": "WIP: ggml-backend-meta : async pinned upload for -sm tensor model load",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-04-28T10:55:46Z",
        "updated_at": "2026-08-04T15:37:24Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22466",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 22475,
        "kind": "pr",
        "state": "open",
        "title": "server: include api_prefix in public_endpoints set",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2026-04-28T13:17:19Z",
        "updated_at": "2026-06-17T07:56:53Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22475",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 22477,
        "kind": "issue",
        "state": "open",
        "title": "Model Request Hy3",
        "labels": [],
        "created_at": "2026-04-28T14:31:43Z",
        "updated_at": "2026-07-29T17:12:53Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/issues/22477",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "training_adapters"
        ]
      },
      {
        "number": 22482,
        "kind": "pr",
        "state": "open",
        "title": "cmake: use cmake's install(EXPORT) to track ggml target properties",
        "labels": [
          "Nvidia GPU",
          "Vulkan",
          "ggml",
          "SYCL",
          "Apple Metal",
          "OpenCL"
        ],
        "created_at": "2026-04-28T17:25:25Z",
        "updated_at": "2026-05-05T17:38:05Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22482",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 22485,
        "kind": "pr",
        "state": "open",
        "title": "gguf-py: shrink layers or embedding vectors for reducing model size.",
        "labels": [
          "python"
        ],
        "created_at": "2026-04-28T18:52:29Z",
        "updated_at": "2026-04-30T03:15:18Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22485",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 22495,
        "kind": "pr",
        "state": "open",
        "title": "examples : add llama-profiler-cpu/gpu for op roofline measurement",
        "labels": [
          "examples"
        ],
        "created_at": "2026-04-29T02:09:03Z",
        "updated_at": "2026-05-04T22:36:50Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22495",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 22509,
        "kind": "pr",
        "state": "open",
        "title": "nix: added dev shells for more backends and updated flake.lock",
        "labels": [
          "nix",
          "devops"
        ],
        "created_at": "2026-04-29T10:05:57Z",
        "updated_at": "2026-08-16T20:33:42Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22509",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 22515,
        "kind": "pr",
        "state": "open",
        "title": "ggml-metal: implement async 2D tensor copy functions",
        "labels": [
          "ggml",
          "Apple Metal"
        ],
        "created_at": "2026-04-29T11:22:54Z",
        "updated_at": "2026-04-29T15:03:00Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22515",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 22519,
        "kind": "pr",
        "state": "open",
        "title": "Update build.md with commands for `nvidia-smi` in `Override Compute Capability Specifications`",
        "labels": [
          "documentation"
        ],
        "created_at": "2026-04-29T14:19:18Z",
        "updated_at": "2026-05-16T11:29:51Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22519",
        "categories": [
          "native_backends",
          "models_architectures",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 22525,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cpu: optimize ggml_gemm_q4_K_8x8_q8_K interleaving/staging for AVX-512 (and AVX2)",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-04-29T15:09:21Z",
        "updated_at": "2026-05-26T06:39:00Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22525",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 22531,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: allow the web UI to be used with other backends",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-04-29T19:45:32Z",
        "updated_at": "2026-07-27T01:17:08Z",
        "closed_at": null,
        "comments": 9,
        "url": "https://github.com/ggml-org/llama.cpp/issues/22531",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 22564,
        "kind": "pr",
        "state": "open",
        "title": "cmake: fix MATH_LIBRARY check on Windows MSVC",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-04-30T19:52:29Z",
        "updated_at": "2026-05-22T13:05:46Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22564",
        "categories": [
          "native_backends",
          "models_architectures",
          "build_platform_ops"
        ]
      },
      {
        "number": 22566,
        "kind": "pr",
        "state": "open",
        "title": "fix: consistent memory breakdown for models loaded with `no_alloc`",
        "labels": [
          "testing"
        ],
        "created_at": "2026-04-30T21:18:22Z",
        "updated_at": "2026-08-08T05:16:41Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22566",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 22568,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cpu: fix msvc c2440 cast error for m512bh and m256bh in sgemm.cpp",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-04-30T21:54:13Z",
        "updated_at": "2026-05-07T22:44:39Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22568",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "loader_conversion",
          "build_platform_ops"
        ]
      },
      {
        "number": 22569,
        "kind": "pr",
        "state": "open",
        "title": "[Draft] feat: implement paged KV cache and attention",
        "labels": [
          "model",
          "testing",
          "Nvidia GPU",
          "examples",
          "ggml"
        ],
        "created_at": "2026-04-30T23:55:46Z",
        "updated_at": "2026-05-10T12:10:57Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22569",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 22571,
        "kind": "pr",
        "state": "open",
        "title": "Swap out F16 for BF16 in Q8_1 activations to avoid overflowing values",
        "labels": [
          "Nvidia GPU",
          "ggml"
        ],
        "created_at": "2026-05-01T01:51:14Z",
        "updated_at": "2026-05-07T15:13:54Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22571",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 22575,
        "kind": "pr",
        "state": "open",
        "title": "server: support input_image in function_call_output (#20663)",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2026-05-01T06:57:26Z",
        "updated_at": "2026-08-21T23:47:20Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22575",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 22576,
        "kind": "pr",
        "state": "open",
        "title": "Support jina-reranker-v3 cross-encoder architecture",
        "labels": [
          "script",
          "testing",
          "Nvidia GPU",
          "examples",
          "python",
          "devops",
          "server",
          "ggml",
          "OpenCL",
          "Hexagon"
        ],
        "created_at": "2026-05-01T07:13:56Z",
        "updated_at": "2026-05-03T16:03:23Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22576",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 22587,
        "kind": "pr",
        "state": "open",
        "title": "CUDA: row-per-warp kernel for GATED_DELTA_NET",
        "labels": [
          "Nvidia GPU",
          "ggml"
        ],
        "created_at": "2026-05-01T18:01:59Z",
        "updated_at": "2026-07-21T12:25:36Z",
        "closed_at": null,
        "comments": 13,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22587",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "speculative_generation",
          "correctness_testing"
        ]
      },
      {
        "number": 22595,
        "kind": "pr",
        "state": "open",
        "title": "metal : pair rsets_add with rsets_rm on buffer free (fix #22593)",
        "labels": [
          "ggml",
          "Apple Metal"
        ],
        "created_at": "2026-05-01T22:19:35Z",
        "updated_at": "2026-08-19T16:43:54Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22595",
        "categories": [
          "native_backends",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 22606,
        "kind": "pr",
        "state": "open",
        "title": "json-schema: prevent out-of-bounds and stack overflow in regex parsing for GBNF",
        "labels": [],
        "created_at": "2026-05-02T04:17:20Z",
        "updated_at": "2026-05-02T04:17:21Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22606",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 22619,
        "kind": "pr",
        "state": "open",
        "title": "server : fix null task dereference after image processing failure",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2026-05-02T15:07:08Z",
        "updated_at": "2026-05-02T15:12:59Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22619",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 22633,
        "kind": "pr",
        "state": "open",
        "title": "convert: avoid expanded dequant scale tensors",
        "labels": [
          "python"
        ],
        "created_at": "2026-05-02T22:24:00Z",
        "updated_at": "2026-05-03T07:48:21Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22633",
        "categories": [
          "quantization",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 22644,
        "kind": "pr",
        "state": "open",
        "title": "Fix openvino virtual devices",
        "labels": [
          "ggml",
          "OpenVINO"
        ],
        "created_at": "2026-05-03T06:35:05Z",
        "updated_at": "2026-05-03T06:35:44Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22644",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 22648,
        "kind": "issue",
        "state": "open",
        "title": "vulkan tensor parallelism support",
        "labels": [
          "enhancement",
          "Vulkan"
        ],
        "created_at": "2026-05-03T11:59:58Z",
        "updated_at": "2026-08-02T08:42:31Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/issues/22648",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 22661,
        "kind": "pr",
        "state": "open",
        "title": "models: fix linear attention state corruption in recurrent layers",
        "labels": [
          "model"
        ],
        "created_at": "2026-05-04T01:25:03Z",
        "updated_at": "2026-05-05T21:13:28Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22661",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 22671,
        "kind": "pr",
        "state": "open",
        "title": "ggml: Add initial MXFP6 CPU implementation",
        "labels": [
          "testing",
          "Nvidia GPU",
          "python",
          "ggml",
          "CUDA",
          "conversion"
        ],
        "created_at": "2026-05-04T08:15:19Z",
        "updated_at": "2026-08-04T22:10:33Z",
        "closed_at": null,
        "comments": 25,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22671",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 22691,
        "kind": "pr",
        "state": "open",
        "title": "tools: add llama-pshard-plan-params for token-tiered placement planning",
        "labels": [
          "Nvidia GPU",
          "examples",
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-05-04T20:17:00Z",
        "updated_at": "2026-06-20T22:27:59Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22691",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 22692,
        "kind": "pr",
        "state": "open",
        "title": "llama: add pshard runtime for plan switching and streamed weights",
        "labels": [
          "Nvidia GPU",
          "examples",
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-05-04T20:21:42Z",
        "updated_at": "2026-06-20T22:29:41Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22692",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 22704,
        "kind": "pr",
        "state": "open",
        "title": "Feat/backward mul mat",
        "labels": [
          "Nvidia GPU",
          "Vulkan",
          "ggml",
          "Apple Metal"
        ],
        "created_at": "2026-05-05T09:07:41Z",
        "updated_at": "2026-05-23T08:32:29Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22704",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "multimodal_embedding",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 22705,
        "kind": "pr",
        "state": "open",
        "title": "Feat/qlora training",
        "labels": [
          "documentation",
          "testing",
          "Nvidia GPU",
          "Vulkan",
          "examples",
          "python",
          "server",
          "ggml",
          "SYCL",
          "Apple Metal",
          "CUDA",
          "jinja parser",
          "conversion"
        ],
        "created_at": "2026-05-05T09:12:24Z",
        "updated_at": "2026-08-10T07:08:24Z",
        "closed_at": null,
        "comments": 17,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22705",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 22725,
        "kind": "pr",
        "state": "open",
        "title": "Filter tools openai server task",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2026-05-05T15:57:17Z",
        "updated_at": "2026-05-05T18:03:07Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22725",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 22761,
        "kind": "pr",
        "state": "open",
        "title": "server: fix /infill prompt placement after FIM_MID",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2026-05-06T13:30:03Z",
        "updated_at": "2026-05-06T13:32:04Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22761",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 22763,
        "kind": "pr",
        "state": "open",
        "title": "android: extract GgufMetadataReader factory to break cyclic dependency",
        "labels": [
          "android",
          "examples"
        ],
        "created_at": "2026-05-06T14:12:07Z",
        "updated_at": "2026-05-06T14:27:28Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22763",
        "categories": [
          "quantization",
          "models_architectures",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 22764,
        "kind": "pr",
        "state": "open",
        "title": "Draft: ggml-opencl: Early proof-of-concept implementation of plans via command buffers",
        "labels": [
          "ggml",
          "OpenCL"
        ],
        "created_at": "2026-05-06T14:20:04Z",
        "updated_at": "2026-05-06T14:37:49Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22764",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 22780,
        "kind": "pr",
        "state": "open",
        "title": "convert : add `--fuse-qkv` flag to fuse Q/K/V into QKV during HF-to-GGUF conversion",
        "labels": [
          "model",
          "python",
          "conversion"
        ],
        "created_at": "2026-05-07T02:24:04Z",
        "updated_at": "2026-08-26T04:27:06Z",
        "closed_at": null,
        "comments": 16,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22780",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 22790,
        "kind": "pr",
        "state": "open",
        "title": "mtmd-cli: load GPU backends before arg parsing to fix false 'no GPU' warning",
        "labels": [
          "examples"
        ],
        "created_at": "2026-05-07T09:07:11Z",
        "updated_at": "2026-05-08T08:38:44Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22790",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 22802,
        "kind": "pr",
        "state": "open",
        "title": "Add new config file options for saving and loading configuration for llama tools in INI format",
        "labels": [],
        "created_at": "2026-05-07T13:04:26Z",
        "updated_at": "2026-05-13T19:11:16Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22802",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "loader_conversion",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 22819,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: WebUI: Pressing Cancel/Escape after editing the System Message deletes it",
        "labels": [
          "bug"
        ],
        "created_at": "2026-05-07T20:13:52Z",
        "updated_at": "2026-06-15T10:28:54Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/22819",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 22835,
        "kind": "pr",
        "state": "open",
        "title": "convert_hf_to_gguf: fix Qwen3.5 linear_num_value_heads overridden by AutoConfig defaults",
        "labels": [
          "python"
        ],
        "created_at": "2026-05-08T10:56:30Z",
        "updated_at": "2026-05-08T14:22:11Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22835",
        "categories": [
          "quantization",
          "models_architectures",
          "loader_conversion",
          "correctness_testing",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 22836,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cpu : add STQ1_0 ternary quantization with ARM NEON vec_dot kernel",
        "labels": [
          "testing",
          "examples",
          "python",
          "ggml",
          "conversion"
        ],
        "created_at": "2026-05-08T11:07:35Z",
        "updated_at": "2026-08-10T06:19:04Z",
        "closed_at": null,
        "comments": 47,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22836",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 22844,
        "kind": "pr",
        "state": "open",
        "title": "test-backend-ops: add more fields to csv output.",
        "labels": [
          "testing"
        ],
        "created_at": "2026-05-08T16:13:06Z",
        "updated_at": "2026-05-08T18:29:59Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22844",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 22848,
        "kind": "pr",
        "state": "open",
        "title": "cli: Add quiet mode",
        "labels": [
          "examples"
        ],
        "created_at": "2026-05-08T19:02:25Z",
        "updated_at": "2026-05-09T15:38:05Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22848",
        "categories": [
          "models_architectures",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 22852,
        "kind": "pr",
        "state": "open",
        "title": "docs: fix suggested cmake flag in `build.md` for including rocwmma",
        "labels": [
          "documentation"
        ],
        "created_at": "2026-05-08T20:48:46Z",
        "updated_at": "2026-05-17T21:01:16Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22852",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 22854,
        "kind": "pr",
        "state": "open",
        "title": "Add --continue-after-failure support to llama-bench for resilient benchmark sweeps",
        "labels": [
          "examples"
        ],
        "created_at": "2026-05-08T21:29:07Z",
        "updated_at": "2026-05-08T21:39:56Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22854",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "api_ui_docs"
        ]
      },
      {
        "number": 22857,
        "kind": "pr",
        "state": "open",
        "title": "security: fix critical integer overflow (CWE-190) in tensor allocation",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-05-08T23:45:20Z",
        "updated_at": "2026-05-12T22:49:40Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22857",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 22874,
        "kind": "pr",
        "state": "open",
        "title": "cli: exit conversation mode on stdin EOF",
        "labels": [
          "examples"
        ],
        "created_at": "2026-05-09T16:54:06Z",
        "updated_at": "2026-05-09T16:54:21Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22874",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 22892,
        "kind": "pr",
        "state": "open",
        "title": "[ggml] Fix Vulkan-Hpp handle usage on 32-bit targets.",
        "labels": [
          "Vulkan",
          "ggml"
        ],
        "created_at": "2026-05-10T04:06:15Z",
        "updated_at": "2026-05-27T12:07:50Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22892",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 22897,
        "kind": "pr",
        "state": "open",
        "title": "fix(quantize): add NVFP4 default type mapping and scale tensors",
        "labels": [
          "examples"
        ],
        "created_at": "2026-05-10T08:34:05Z",
        "updated_at": "2026-07-09T23:33:13Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22897",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "api_ui_docs"
        ]
      },
      {
        "number": 22921,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Allow independent control of slots (-ns) and parallelism (-np)",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-05-10T20:25:01Z",
        "updated_at": "2026-08-02T16:10:55Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/issues/22921",
        "categories": [
          "models_architectures",
          "scheduling_serving"
        ]
      },
      {
        "number": 22922,
        "kind": "pr",
        "state": "open",
        "title": "common: improve --fit host-memory accounting for CPU and iGPU",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-05-10T21:04:28Z",
        "updated_at": "2026-05-14T16:34:19Z",
        "closed_at": null,
        "comments": 9,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22922",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 22928,
        "kind": "pr",
        "state": "open",
        "title": "kv-cache: use `-t` threads for IQ4 packing from ggml code",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-05-11T01:06:05Z",
        "updated_at": "2026-05-11T01:06:21Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22928",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing"
        ]
      },
      {
        "number": 22934,
        "kind": "pr",
        "state": "open",
        "title": "tests: support multi-op perf groups in test-backend-ops",
        "labels": [
          "testing"
        ],
        "created_at": "2026-05-11T07:09:43Z",
        "updated_at": "2026-05-11T08:46:47Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22934",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing"
        ]
      },
      {
        "number": 22939,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cpu: avoid treating all host RAM as free",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-05-11T09:45:39Z",
        "updated_at": "2026-07-31T19:20:08Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22939",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 22950,
        "kind": "pr",
        "state": "open",
        "title": "[Tensor Parallel] Enable Auto parameter fitting in split-mode tensor",
        "labels": [
          "documentation",
          "ggml"
        ],
        "created_at": "2026-05-11T16:59:24Z",
        "updated_at": "2026-06-05T02:39:27Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22950",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "api_ui_docs"
        ]
      },
      {
        "number": 22951,
        "kind": "pr",
        "state": "open",
        "title": "vulkan: Pad Q3_K/Q6_K tensors out to 32-bit alignment",
        "labels": [
          "Vulkan",
          "ggml"
        ],
        "created_at": "2026-05-11T17:57:10Z",
        "updated_at": "2026-05-14T14:27:18Z",
        "closed_at": null,
        "comments": 13,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22951",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing"
        ]
      },
      {
        "number": 22960,
        "kind": "pr",
        "state": "open",
        "title": "server : emit empty input field in anthropic streaming tool_use content_block_start",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2026-05-12T06:57:33Z",
        "updated_at": "2026-05-12T07:30:37Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22960",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 22970,
        "kind": "pr",
        "state": "open",
        "title": "vulkan : transpose A-matrix data layout for K-quant mul_mat performance",
        "labels": [
          "documentation",
          "Vulkan",
          "python",
          "ggml"
        ],
        "created_at": "2026-05-12T10:30:32Z",
        "updated_at": "2026-08-25T07:49:48Z",
        "closed_at": null,
        "comments": 39,
        "url": "https://github.com/ggml-org/llama.cpp/pull/22970",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 23009,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cpu : fix riscv xtheadvector builds and add a q1_0 vec dot kernel",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-05-13T12:17:10Z",
        "updated_at": "2026-05-14T16:08:45Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23009",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 23019,
        "kind": "pr",
        "state": "open",
        "title": "feat: add SQ4 (Semantic Quantization 4-bit) quantization type",
        "labels": [
          "examples",
          "ggml"
        ],
        "created_at": "2026-05-13T18:14:24Z",
        "updated_at": "2026-05-15T12:16:49Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23019",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 23029,
        "kind": "pr",
        "state": "open",
        "title": "chat : add Nemotron Nano v2 specialized parser",
        "labels": [
          "testing"
        ],
        "created_at": "2026-05-14T00:04:02Z",
        "updated_at": "2026-06-03T19:46:48Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23029",
        "categories": [
          "models_architectures",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 23033,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: missing tensor 'blk.40.ssm_conv1d.weight'",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-05-14T02:37:51Z",
        "updated_at": "2026-07-31T08:54:00Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/issues/23033",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 23037,
        "kind": "pr",
        "state": "open",
        "title": "model: discover hybrid layer types via tensor presence",
        "labels": [
          "model"
        ],
        "created_at": "2026-05-14T04:02:03Z",
        "updated_at": "2026-05-14T04:32:31Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23037",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 23112,
        "kind": "pr",
        "state": "open",
        "title": "[DRAFT] Support for Zaya1 8B model",
        "labels": [
          "model",
          "testing",
          "Nvidia GPU",
          "examples",
          "python",
          "ggml",
          "CUDA",
          "conversion"
        ],
        "created_at": "2026-05-15T16:59:17Z",
        "updated_at": "2026-07-31T08:23:33Z",
        "closed_at": null,
        "comments": 32,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23112",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 23122,
        "kind": "pr",
        "state": "open",
        "title": "ggml: add DeepSeek V4 hyperconnection + KV ops (CPU)",
        "labels": [
          "model",
          "testing",
          "examples",
          "python",
          "ggml"
        ],
        "created_at": "2026-05-15T21:55:13Z",
        "updated_at": "2026-08-01T19:45:41Z",
        "closed_at": null,
        "comments": 22,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23122",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 23128,
        "kind": "pr",
        "state": "open",
        "title": "server: redact URL path/query in handle_media INFO log",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2026-05-16T01:06:00Z",
        "updated_at": "2026-05-16T01:45:26Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23128",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 23161,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Model type gemma4_assistant not supported",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-05-16T19:31:43Z",
        "updated_at": "2026-08-17T01:07:41Z",
        "closed_at": null,
        "comments": 13,
        "url": "https://github.com/ggml-org/llama.cpp/issues/23161",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 23188,
        "kind": "pr",
        "state": "open",
        "title": "ggml-rpc: opt-in OpenTelemetry tracing + protocol hardening",
        "labels": [
          "documentation",
          "examples",
          "ggml"
        ],
        "created_at": "2026-05-17T05:49:01Z",
        "updated_at": "2026-05-17T17:04:54Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23188",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 23201,
        "kind": "pr",
        "state": "open",
        "title": "contrib: require explicit agreement for including external code",
        "labels": [],
        "created_at": "2026-05-17T11:29:20Z",
        "updated_at": "2026-05-18T11:28:36Z",
        "closed_at": null,
        "comments": 14,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23201",
        "categories": [
          "models_architectures",
          "build_platform_ops"
        ]
      },
      {
        "number": 23208,
        "kind": "pr",
        "state": "open",
        "title": "llama-bench: add VRAM and RAM reporting",
        "labels": [
          "examples"
        ],
        "created_at": "2026-05-17T13:54:42Z",
        "updated_at": "2026-05-20T20:37:59Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23208",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 23210,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: llama-server crashes on CUDA with Qwen3.6-27B",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-05-17T14:13:17Z",
        "updated_at": "2026-08-26T14:58:21Z",
        "closed_at": null,
        "comments": 13,
        "url": "https://github.com/ggml-org/llama.cpp/issues/23210",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 23228,
        "kind": "pr",
        "state": "open",
        "title": "vulkan: fix GDN shader on MoltenVK by replacing gl_SubgroupInvocationID with gl_LocalInvocationIndex",
        "labels": [
          "Vulkan",
          "ggml"
        ],
        "created_at": "2026-05-17T19:10:56Z",
        "updated_at": "2026-05-17T23:07:12Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23228",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 23232,
        "kind": "pr",
        "state": "open",
        "title": "fit : add --fit-show-mem to print probe table at INFO",
        "labels": [
          "examples"
        ],
        "created_at": "2026-05-17T21:46:35Z",
        "updated_at": "2026-05-18T14:26:14Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23232",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 23235,
        "kind": "pr",
        "state": "open",
        "title": "server : implement namespaced tools in responses API",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2026-05-17T22:08:03Z",
        "updated_at": "2026-05-17T23:40:26Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23235",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 23247,
        "kind": "pr",
        "state": "open",
        "title": "common : support schema-constrained decoding for Gemma 4 tool calls",
        "labels": [
          "testing"
        ],
        "created_at": "2026-05-18T03:28:38Z",
        "updated_at": "2026-06-30T16:57:45Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23247",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 23309,
        "kind": "pr",
        "state": "open",
        "title": "Improve performance of ggml_gemv_q4_K_8x8_q8_K for +12-23% tok/s on AVX-VNNI systems",
        "labels": [
          "testing",
          "ggml"
        ],
        "created_at": "2026-05-19T02:20:04Z",
        "updated_at": "2026-06-15T16:21:07Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23309",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 23340,
        "kind": "pr",
        "state": "open",
        "title": "server : add --slot-context for per-slot context size configuration",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2026-05-19T14:10:48Z",
        "updated_at": "2026-05-19T22:23:20Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23340",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 23343,
        "kind": "pr",
        "state": "open",
        "title": "nemotron : support non-standard GQA, BPE tokenizer and domyn-small pre-type",
        "labels": [
          "model",
          "python"
        ],
        "created_at": "2026-05-19T14:52:00Z",
        "updated_at": "2026-05-27T13:18:09Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23343",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 23358,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cpu: Optimized Arm NEON cpu q1_0 dot (with plain/DP/I8MM)",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-05-19T18:21:41Z",
        "updated_at": "2026-06-05T16:16:39Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23358",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 23373,
        "kind": "pr",
        "state": "open",
        "title": "readme : add FootprintAI/go-nativeml to Bindings",
        "labels": [],
        "created_at": "2026-05-20T01:46:02Z",
        "updated_at": "2026-05-20T01:46:03Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23373",
        "categories": [
          "native_backends",
          "models_architectures",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 23374,
        "kind": "pr",
        "state": "open",
        "title": "Tests: improve split-max-size coverage",
        "labels": [
          "examples"
        ],
        "created_at": "2026-05-20T02:02:06Z",
        "updated_at": "2026-05-20T09:31:26Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23374",
        "categories": [
          "quantization",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 23388,
        "kind": "pr",
        "state": "open",
        "title": "common : default embedding preset",
        "labels": [],
        "created_at": "2026-05-20T07:12:49Z",
        "updated_at": "2026-05-20T07:31:55Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23388",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "api_ui_docs"
        ]
      },
      {
        "number": 23401,
        "kind": "pr",
        "state": "open",
        "title": "ggml-metal: fix deployment target compatibility for waitUntilSignaledValue",
        "labels": [
          "ggml",
          "Apple Metal"
        ],
        "created_at": "2026-05-20T09:14:06Z",
        "updated_at": "2026-05-20T12:42:09Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23401",
        "categories": [
          "native_backends",
          "build_platform_ops"
        ]
      },
      {
        "number": 23402,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cpu:Optimized risc-v cpu nvfp4",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-05-20T09:20:05Z",
        "updated_at": "2026-05-20T13:30:46Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23402",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 23407,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Multi-modal Qwen3.5 on llama.cpp triggered a CUDA error",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-05-20T09:55:17Z",
        "updated_at": "2026-08-05T01:10:56Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/23407",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 23436,
        "kind": "pr",
        "state": "open",
        "title": "json-schema-to-grammar: expand PCRE shorthands in pattern strings",
        "labels": [
          "testing"
        ],
        "created_at": "2026-05-20T18:31:11Z",
        "updated_at": "2026-08-25T12:44:45Z",
        "closed_at": null,
        "comments": 12,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23436",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 23439,
        "kind": "pr",
        "state": "open",
        "title": "ggml/cpu: skip zero-scale blocks in TQ1_0 and TQ2_0 vec_dot kernels",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-05-20T20:15:56Z",
        "updated_at": "2026-05-21T10:46:45Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23439",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 23440,
        "kind": "pr",
        "state": "open",
        "title": "MoE disk offloading for Metal",
        "labels": [
          "ggml",
          "Apple Metal"
        ],
        "created_at": "2026-05-20T20:30:33Z",
        "updated_at": "2026-07-28T08:53:37Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23440",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 23460,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Unable to pass samplers to models-preset in server router mode",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-05-21T05:12:44Z",
        "updated_at": "2026-07-28T17:59:33Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/23460",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 23474,
        "kind": "pr",
        "state": "open",
        "title": "Optimize ggml_vec_dot_q4_K_q8_K_generic",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-05-21T11:11:53Z",
        "updated_at": "2026-08-13T06:50:41Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23474",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 23476,
        "kind": "pr",
        "state": "open",
        "title": "imatrix: optionally activate MTP/NextN draft head during collection",
        "labels": [
          "examples"
        ],
        "created_at": "2026-05-21T12:58:54Z",
        "updated_at": "2026-08-15T21:12:08Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23476",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "api_ui_docs"
        ]
      },
      {
        "number": 23489,
        "kind": "pr",
        "state": "open",
        "title": "common/download: prevent duplicate MTP draft model downloads",
        "labels": [],
        "created_at": "2026-05-21T18:22:06Z",
        "updated_at": "2026-05-21T19:27:40Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23489",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 23492,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cpu: ARM Repack kernels for Q1_0",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-05-21T19:21:27Z",
        "updated_at": "2026-06-04T08:29:02Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23492",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 23510,
        "kind": "pr",
        "state": "open",
        "title": "Skip hash probe when no cache is present",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-05-22T07:19:07Z",
        "updated_at": "2026-05-22T15:41:35Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23510",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 23515,
        "kind": "pr",
        "state": "open",
        "title": "fix no timeout flag handling in router mode",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2026-05-22T09:14:17Z",
        "updated_at": "2026-05-23T23:55:49Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23515",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 23572,
        "kind": "pr",
        "state": "open",
        "title": "CUDA: native 4-bit float quant (Blackwell PP +40%)",
        "labels": [
          "examples"
        ],
        "created_at": "2026-05-23T12:59:14Z",
        "updated_at": "2026-08-18T09:31:37Z",
        "closed_at": null,
        "comments": 11,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23572",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 23575,
        "kind": "pr",
        "state": "open",
        "title": "llama-quantize: use static quantization level for tensors missing from imatrix data",
        "labels": [
          "examples"
        ],
        "created_at": "2026-05-23T13:33:04Z",
        "updated_at": "2026-07-29T13:56:42Z",
        "closed_at": null,
        "comments": 23,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23575",
        "categories": [
          "quantization",
          "models_architectures",
          "speculative_generation",
          "api_ui_docs"
        ]
      },
      {
        "number": 23577,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: MTP with Qwen3.6 27B outputs repeated //// after long session",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-05-23T14:39:09Z",
        "updated_at": "2026-08-09T20:19:55Z",
        "closed_at": null,
        "comments": 31,
        "url": "https://github.com/ggml-org/llama.cpp/issues/23577",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 23580,
        "kind": "pr",
        "state": "open",
        "title": "cmake : error when LLAMA_BUILD_APP=ON and LLAMA_BUILD_TOOLS=OFF",
        "labels": [
          "build"
        ],
        "created_at": "2026-05-23T17:00:48Z",
        "updated_at": "2026-05-23T17:06:15Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23580",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 23593,
        "kind": "pr",
        "state": "open",
        "title": "ggml: fix AVX-512 BF16 build with clang-cl",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-05-24T05:05:38Z",
        "updated_at": "2026-05-27T16:58:04Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23593",
        "categories": [
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 23602,
        "kind": "pr",
        "state": "open",
        "title": "common: preserve horizontal whitespace in tool calls",
        "labels": [
          "testing"
        ],
        "created_at": "2026-05-24T09:02:05Z",
        "updated_at": "2026-05-28T03:49:45Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23602",
        "categories": [
          "models_architectures",
          "structured_generation_tools",
          "correctness_testing"
        ]
      },
      {
        "number": 23609,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Expose Custom Parameters in Web GUI",
        "labels": [],
        "created_at": "2026-05-24T12:01:30Z",
        "updated_at": "2026-08-05T01:10:54Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/23609",
        "categories": [
          "scheduling_serving",
          "api_ui_docs"
        ]
      },
      {
        "number": 23618,
        "kind": "pr",
        "state": "open",
        "title": "gguf-py: preserve MoE size labels for mmproj metadata",
        "labels": [
          "python"
        ],
        "created_at": "2026-05-24T15:30:17Z",
        "updated_at": "2026-06-05T19:46:23Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23618",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "loader_conversion",
          "correctness_testing",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 23622,
        "kind": "issue",
        "state": "open",
        "title": "UI Bug: Editing assistant message duplicates reasoning block and corrupts stored content",
        "labels": [],
        "created_at": "2026-05-24T16:16:19Z",
        "updated_at": "2026-08-05T01:10:53Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/23622",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "multimodal_embedding",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 23644,
        "kind": "pr",
        "state": "open",
        "title": "fix: sanitize sampling and mirostat parameters to prevent unstable states",
        "labels": [],
        "created_at": "2026-05-25T04:34:28Z",
        "updated_at": "2026-05-26T14:59:09Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23644",
        "categories": [
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing"
        ]
      },
      {
        "number": 23666,
        "kind": "pr",
        "state": "open",
        "title": "server: enhance FIFO prompt cache eviction with second-chance algorithm",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2026-05-25T13:18:07Z",
        "updated_at": "2026-06-18T08:42:29Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23666",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 23671,
        "kind": "pr",
        "state": "open",
        "title": "ggml : add `alloc_buffer_n` to buffer type interface",
        "labels": [
          "Nvidia GPU",
          "Vulkan",
          "ggml",
          "SYCL",
          "Apple Metal",
          "Ascend NPU",
          "OpenCL",
          "IBM zDNN",
          "Hexagon",
          "CUDA",
          "OpenVINO",
          "WebGPU"
        ],
        "created_at": "2026-05-25T14:28:56Z",
        "updated_at": "2026-06-15T13:16:36Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23671",
        "categories": [
          "native_backends",
          "models_architectures",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 23683,
        "kind": "pr",
        "state": "open",
        "title": "ggml.h: Fix warnings due to enum arithmethic",
        "labels": [
          "testing",
          "examples",
          "ggml",
          "mtmd"
        ],
        "created_at": "2026-05-25T19:35:15Z",
        "updated_at": "2026-08-11T15:25:10Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23683",
        "categories": [
          "models_architectures",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 23685,
        "kind": "pr",
        "state": "open",
        "title": "HIP: Adds 4x packed Q8_1 activation for Q4_K_M models in MMVQ Extension to Q5_K and Q6_K Models ",
        "labels": [
          "Nvidia GPU",
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-05-25T20:57:04Z",
        "updated_at": "2026-08-21T20:13:45Z",
        "closed_at": null,
        "comments": 18,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23685",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 23692,
        "kind": "pr",
        "state": "open",
        "title": "quants: MSE scale search to minimize NVFP4 reconstruction error",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-05-26T00:32:57Z",
        "updated_at": "2026-08-18T09:58:25Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23692",
        "categories": [
          "quantization",
          "models_architectures",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 23699,
        "kind": "pr",
        "state": "open",
        "title": "CI: test-backend-ops -j8",
        "labels": [
          "script",
          "testing",
          "python",
          "devops"
        ],
        "created_at": "2026-05-26T05:11:50Z",
        "updated_at": "2026-06-26T15:01:37Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23699",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 23704,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: server in Router Mode: Multiple presets for the same model (loaded once)",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-05-26T07:27:46Z",
        "updated_at": "2026-08-24T14:18:46Z",
        "closed_at": null,
        "comments": 16,
        "url": "https://github.com/ggml-org/llama.cpp/issues/23704",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 23707,
        "kind": "pr",
        "state": "open",
        "title": "model-loader: Split mmap buffers to avoid mapping unneeded data to the GPU",
        "labels": [],
        "created_at": "2026-05-26T07:56:34Z",
        "updated_at": "2026-05-26T07:56:35Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23707",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 23724,
        "kind": "pr",
        "state": "open",
        "title": "[Metal] Partially support OUT_PROD op",
        "labels": [
          "documentation",
          "testing",
          "ggml",
          "Apple Metal"
        ],
        "created_at": "2026-05-26T14:22:08Z",
        "updated_at": "2026-07-28T14:08:28Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23724",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 23737,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: GGML_ASSERT(tensor->data != NULL) on Vulkan since b9318",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-05-26T17:42:06Z",
        "updated_at": "2026-08-26T13:40:24Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/issues/23737",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 23745,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cpu: add RVV repack GEMM and GEMV for Q3_K, Q6_K",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-05-26T23:07:12Z",
        "updated_at": "2026-07-01T07:36:36Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23745",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 23762,
        "kind": "pr",
        "state": "open",
        "title": "vulkan: fix UMA performance by preferring cached host memory and handling non…",
        "labels": [
          "Vulkan",
          "ggml"
        ],
        "created_at": "2026-05-27T06:26:08Z",
        "updated_at": "2026-06-18T05:05:15Z",
        "closed_at": null,
        "comments": 9,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23762",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 23768,
        "kind": "pr",
        "state": "open",
        "title": "fix: duplicated \"the\" in compare-llama-bench and minicpmv-surgery comments",
        "labels": [
          "script",
          "examples",
          "python"
        ],
        "created_at": "2026-05-27T07:33:16Z",
        "updated_at": "2026-05-27T07:33:42Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23768",
        "categories": [
          "models_architectures",
          "api_ui_docs"
        ]
      },
      {
        "number": 23769,
        "kind": "issue",
        "state": "open",
        "title": "Crash with Vulcan on Intel B70 with MoE models (Qwen3.6-35B-A3B-MTP and Qwen3.6-35B-A3B)",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-05-27T08:16:50Z",
        "updated_at": "2026-08-11T15:01:34Z",
        "closed_at": null,
        "comments": 11,
        "url": "https://github.com/ggml-org/llama.cpp/issues/23769",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 23774,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Huge performance degradation with MTP on Vulkan",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-05-27T09:23:16Z",
        "updated_at": "2026-08-21T02:13:29Z",
        "closed_at": null,
        "comments": 20,
        "url": "https://github.com/ggml-org/llama.cpp/issues/23774",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 23787,
        "kind": "pr",
        "state": "open",
        "title": "common: fix HTTPS handshake on Windows, harden HTTP client",
        "labels": [],
        "created_at": "2026-05-27T16:49:44Z",
        "updated_at": "2026-06-19T20:20:16Z",
        "closed_at": null,
        "comments": 10,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23787",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "loader_conversion",
          "build_platform_ops"
        ]
      },
      {
        "number": 23793,
        "kind": "pr",
        "state": "open",
        "title": "Speed up ggml_gemv_q4_K_8x8_q8_K",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-05-28T01:35:18Z",
        "updated_at": "2026-06-09T13:08:50Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23793",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 23799,
        "kind": "pr",
        "state": "open",
        "title": "Sync zDNN branch lineage",
        "labels": [
          "script",
          "python",
          "ggml",
          "IBM zDNN"
        ],
        "created_at": "2026-05-28T05:22:47Z",
        "updated_at": "2026-06-08T11:36:21Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23799",
        "categories": [
          "native_backends",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 23821,
        "kind": "pr",
        "state": "open",
        "title": "Bug fix: Hexagon support for llama-cli and llama-server",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2026-05-28T14:13:34Z",
        "updated_at": "2026-06-02T04:28:54Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23821",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 23828,
        "kind": "pr",
        "state": "open",
        "title": "fix(tps): correct off-by-one in decode token count for generation TPS",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2026-05-28T16:19:28Z",
        "updated_at": "2026-05-28T16:42:02Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23828",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 23836,
        "kind": "pr",
        "state": "open",
        "title": "common: use common_cpu_get_num_math() for --threads -1",
        "labels": [],
        "created_at": "2026-05-28T19:48:34Z",
        "updated_at": "2026-08-24T15:12:49Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23836",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 23848,
        "kind": "pr",
        "state": "open",
        "title": "agentic: question tool + shared plumbing",
        "labels": [
          "documentation",
          "examples",
          "python",
          "server",
          "server/ui"
        ],
        "created_at": "2026-05-29T02:42:00Z",
        "updated_at": "2026-07-22T12:59:29Z",
        "closed_at": null,
        "comments": 9,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23848",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 23854,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: The title should ideally be less than 64 characters in length",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-05-29T05:14:46Z",
        "updated_at": "2026-08-05T01:10:52Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/issues/23854",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 23900,
        "kind": "pr",
        "state": "open",
        "title": "fix: VMM pool cuMemSetAccess for ROCm gfx1151 APU",
        "labels": [
          "Nvidia GPU",
          "ggml"
        ],
        "created_at": "2026-05-30T08:32:07Z",
        "updated_at": "2026-05-30T08:36:54Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23900",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing"
        ]
      },
      {
        "number": 23905,
        "kind": "pr",
        "state": "open",
        "title": "lama-bench : fix default thread count evaluated at static init time",
        "labels": [
          "examples"
        ],
        "created_at": "2026-05-30T09:39:27Z",
        "updated_at": "2026-05-30T14:28:01Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23905",
        "categories": [
          "models_architectures",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 23908,
        "kind": "pr",
        "state": "open",
        "title": "build: Add vulkan building script",
        "labels": [],
        "created_at": "2026-05-30T11:08:34Z",
        "updated_at": "2026-05-30T14:30:30Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23908",
        "categories": [
          "native_backends",
          "models_architectures",
          "build_platform_ops"
        ]
      },
      {
        "number": 23909,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: ADD Q2_0",
        "labels": [
          "enhancement",
          "stale"
        ],
        "created_at": "2026-05-30T11:42:08Z",
        "updated_at": "2026-08-20T02:06:45Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/23909",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 23915,
        "kind": "pr",
        "state": "open",
        "title": "loader: increase async upload staging buffer to 4 MiB",
        "labels": [],
        "created_at": "2026-05-30T16:29:35Z",
        "updated_at": "2026-05-30T17:15:55Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23915",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 23932,
        "kind": "pr",
        "state": "open",
        "title": "common: use physical core count for --threads -1 default",
        "labels": [],
        "created_at": "2026-05-31T05:48:32Z",
        "updated_at": "2026-05-31T05:48:32Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23932",
        "categories": [
          "other"
        ]
      },
      {
        "number": 23933,
        "kind": "pr",
        "state": "open",
        "title": "common : retry HTTP requests over IPv4 when IPv6 connect fails",
        "labels": [],
        "created_at": "2026-05-31T06:03:15Z",
        "updated_at": "2026-06-18T22:20:24Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23933",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 23948,
        "kind": "issue",
        "state": "open",
        "title": "Feature request: (server) file upload API",
        "labels": [],
        "created_at": "2026-05-31T21:30:35Z",
        "updated_at": "2026-08-01T01:14:47Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/23948",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 23952,
        "kind": "pr",
        "state": "open",
        "title": "server: merge custom preset aliases into existing HF cached model entries",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2026-06-01T02:25:08Z",
        "updated_at": "2026-06-01T02:25:26Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23952",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 23984,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Unable to use the OpenVINO NPU backend",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-06-01T17:27:32Z",
        "updated_at": "2026-08-21T02:13:27Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/23984",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 23992,
        "kind": "pr",
        "state": "open",
        "title": "metal: optimize pad_reflect_1d_f32 kernel",
        "labels": [
          "testing",
          "ggml",
          "Apple Metal"
        ],
        "created_at": "2026-06-01T19:58:11Z",
        "updated_at": "2026-07-01T01:42:45Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23992",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing"
        ]
      },
      {
        "number": 23993,
        "kind": "pr",
        "state": "open",
        "title": "server/common: Fix `response_format: json_schema` & prefill parsing bug",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2026-06-01T20:04:08Z",
        "updated_at": "2026-08-02T00:39:34Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/23993",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 24010,
        "kind": "pr",
        "state": "open",
        "title": "server: add KV cache metrics",
        "labels": [
          "examples",
          "python",
          "server"
        ],
        "created_at": "2026-06-02T06:41:23Z",
        "updated_at": "2026-08-11T18:30:03Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24010",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24014,
        "kind": "pr",
        "state": "open",
        "title": "mtmd: correct gemma4 min/max tokens",
        "labels": [
          "examples"
        ],
        "created_at": "2026-06-02T09:08:24Z",
        "updated_at": "2026-06-02T09:37:41Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24014",
        "categories": [
          "native_backends",
          "models_architectures",
          "multimodal_embedding",
          "api_ui_docs"
        ]
      },
      {
        "number": 24015,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: ggml-backend-meta.cpp:541: GGML_ASSERT(src_ss[0].axis != GGML_BACKEND_SPLIT_AXIS_0) failed",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-06-02T09:53:07Z",
        "updated_at": "2026-08-23T01:07:43Z",
        "closed_at": null,
        "comments": 10,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24015",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 24023,
        "kind": "pr",
        "state": "open",
        "title": "tools/ui: add OAuth support for MCP servers",
        "labels": [
          "examples",
          "server/ui"
        ],
        "created_at": "2026-06-02T13:06:20Z",
        "updated_at": "2026-06-02T13:12:29Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24023",
        "categories": [
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24028,
        "kind": "pr",
        "state": "open",
        "title": "server: persist slot checkpoints to .ckpt sidecar (LSCKPT2)",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2026-06-02T15:08:23Z",
        "updated_at": "2026-07-15T18:36:39Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24028",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 24033,
        "kind": "pr",
        "state": "open",
        "title": "jinja: implement map('filter')",
        "labels": [
          "testing",
          "jinja parser"
        ],
        "created_at": "2026-06-02T16:08:47Z",
        "updated_at": "2026-08-05T18:47:58Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24033",
        "categories": [
          "native_backends",
          "models_architectures",
          "structured_generation_tools",
          "correctness_testing"
        ]
      },
      {
        "number": 24035,
        "kind": "pr",
        "state": "open",
        "title": "server: preserve reusable checkpoints for recurrent / hybrid prompts",
        "labels": [
          "testing",
          "examples",
          "server"
        ],
        "created_at": "2026-06-02T16:52:23Z",
        "updated_at": "2026-07-27T10:19:38Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24035",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24054,
        "kind": "pr",
        "state": "open",
        "title": "log: add condition of debug log",
        "labels": [],
        "created_at": "2026-06-03T02:10:56Z",
        "updated_at": "2026-06-03T02:10:57Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24054",
        "categories": [
          "correctness_testing"
        ]
      },
      {
        "number": 24055,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Context checkpoints always invalidated on hybrid/recurrent models ",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-06-03T02:13:37Z",
        "updated_at": "2026-08-18T12:11:35Z",
        "closed_at": null,
        "comments": 19,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24055",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 24058,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cpu: add wasm simd path for iq4_nl_q8_0",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-06-03T06:57:03Z",
        "updated_at": "2026-06-03T07:14:49Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24058",
        "categories": [
          "quantization",
          "models_architectures",
          "loader_conversion",
          "correctness_testing"
        ]
      },
      {
        "number": 24066,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Vulkan: performance drop in recent builds",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-06-03T10:06:36Z",
        "updated_at": "2026-08-25T07:49:48Z",
        "closed_at": null,
        "comments": 41,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24066",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 24076,
        "kind": "pr",
        "state": "open",
        "title": "server: reject partial media truncation",
        "labels": [
          "testing",
          "examples",
          "python",
          "server"
        ],
        "created_at": "2026-06-03T14:16:57Z",
        "updated_at": "2026-06-25T10:59:20Z",
        "closed_at": null,
        "comments": 10,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24076",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24083,
        "kind": "pr",
        "state": "open",
        "title": "Add nvidia/llama-nemotron-rerank-1b-v2 cross-encoder reraranker",
        "labels": [
          "model",
          "examples",
          "python",
          "server",
          "conversion"
        ],
        "created_at": "2026-06-03T17:18:00Z",
        "updated_at": "2026-08-21T14:07:54Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24083",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24090,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: WebUI - Projects / Folders for Conversations",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-06-03T20:47:54Z",
        "updated_at": "2026-08-04T01:15:39Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24090",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24094,
        "kind": "pr",
        "state": "open",
        "title": "fix: don't build AMX by default with Apple clang",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-06-03T23:12:20Z",
        "updated_at": "2026-06-03T23:58:10Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24094",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24113,
        "kind": "pr",
        "state": "open",
        "title": "test(server): fix flaky completion test",
        "labels": [
          "examples",
          "python",
          "server"
        ],
        "created_at": "2026-06-04T11:12:06Z",
        "updated_at": "2026-06-04T11:12:34Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24113",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 24114,
        "kind": "pr",
        "state": "open",
        "title": "test(server): homogeneize slow tests eviction",
        "labels": [
          "examples",
          "python",
          "devops",
          "server"
        ],
        "created_at": "2026-06-04T11:15:10Z",
        "updated_at": "2026-06-04T11:25:36Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24114",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24121,
        "kind": "pr",
        "state": "open",
        "title": "server: allow missing/null content key in OpenAI Responses API",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2026-06-04T12:22:51Z",
        "updated_at": "2026-06-04T12:23:14Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24121",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "loader_conversion",
          "api_ui_docs"
        ]
      },
      {
        "number": 24122,
        "kind": "pr",
        "state": "open",
        "title": "rpc: reduce small-message overhead and tune cache probing",
        "labels": [
          "examples",
          "python",
          "devops",
          "ggml"
        ],
        "created_at": "2026-06-04T12:34:37Z",
        "updated_at": "2026-08-24T09:10:59Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24122",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24124,
        "kind": "pr",
        "state": "open",
        "title": "Add ctx-per-slot argument for unified KV cache",
        "labels": [
          "documentation",
          "examples",
          "server"
        ],
        "created_at": "2026-06-04T12:49:39Z",
        "updated_at": "2026-07-24T18:00:40Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24124",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 24132,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: gemma-4-E4B-it-Q4_0.gguf  ggml-backend.cpp:1367: GGML_ASSERT(n_inputs < GGML_SCHED_MAX_SPLIT_INPUTS) failed",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-06-04T15:49:20Z",
        "updated_at": "2026-08-11T10:09:11Z",
        "closed_at": null,
        "comments": 13,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24132",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 24139,
        "kind": "issue",
        "state": "open",
        "title": "Research: KVarN (variance-normalized KV-cache quantization)",
        "labels": [
          "research 🔬"
        ],
        "created_at": "2026-06-04T18:44:08Z",
        "updated_at": "2026-06-13T21:43:13Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24139",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 24143,
        "kind": "pr",
        "state": "open",
        "title": "Fix/server prompt cache no consume on load",
        "labels": [
          "examples",
          "python",
          "server"
        ],
        "created_at": "2026-06-04T19:55:03Z",
        "updated_at": "2026-06-04T20:18:49Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24143",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24156,
        "kind": "pr",
        "state": "open",
        "title": "model-loader: add --reclaim-mmap-source to drop dormant mmap pages (Fixes #16761)",
        "labels": [],
        "created_at": "2026-06-05T05:27:43Z",
        "updated_at": "2026-07-10T20:30:47Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24156",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24163,
        "kind": "pr",
        "state": "open",
        "title": "mtmd : add Apple CoreML backend for vision encoding",
        "labels": [
          "examples",
          "python"
        ],
        "created_at": "2026-06-05T08:06:00Z",
        "updated_at": "2026-07-22T11:26:36Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24163",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 24164,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: WEBUI does not uploads audio ogg files",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-06-05T08:14:52Z",
        "updated_at": "2026-08-05T01:10:49Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24164",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 24168,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: [SYCL]  empty/gibberish output on hybrid models + ggml_sycl_op_mul_mat crash (qwen3next/qwen35 arch) on Intel Arc Pro B60 — regression between *pinpointed b9128- b9159* (2026-03-23) and server-intel latest (2026-06-03)",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-06-05T08:21:45Z",
        "updated_at": "2026-08-04T09:39:31Z",
        "closed_at": null,
        "comments": 25,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24168",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24173,
        "kind": "pr",
        "state": "open",
        "title": "refactor: replace embd_normalize integer parameter with enum",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2026-06-05T10:36:58Z",
        "updated_at": "2026-06-05T10:52:38Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24173",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24177,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Using RPC: top_k backend sampling crashes with GGML_ASSERT(shared_mem <= smpb) in argsort.cu (affects all AMD GPUs)",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-06-05T11:44:57Z",
        "updated_at": "2026-07-29T14:33:11Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24177",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "correctness_testing"
        ]
      },
      {
        "number": 24185,
        "kind": "pr",
        "state": "open",
        "title": "Add ROCmFP4 CPU quantization support",
        "labels": [
          "examples",
          "ggml"
        ],
        "created_at": "2026-06-05T13:42:53Z",
        "updated_at": "2026-08-03T15:18:47Z",
        "closed_at": null,
        "comments": 13,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24185",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24197,
        "kind": "pr",
        "state": "open",
        "title": "docs: link function calling guide from README",
        "labels": [],
        "created_at": "2026-06-05T16:45:52Z",
        "updated_at": "2026-06-05T16:48:42Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24197",
        "categories": [
          "structured_generation_tools",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 24199,
        "kind": "issue",
        "state": "open",
        "title": "Intel Arc A770 + llama.cpp Vulkan: 9B Model at 256K Context, 49 t/s — Practical Guide",
        "labels": [
          "stale"
        ],
        "created_at": "2026-06-05T17:14:44Z",
        "updated_at": "2026-08-12T01:08:51Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24199",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24210,
        "kind": "pr",
        "state": "open",
        "title": "server: context shift",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2026-06-05T20:44:03Z",
        "updated_at": "2026-06-07T16:07:00Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24210",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 24247,
        "kind": "pr",
        "state": "open",
        "title": "server : add token healing support",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2026-06-06T20:19:33Z",
        "updated_at": "2026-06-06T20:20:38Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24247",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 24255,
        "kind": "pr",
        "state": "open",
        "title": "Cast manually for _mm_prefetch() to avoid type mismatch error on Clang",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-06-07T06:07:16Z",
        "updated_at": "2026-06-07T06:07:34Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24255",
        "categories": [
          "models_architectures",
          "build_platform_ops"
        ]
      },
      {
        "number": 24292,
        "kind": "pr",
        "state": "open",
        "title": "rpc : fix UAF in graph_recompute leading to remote code execution",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-06-08T07:17:48Z",
        "updated_at": "2026-06-10T04:58:05Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24292",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24295,
        "kind": "issue",
        "state": "open",
        "title": "Responses API: silently drops 'namespace' and 'web_search' tool types → Codex CLI MCP tools unreachable",
        "labels": [],
        "created_at": "2026-06-08T09:47:13Z",
        "updated_at": "2026-08-22T16:00:06Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24295",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "api_ui_docs"
        ]
      },
      {
        "number": 24301,
        "kind": "pr",
        "state": "open",
        "title": "Docs: Adds AI Badgr as an optional hosted GPU launch path for running a `llama.cpp` server.",
        "labels": [],
        "created_at": "2026-06-08T11:20:49Z",
        "updated_at": "2026-06-08T11:25:15Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24301",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24303,
        "kind": "issue",
        "state": "open",
        "title": "[BUG] Qwen3.6-35B-A3B / llama-server merges consecutive images into 2 frames, causing incorrect image count and partial image understanding",
        "labels": [],
        "created_at": "2026-06-08T12:45:54Z",
        "updated_at": "2026-08-21T09:42:22Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24303",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24309,
        "kind": "issue",
        "state": "open",
        "title": "[bug] --split-mode tensor crashes on GLM-4.5 Air MoE nextn draft head tensors (GGML_ASSERT tensor_axis_0 != nullptr)",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-06-08T15:46:23Z",
        "updated_at": "2026-08-23T01:07:42Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24309",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 24318,
        "kind": "pr",
        "state": "open",
        "title": "args: add --video-* CLI arguments",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2026-06-08T20:12:32Z",
        "updated_at": "2026-06-09T10:19:52Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24318",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "api_ui_docs"
        ]
      },
      {
        "number": 24324,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: llama.cpp-b9568/ggml/src/ggml-cuda/fattn.cu:579: fatal error",
        "labels": [
          "bug"
        ],
        "created_at": "2026-06-09T01:19:17Z",
        "updated_at": "2026-08-25T15:52:08Z",
        "closed_at": null,
        "comments": 9,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24324",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 24325,
        "kind": "pr",
        "state": "open",
        "title": "RPC: query remote backend op support instead of assuming all ops are supported",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-06-09T04:12:59Z",
        "updated_at": "2026-07-08T13:26:19Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24325",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 24338,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: In new chat `/` route draft message is being deleted when adding a system prompt",
        "labels": [
          "bug"
        ],
        "created_at": "2026-06-09T08:45:23Z",
        "updated_at": "2026-06-09T08:45:23Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24338",
        "categories": [
          "models_architectures",
          "correctness_testing"
        ]
      },
      {
        "number": 24343,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: E llama_init_from_model: failed to initialize the context: Gemma4Assistant",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-06-09T09:03:13Z",
        "updated_at": "2026-08-13T23:32:01Z",
        "closed_at": null,
        "comments": 12,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24343",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 24345,
        "kind": "issue",
        "state": "open",
        "title": "Feature: Improve Message Steering & Queuing features",
        "labels": [],
        "created_at": "2026-06-09T09:22:19Z",
        "updated_at": "2026-08-05T01:10:46Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24345",
        "categories": [
          "other"
        ]
      },
      {
        "number": 24348,
        "kind": "issue",
        "state": "open",
        "title": "Feature: Improve Tool Permissions management & logic",
        "labels": [],
        "created_at": "2026-06-09T09:26:08Z",
        "updated_at": "2026-08-05T01:10:46Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24348",
        "categories": [
          "other"
        ]
      },
      {
        "number": 24361,
        "kind": "pr",
        "state": "open",
        "title": "mtmd, llama: (demo) shared backend sched",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2026-06-09T13:51:05Z",
        "updated_at": "2026-06-18T13:38:52Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24361",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24364,
        "kind": "pr",
        "state": "open",
        "title": "Force NVFP4 W4A8 path for NVFP4_W4A16 layers on Blackwell, where NVFP4 normally uses the native W4A4 path. ",
        "labels": [
          "documentation",
          "testing",
          "Nvidia GPU",
          "python",
          "ggml",
          "CUDA",
          "conversion"
        ],
        "created_at": "2026-06-09T14:32:56Z",
        "updated_at": "2026-08-11T12:37:13Z",
        "closed_at": null,
        "comments": 13,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24364",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24368,
        "kind": "pr",
        "state": "open",
        "title": "metal : wind down leftover residency sets at teardown instead of aborting",
        "labels": [
          "ggml",
          "Apple Metal"
        ],
        "created_at": "2026-06-09T15:54:01Z",
        "updated_at": "2026-06-09T15:58:41Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24368",
        "categories": [
          "native_backends",
          "models_architectures",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 24382,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Special Token Injection",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-06-09T22:19:01Z",
        "updated_at": "2026-08-07T09:11:45Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24382",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24383,
        "kind": "pr",
        "state": "open",
        "title": "UI: Add support for calling API endpoints on remote llama-server",
        "labels": [
          "examples",
          "server/ui"
        ],
        "created_at": "2026-06-09T23:03:33Z",
        "updated_at": "2026-07-07T19:06:36Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24383",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24386,
        "kind": "pr",
        "state": "open",
        "title": "ggml: tune RDNA4 MMVQ warps for K-quants",
        "labels": [
          "Nvidia GPU",
          "ggml"
        ],
        "created_at": "2026-06-10T01:04:59Z",
        "updated_at": "2026-06-10T01:08:45Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24386",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 24406,
        "kind": "pr",
        "state": "open",
        "title": "vulkan: add Intel Xe flash attention optimization kernels (2/3, Xe-LPG Plus/Xe2/Xe3)",
        "labels": [
          "Vulkan",
          "ggml"
        ],
        "created_at": "2026-06-10T09:32:14Z",
        "updated_at": "2026-08-26T12:54:08Z",
        "closed_at": null,
        "comments": 21,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24406",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 24407,
        "kind": "pr",
        "state": "open",
        "title": "vulkan: GEMM/Group GEMM optimizations on Intel Xe (3/3, Xe-LPG Plus/Xe2/Xe3)",
        "labels": [
          "model",
          "Vulkan",
          "examples",
          "ggml"
        ],
        "created_at": "2026-06-10T09:34:42Z",
        "updated_at": "2026-07-09T12:57:45Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24407",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24408,
        "kind": "pr",
        "state": "open",
        "title": "vulkan: Intel Xe flash attention, GEMM optimizations(Xe-LPG Plus/Xe2/Xe3) [MEGA PR]",
        "labels": [
          "model",
          "Vulkan",
          "examples",
          "ggml"
        ],
        "created_at": "2026-06-10T09:36:51Z",
        "updated_at": "2026-07-02T11:55:12Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24408",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24409,
        "kind": "pr",
        "state": "open",
        "title": "server : add request path to 404 error message",
        "labels": [
          "examples",
          "python",
          "server"
        ],
        "created_at": "2026-06-10T10:27:54Z",
        "updated_at": "2026-06-10T10:28:25Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24409",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 24415,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: can't load gemma-4-12B with OpenVINO (CPU, GPU and NPU)",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-06-10T13:54:25Z",
        "updated_at": "2026-08-12T01:17:37Z",
        "closed_at": null,
        "comments": 11,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24415",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing"
        ]
      },
      {
        "number": 24423,
        "kind": "pr",
        "state": "open",
        "title": "DiffusionGemma",
        "labels": [
          "model",
          "testing",
          "Nvidia GPU",
          "examples",
          "python",
          "ggml",
          "CUDA",
          "conversion"
        ],
        "created_at": "2026-06-10T15:56:37Z",
        "updated_at": "2026-08-26T10:38:03Z",
        "closed_at": null,
        "comments": 84,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24423",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24427,
        "kind": "pr",
        "state": "open",
        "title": "Add diffusion-gemma block-diffusion support",
        "labels": [
          "model",
          "testing",
          "Nvidia GPU",
          "examples",
          "python",
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-06-10T16:36:07Z",
        "updated_at": "2026-08-24T04:00:27Z",
        "closed_at": null,
        "comments": 17,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24427",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24429,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: mtmd video input hangs on Windows — probe() deadlocks on faststart MP4, decode emits 0 frames when MOOV at end",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-06-10T16:55:59Z",
        "updated_at": "2026-08-24T17:22:08Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24429",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 24430,
        "kind": "pr",
        "state": "open",
        "title": "server : fix SSL env vars leaking to router child processes",
        "labels": [
          "examples",
          "python",
          "server"
        ],
        "created_at": "2026-06-10T18:14:02Z",
        "updated_at": "2026-06-10T18:17:26Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24430",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24435,
        "kind": "pr",
        "state": "open",
        "title": "Bert dense modules (ColBERT dense projection for bert / modern-bert)",
        "labels": [
          "model",
          "examples",
          "python"
        ],
        "created_at": "2026-06-10T20:24:37Z",
        "updated_at": "2026-06-11T21:03:30Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24435",
        "categories": [
          "quantization",
          "models_architectures",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 24437,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: HIP: GGML_HIP_ROCWMMA_FATTN=ON causes severe prefill regression with flash attention on gfx1151 (Strix Halo), up to −41% at long context",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-06-10T20:58:11Z",
        "updated_at": "2026-08-16T00:50:28Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24437",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 24438,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: ROCm/HIP backend achieves only ~40% of memory bandwidth on gfx1151 (Strix Halo) for MoE token generation",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-06-10T21:22:13Z",
        "updated_at": "2026-08-18T02:05:48Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24438",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24440,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: llama server crashes(fattn.cu:579: fatal error) after editing system message when using Gemma 4 31B with MTP and -sm tensor",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-06-10T22:09:10Z",
        "updated_at": "2026-07-29T16:32:41Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24440",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 24456,
        "kind": "pr",
        "state": "open",
        "title": "ggml: improve RVV q4_0 GEMM prefill locality",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-06-11T05:33:35Z",
        "updated_at": "2026-07-15T03:44:23Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24456",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing"
        ]
      },
      {
        "number": 24460,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: llama-convert-hf-to-gguf.py fails with: AttributeError: 'list' object has no attribute 'keys'",
        "labels": [
          "bug"
        ],
        "created_at": "2026-06-11T08:01:13Z",
        "updated_at": "2026-08-17T08:05:08Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24460",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "loader_conversion",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 24467,
        "kind": "pr",
        "state": "open",
        "title": "docs(server): add llama-server WebUI settings complete guide",
        "labels": [
          "documentation"
        ],
        "created_at": "2026-06-11T11:10:50Z",
        "updated_at": "2026-06-11T11:15:15Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24467",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24473,
        "kind": "issue",
        "state": "open",
        "title": "Feature: Compact Conversation Action",
        "labels": [],
        "created_at": "2026-06-11T13:31:51Z",
        "updated_at": "2026-08-19T18:17:37Z",
        "closed_at": null,
        "comments": 9,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24473",
        "categories": [
          "memory_kv_context",
          "api_ui_docs"
        ]
      },
      {
        "number": 24492,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Gemma 4 31B MTP (draft-mtp) crashes on Vulkan backend, pre-allocated tensor cannot run operation NONE",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-06-11T21:21:55Z",
        "updated_at": "2026-08-18T10:20:05Z",
        "closed_at": null,
        "comments": 13,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24492",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 24494,
        "kind": "pr",
        "state": "open",
        "title": "common: add sampling env vars",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2026-06-11T21:54:25Z",
        "updated_at": "2026-06-12T06:45:28Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24494",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "api_ui_docs"
        ]
      },
      {
        "number": 24519,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: --no-kv-offload causes immediate EOS generation with Qwen3.6-27B on Vulkan (works without flag)",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-06-12T13:44:06Z",
        "updated_at": "2026-08-16T12:27:13Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24519",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 24546,
        "kind": "pr",
        "state": "open",
        "title": "CUDA: size routed MoE MMQ N-tiles from typical expert width on RDNA3",
        "labels": [
          "Nvidia GPU",
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-06-12T21:01:19Z",
        "updated_at": "2026-08-24T18:57:51Z",
        "closed_at": null,
        "comments": 14,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24546",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 24549,
        "kind": "pr",
        "state": "open",
        "title": "llama : disable graph reuse when contexts share memory under SPLIT_MODE_TENSOR",
        "labels": [],
        "created_at": "2026-06-12T23:52:47Z",
        "updated_at": "2026-08-26T14:57:35Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24549",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "training_adapters"
        ]
      },
      {
        "number": 24554,
        "kind": "pr",
        "state": "open",
        "title": "Fix 24486: TP: allows the usage of 4-10 gpus for stepfun and laguna",
        "labels": [],
        "created_at": "2026-06-13T01:18:10Z",
        "updated_at": "2026-08-20T00:14:40Z",
        "closed_at": null,
        "comments": 11,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24554",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing"
        ]
      },
      {
        "number": 24559,
        "kind": "pr",
        "state": "open",
        "title": "ggml-alloc : check realloc result in alloc_tensor_range",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-06-13T02:03:40Z",
        "updated_at": "2026-06-13T02:22:16Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24559",
        "categories": [
          "native_backends",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 24565,
        "kind": "pr",
        "state": "open",
        "title": "[fattn-tune] Add Blackwell MMA config",
        "labels": [
          "Nvidia GPU",
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-06-13T07:04:40Z",
        "updated_at": "2026-07-27T15:13:23Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24565",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 24566,
        "kind": "pr",
        "state": "open",
        "title": "EXPERIMENT: meta: key external view cache by backend context",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-06-13T08:45:39Z",
        "updated_at": "2026-06-15T01:51:16Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24566",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24569,
        "kind": "pr",
        "state": "open",
        "title": "CUDA: Add conv3d.",
        "labels": [
          "Nvidia GPU",
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-06-13T11:24:17Z",
        "updated_at": "2026-07-03T12:10:49Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24569",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 24577,
        "kind": "pr",
        "state": "open",
        "title": "ui: (demo) access server remotely via webrtc",
        "labels": [
          "examples",
          "server/ui"
        ],
        "created_at": "2026-06-13T14:45:51Z",
        "updated_at": "2026-08-16T12:28:57Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24577",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24590,
        "kind": "pr",
        "state": "open",
        "title": " llama : suppress misleading Gemma4Assistant error during memory fitting",
        "labels": [],
        "created_at": "2026-06-13T19:26:22Z",
        "updated_at": "2026-06-18T18:33:52Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24590",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 24591,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Passing non-unique ids to CUDA ggml_mul_mat_id() implementation results in crash",
        "labels": [
          "bug",
          "wontfix",
          "CUDA"
        ],
        "created_at": "2026-06-13T20:16:39Z",
        "updated_at": "2026-08-05T10:49:34Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24591",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24616,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: add dp4a emulation via dp2a on cuda",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-06-14T16:29:05Z",
        "updated_at": "2026-08-10T22:07:13Z",
        "closed_at": null,
        "comments": 14,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24616",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24627,
        "kind": "pr",
        "state": "open",
        "title": "convert : reorder V heads for LoraTorchTensor",
        "labels": [
          "python"
        ],
        "created_at": "2026-06-14T21:22:59Z",
        "updated_at": "2026-06-14T22:54:54Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24627",
        "categories": [
          "native_backends",
          "models_architectures",
          "loader_conversion",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 24657,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: GGML_ASSERT(n_inputs < GGML_SCHED_MAX_SPLIT_INPUTS) - On more than one GPU",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-06-15T15:29:00Z",
        "updated_at": "2026-08-04T01:15:32Z",
        "closed_at": null,
        "comments": 13,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24657",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 24669,
        "kind": "pr",
        "state": "open",
        "title": "llama: add llama_batch_ext",
        "labels": [
          "testing"
        ],
        "created_at": "2026-06-15T21:15:21Z",
        "updated_at": "2026-08-25T16:50:41Z",
        "closed_at": null,
        "comments": 12,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24669",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 24670,
        "kind": "issue",
        "state": "open",
        "title": "draft-mtp speculative decoding not activating on Turing (sm_75) with hybrid SSM+attention model (Qwen3.6-35B-A3B)",
        "labels": [],
        "created_at": "2026-06-15T21:44:48Z",
        "updated_at": "2026-08-25T07:49:46Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24670",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "build_platform_ops"
        ]
      },
      {
        "number": 24680,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Web-UI uploaded images are not width-constrained",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-06-16T05:31:02Z",
        "updated_at": "2026-07-27T01:16:18Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24680",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 24681,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug:",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-06-16T05:45:36Z",
        "updated_at": "2026-08-12T01:08:47Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24681",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 24696,
        "kind": "pr",
        "state": "open",
        "title": "llama-eval : Add support for API keys",
        "labels": [
          "examples",
          "python"
        ],
        "created_at": "2026-06-16T16:54:20Z",
        "updated_at": "2026-07-11T09:30:42Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24696",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24700,
        "kind": "pr",
        "state": "open",
        "title": "ci(snapdragon): add Windows on ARM64 build and test via Qualcomm Device Cloud",
        "labels": [
          "documentation",
          "script",
          "python",
          "devops"
        ],
        "created_at": "2026-06-16T18:39:49Z",
        "updated_at": "2026-06-16T20:50:09Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24700",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24701,
        "kind": "issue",
        "state": "open",
        "title": "UI: Improve accessibility for screenreaders: Not able to delete old conversations",
        "labels": [],
        "created_at": "2026-06-16T19:03:50Z",
        "updated_at": "2026-07-27T01:16:16Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24701",
        "categories": [
          "models_architectures",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24712,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Warning Message - sched_reserve: layer 0 is assigned to device CPU but the fused Gated Delta Net tensor is assigned to device CUDA0 (usually due to missing support)",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-06-17T02:27:57Z",
        "updated_at": "2026-08-18T02:12:59Z",
        "closed_at": null,
        "comments": 16,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24712",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24714,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: forcing full prompt re-processing due to lack of cache data unsloth/Qwen3.5-2B-MTP-GGUF:Q4_K_XL",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-06-17T04:56:06Z",
        "updated_at": "2026-08-12T01:08:45Z",
        "closed_at": null,
        "comments": 18,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24714",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 24716,
        "kind": "pr",
        "state": "open",
        "title": "Feat/add modelscope -ms args support",
        "labels": [
          "server"
        ],
        "created_at": "2026-06-17T07:54:09Z",
        "updated_at": "2026-08-24T06:59:06Z",
        "closed_at": null,
        "comments": 10,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24716",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24720,
        "kind": "pr",
        "state": "open",
        "title": "vulkan: command buffer graph reuse",
        "labels": [
          "Vulkan",
          "ggml"
        ],
        "created_at": "2026-06-17T12:47:10Z",
        "updated_at": "2026-06-20T18:21:08Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24720",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 24729,
        "kind": "issue",
        "state": "open",
        "title": "Windows OpenVINO builds fail to start due to missing OpenSSL dependencies and incorrect release packaging",
        "labels": [],
        "created_at": "2026-06-17T16:19:20Z",
        "updated_at": "2026-08-05T01:10:42Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24729",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 24734,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Vulkan regression on AMD Strix Halo UMA between b9660 and b9672: model loads on GPU but inference shows no GPU compute",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-06-17T20:27:15Z",
        "updated_at": "2026-08-25T01:07:36Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24734",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing"
        ]
      },
      {
        "number": 24737,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Qwen3.5-4B: GGUF conversion/load expects 33 blocks, model only has 32",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-06-17T22:39:27Z",
        "updated_at": "2026-08-06T01:09:33Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24737",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 24738,
        "kind": "pr",
        "state": "open",
        "title": "convert : MXFP4 expert passthrough for Gemma4 (vllm_fused_moe)",
        "labels": [
          "python"
        ],
        "created_at": "2026-06-17T22:43:39Z",
        "updated_at": "2026-06-17T23:03:35Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24738",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24740,
        "kind": "pr",
        "state": "open",
        "title": "server : add GET/POST /cvectors for control vector hot-swap",
        "labels": [
          "examples",
          "python",
          "server"
        ],
        "created_at": "2026-06-18T00:09:17Z",
        "updated_at": "2026-06-18T00:12:23Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24740",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 24743,
        "kind": "pr",
        "state": "open",
        "title": "support MiniCPM5 XML tool calls",
        "labels": [
          "python",
          "jinja parser"
        ],
        "created_at": "2026-06-18T01:37:41Z",
        "updated_at": "2026-06-23T07:03:23Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24743",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24749,
        "kind": "pr",
        "state": "open",
        "title": "mtmd: add NVIDIA LocateAnything-3B vision support",
        "labels": [
          "examples",
          "python"
        ],
        "created_at": "2026-06-18T05:20:44Z",
        "updated_at": "2026-06-26T19:59:56Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24749",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24759,
        "kind": "pr",
        "state": "open",
        "title": "gguf: guard GGML_PAD inputs against integer overflow",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-06-18T09:05:23Z",
        "updated_at": "2026-06-18T10:15:05Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24759",
        "categories": [
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24761,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Bug of how web-ui provides MCP server to model",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-06-18T09:58:28Z",
        "updated_at": "2026-07-27T01:16:12Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24761",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24767,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug:",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-06-18T15:37:05Z",
        "updated_at": "2026-08-12T01:08:44Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24767",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 24768,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Gemma 4 MTP \"heuristic\" n-max",
        "labels": [
          "enhancement",
          "stale"
        ],
        "created_at": "2026-06-18T16:02:23Z",
        "updated_at": "2026-08-15T01:23:21Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24768",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24785,
        "kind": "pr",
        "state": "open",
        "title": "server: add recurrent state shrink/expand for prompt cache (#22746)",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2026-06-18T23:10:41Z",
        "updated_at": "2026-08-25T07:49:48Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24785",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24795,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: gemma4-assistant MTP draft model fails to load — \"invalid vector subscript\" (regression: works on b9553, broken on   b9702/b9717)",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-06-19T08:37:26Z",
        "updated_at": "2026-08-24T22:00:04Z",
        "closed_at": null,
        "comments": 10,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24795",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 24800,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Supports PP-OCRv6",
        "labels": [
          "enhancement",
          "stale"
        ],
        "created_at": "2026-06-19T10:37:51Z",
        "updated_at": "2026-08-13T01:08:43Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24800",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24805,
        "kind": "pr",
        "state": "open",
        "title": "Expose vocab.max_token_len to C-style API",
        "labels": [],
        "created_at": "2026-06-19T13:11:33Z",
        "updated_at": "2026-06-19T13:14:58Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24805",
        "categories": [
          "models_architectures",
          "api_ui_docs"
        ]
      },
      {
        "number": 24812,
        "kind": "issue",
        "state": "open",
        "title": "Qwen3.5 on Vulkan/RADV RX590 produces clustered garbage / question-mark output in warm llama-server streaming",
        "labels": [
          "stale"
        ],
        "created_at": "2026-06-19T17:52:50Z",
        "updated_at": "2026-08-20T02:06:37Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24812",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24822,
        "kind": "issue",
        "state": "open",
        "title": "Server: improve progress reporting",
        "labels": [],
        "created_at": "2026-06-19T20:28:08Z",
        "updated_at": "2026-08-20T12:39:53Z",
        "closed_at": null,
        "comments": 9,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24822",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "loader_conversion",
          "api_ui_docs"
        ]
      },
      {
        "number": 24836,
        "kind": "issue",
        "state": "open",
        "title": "Misc bug: CUDA/HIP: llama.cpp will have to find a different way on HIP target to calculate free VRAM",
        "labels": [
          "stale"
        ],
        "created_at": "2026-06-20T11:10:14Z",
        "updated_at": "2026-08-23T01:07:38Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24836",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 24840,
        "kind": "issue",
        "state": "open",
        "title": "Bug: server - update_slots(): batch_view offset is not propagated to ctx_dft",
        "labels": [],
        "created_at": "2026-06-20T14:09:34Z",
        "updated_at": "2026-07-28T19:28:32Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24840",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "correctness_testing"
        ]
      },
      {
        "number": 24844,
        "kind": "pr",
        "state": "open",
        "title": "metal : add DIAG_MASK_INF op",
        "labels": [
          "documentation",
          "ggml",
          "Apple Metal"
        ],
        "created_at": "2026-06-20T19:06:21Z",
        "updated_at": "2026-06-25T05:34:12Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24844",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 24850,
        "kind": "pr",
        "state": "open",
        "title": "WIP: server: expose cache, speculative-decode, and resource metrics",
        "labels": [
          "examples",
          "python",
          "server"
        ],
        "created_at": "2026-06-21T01:02:31Z",
        "updated_at": "2026-06-21T09:21:36Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24850",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24851,
        "kind": "pr",
        "state": "open",
        "title": "WIP: server: fan out /metrics across all models in router mode",
        "labels": [
          "examples",
          "python",
          "server"
        ],
        "created_at": "2026-06-21T01:02:37Z",
        "updated_at": "2026-06-21T09:23:51Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24851",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24855,
        "kind": "pr",
        "state": "open",
        "title": "vulkan: fix MoE models correctness after experts offloading on MoltenVK and revert old workarounds",
        "labels": [
          "Vulkan",
          "ggml"
        ],
        "created_at": "2026-06-21T02:14:49Z",
        "updated_at": "2026-07-28T10:19:35Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24855",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 24871,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: sampling penalties causes up to 20% text generation overhead",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-06-21T12:54:27Z",
        "updated_at": "2026-08-19T02:07:41Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24871",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24888,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Malformed tokenizer metadata can cause ASan heap-buffer-overflow reads in vocab load",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-06-22T01:14:47Z",
        "updated_at": "2026-08-01T01:14:23Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24888",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 24891,
        "kind": "pr",
        "state": "open",
        "title": "server: fix checkpoint invalidation after tool requests in multi-turn conversation",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2026-06-22T02:32:22Z",
        "updated_at": "2026-08-24T02:50:24Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24891",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24896,
        "kind": "pr",
        "state": "open",
        "title": "opencl: q8_0 GEMM using dot8 + local memory broadcast for Adreno",
        "labels": [
          "ggml",
          "OpenCL"
        ],
        "created_at": "2026-06-22T06:26:21Z",
        "updated_at": "2026-06-23T03:29:50Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24896",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 24898,
        "kind": "pr",
        "state": "open",
        "title": "server : support \"dimensions\" parameter in embeddings endpoint",
        "labels": [
          "documentation",
          "examples",
          "python",
          "server"
        ],
        "created_at": "2026-06-22T08:31:05Z",
        "updated_at": "2026-07-01T19:30:59Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24898",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 24901,
        "kind": "pr",
        "state": "open",
        "title": "POC/server disagg prefill decode",
        "labels": [
          "documentation",
          "examples",
          "server"
        ],
        "created_at": "2026-06-22T10:09:12Z",
        "updated_at": "2026-07-13T10:51:18Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24901",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 24902,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: pre-select models in the webui using alias",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-06-22T10:10:50Z",
        "updated_at": "2026-08-10T01:08:22Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24902",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 24905,
        "kind": "pr",
        "state": "open",
        "title": "vulkan: split ggml-vulkan.cpp file",
        "labels": [
          "Vulkan",
          "ggml"
        ],
        "created_at": "2026-06-22T14:04:37Z",
        "updated_at": "2026-08-07T16:08:41Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24905",
        "categories": [
          "native_backends",
          "models_architectures",
          "api_ui_docs"
        ]
      },
      {
        "number": 24906,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: ROCm backend reports incorrect free VRAM",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-06-22T14:31:21Z",
        "updated_at": "2026-08-14T01:08:40Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24906",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 24910,
        "kind": "pr",
        "state": "open",
        "title": "vocab : validate tokenizer metadata before typed reads",
        "labels": [
          "testing"
        ],
        "created_at": "2026-06-22T17:01:53Z",
        "updated_at": "2026-08-18T11:41:14Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24910",
        "categories": [
          "quantization",
          "models_architectures",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 24914,
        "kind": "pr",
        "state": "open",
        "title": "server: run sampling in a threadpool ",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2026-06-22T17:42:31Z",
        "updated_at": "2026-06-25T17:06:27Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24914",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "api_ui_docs"
        ]
      },
      {
        "number": 24916,
        "kind": "pr",
        "state": "open",
        "title": "vocab : reject invalid tokenizer metadata",
        "labels": [
          "testing"
        ],
        "created_at": "2026-06-22T18:35:32Z",
        "updated_at": "2026-06-23T23:51:44Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24916",
        "categories": [
          "quantization",
          "models_architectures",
          "loader_conversion",
          "correctness_testing"
        ]
      },
      {
        "number": 24925,
        "kind": "pr",
        "state": "open",
        "title": "Minimax M3 EAGLE3 Support",
        "labels": [
          "model",
          "testing",
          "python"
        ],
        "created_at": "2026-06-23T00:28:18Z",
        "updated_at": "2026-06-24T19:03:01Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24925",
        "categories": [
          "quantization",
          "models_architectures",
          "speculative_generation",
          "structured_generation_tools",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 24929,
        "kind": "pr",
        "state": "open",
        "title": "docs: note ROCm HIP SDK 7.1.1 + MSVC >=14.40 `<cmath>` build failure and workaround (Windows)",
        "labels": [
          "documentation"
        ],
        "created_at": "2026-06-23T02:50:42Z",
        "updated_at": "2026-06-23T02:51:09Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24929",
        "categories": [
          "native_backends",
          "models_architectures",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24932,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Agent mode, shelling command executing can fail caused by whitespace in path",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-06-23T04:13:51Z",
        "updated_at": "2026-08-25T01:07:33Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24932",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24934,
        "kind": "pr",
        "state": "open",
        "title": "Lllama sampler speedups/fixes",
        "labels": [
          "testing"
        ],
        "created_at": "2026-06-23T07:28:02Z",
        "updated_at": "2026-08-23T20:38:33Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24934",
        "categories": [
          "models_architectures",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 24937,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: CUDA: llama-bench prefill crashes on MoE/SSM models (ggml_cuda_mul_mat_q / mm_ids_helper) — regression since a6cc43c28; llama-cli OK",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-06-23T08:20:49Z",
        "updated_at": "2026-08-15T01:23:19Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24937",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24946,
        "kind": "issue",
        "state": "open",
        "title": "[SYCL/xe] -cb pins GPU at gt-c0 on Battlemage, prevents idle power savings",
        "labels": [],
        "created_at": "2026-06-23T13:20:14Z",
        "updated_at": "2026-08-12T15:40:58Z",
        "closed_at": null,
        "comments": 15,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24946",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24953,
        "kind": "pr",
        "state": "open",
        "title": "bench: Fix misc. bug #24951 - Standard Deviation issues",
        "labels": [
          "examples"
        ],
        "created_at": "2026-06-23T19:10:42Z",
        "updated_at": "2026-06-24T14:28:22Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24953",
        "categories": [
          "models_architectures",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24956,
        "kind": "pr",
        "state": "open",
        "title": "server : create context checkpoint on slot restore",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2026-06-23T20:33:21Z",
        "updated_at": "2026-06-23T20:33:47Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24956",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24957,
        "kind": "pr",
        "state": "open",
        "title": "Improve Server OAI Responses API streaming compatibility",
        "labels": [
          "examples",
          "server"
        ],
        "created_at": "2026-06-23T20:55:08Z",
        "updated_at": "2026-07-22T22:25:19Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24957",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24966,
        "kind": "pr",
        "state": "open",
        "title": "vulkan: disable MMVQ on AMD UMA devices",
        "labels": [
          "Vulkan",
          "ggml"
        ],
        "created_at": "2026-06-24T06:29:55Z",
        "updated_at": "2026-07-13T06:22:18Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24966",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 24975,
        "kind": "pr",
        "state": "open",
        "title": "llama: R-SWA reference sliding window attention for Unlimited-OCR",
        "labels": [
          "model",
          "examples",
          "python",
          "mtmd",
          "conversion"
        ],
        "created_at": "2026-06-24T14:32:21Z",
        "updated_at": "2026-08-25T07:11:10Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24975",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24988,
        "kind": "pr",
        "state": "open",
        "title": "gguf : reject non-u32 general.alignment",
        "labels": [
          "testing",
          "ggml"
        ],
        "created_at": "2026-06-24T21:22:53Z",
        "updated_at": "2026-08-10T20:19:51Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24988",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 24989,
        "kind": "pr",
        "state": "open",
        "title": "arg: detect console width dynamically for CLI help wrapping",
        "labels": [],
        "created_at": "2026-06-24T21:54:24Z",
        "updated_at": "2026-06-29T19:05:16Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24989",
        "categories": [
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 24992,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Tools doesn't work in WebUI when running in router mode",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-06-25T02:28:29Z",
        "updated_at": "2026-08-26T01:07:50Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24992",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 24997,
        "kind": "pr",
        "state": "open",
        "title": "gguf: add upper bound check for general.alignment (CWE-1284)",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-06-25T09:12:28Z",
        "updated_at": "2026-06-30T12:53:27Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/24997",
        "categories": [
          "quantization",
          "models_architectures",
          "loader_conversion"
        ]
      },
      {
        "number": 24999,
        "kind": "issue",
        "state": "open",
        "title": "CUDA error: the requested functionality is not supported",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-06-25T09:55:25Z",
        "updated_at": "2026-08-25T19:14:51Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/24999",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25004,
        "kind": "pr",
        "state": "open",
        "title": "recurrent : support equal splits for recurrent-state rollback",
        "labels": [
          "model",
          "testing"
        ],
        "created_at": "2026-06-25T12:40:31Z",
        "updated_at": "2026-08-08T15:49:24Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25004",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25015,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: backend sampling has to be explicitly enabled in server ui even when specified via cli",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-06-25T17:33:10Z",
        "updated_at": "2026-07-27T01:15:59Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25015",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 25030,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: add builds for arm64 windows with CUDA",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-06-26T06:29:51Z",
        "updated_at": "2026-08-20T19:22:19Z",
        "closed_at": null,
        "comments": 16,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25030",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25032,
        "kind": "pr",
        "state": "open",
        "title": "ggml : unify tie-breaking to first index across all backends",
        "labels": [
          "testing",
          "Vulkan",
          "ggml",
          "SYCL",
          "Apple Metal",
          "CUDA",
          "WebGPU"
        ],
        "created_at": "2026-06-26T08:28:53Z",
        "updated_at": "2026-06-26T09:50:06Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25032",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing"
        ]
      },
      {
        "number": 25034,
        "kind": "pr",
        "state": "open",
        "title": "ui: Skills library + improvements for System Message, MCP Prompts & Resources UI/UX",
        "labels": [
          "documentation",
          "server/ui"
        ],
        "created_at": "2026-06-26T08:46:50Z",
        "updated_at": "2026-07-16T07:11:29Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25034",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25041,
        "kind": "pr",
        "state": "open",
        "title": "[SYCL] update Q&A for saving power",
        "labels": [
          "documentation",
          "SYCL"
        ],
        "created_at": "2026-06-26T10:41:35Z",
        "updated_at": "2026-07-07T08:22:05Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25041",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "api_ui_docs"
        ]
      },
      {
        "number": 25048,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cpu: replace cyclic chunk distribution with atomic work-stealing",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-06-26T12:03:46Z",
        "updated_at": "2026-08-20T19:23:48Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25048",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25051,
        "kind": "pr",
        "state": "open",
        "title": "vulkan: add allreduce function with cross-device CPU proxy and fix Tensor Parallel crash [EXPERIMENTAL]",
        "labels": [
          "Vulkan",
          "ggml"
        ],
        "created_at": "2026-06-26T12:37:56Z",
        "updated_at": "2026-08-16T13:44:58Z",
        "closed_at": null,
        "comments": 48,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25051",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25052,
        "kind": "pr",
        "state": "open",
        "title": "server : drop the need for vendor subprocess.h",
        "labels": [
          "server"
        ],
        "created_at": "2026-06-26T12:46:59Z",
        "updated_at": "2026-06-26T21:14:50Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25052",
        "categories": [
          "models_architectures",
          "scheduling_serving"
        ]
      },
      {
        "number": 25060,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Blackwell GGML-CUDA SOFT_MAX Crash",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-06-26T17:37:43Z",
        "updated_at": "2026-08-20T05:44:22Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25060",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25061,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: CUDA error: unsupported value or parameter in cublasSgemm_v2 during large context processing",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-06-26T17:43:58Z",
        "updated_at": "2026-08-25T01:07:32Z",
        "closed_at": null,
        "comments": 10,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25061",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25067,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Premature \"reasoning-budget: deactivated (natural end)\", even BEFORE prompt processing",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-06-26T23:06:03Z",
        "updated_at": "2026-07-29T07:00:04Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25067",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 25068,
        "kind": "pr",
        "state": "open",
        "title": "docs: add Apple A-chipsets model size estimation guide",
        "labels": [
          "documentation"
        ],
        "created_at": "2026-06-26T23:22:09Z",
        "updated_at": "2026-06-26T23:26:22Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25068",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25082,
        "kind": "issue",
        "state": "open",
        "title": "HIP/ROCm: system RAM grows unbounded with parallel slots due to CUDA graph cache never being evicted",
        "labels": [],
        "created_at": "2026-06-27T17:20:15Z",
        "updated_at": "2026-08-03T01:15:01Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25082",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25088,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Qwen3-VL image embedding doesn't work",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-06-28T02:02:11Z",
        "updated_at": "2026-08-14T01:08:38Z",
        "closed_at": null,
        "comments": 17,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25088",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing"
        ]
      },
      {
        "number": 25089,
        "kind": "pr",
        "state": "open",
        "title": "sycl: fix check_graph_compatibility() to allow graphs for MoE decode (CONCAT dim!=3, MUL_MAT_ID fused path)",
        "labels": [
          "ggml",
          "SYCL"
        ],
        "created_at": "2026-06-28T04:19:31Z",
        "updated_at": "2026-07-28T05:02:06Z",
        "closed_at": null,
        "comments": 9,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25089",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 25109,
        "kind": "issue",
        "state": "open",
        "title": "Research: BGE-M3 multi head support available",
        "labels": [
          "research 🔬"
        ],
        "created_at": "2026-06-28T17:20:38Z",
        "updated_at": "2026-06-28T18:06:14Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25109",
        "categories": [
          "quantization",
          "models_architectures",
          "multimodal_embedding",
          "api_ui_docs"
        ]
      },
      {
        "number": 25117,
        "kind": "issue",
        "state": "open",
        "title": "DFlash performance regression on AMD APU + quantized MoE target: ~2x slower than baseline",
        "labels": [],
        "created_at": "2026-06-29T03:53:29Z",
        "updated_at": "2026-08-12T13:20:53Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25117",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 25119,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cpu: fix NEON build compilation on 32-bit ARMv7 architectures without hardware FP16",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-06-29T04:30:50Z",
        "updated_at": "2026-08-17T17:43:29Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25119",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25129,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: MTP breaks multimodality in StepFun Step-3.7-Flash \"inconsistent sequence positions\", \"failed to process speculative batch\"",
        "labels": [
          "bug"
        ],
        "created_at": "2026-06-29T10:36:56Z",
        "updated_at": "2026-06-29T17:22:31Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25129",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25133,
        "kind": "pr",
        "state": "open",
        "title": "llama : add position-relocatable KV range save/load",
        "labels": [
          "model",
          "testing",
          "examples"
        ],
        "created_at": "2026-06-29T14:36:10Z",
        "updated_at": "2026-08-12T22:51:47Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25133",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25144,
        "kind": "pr",
        "state": "open",
        "title": "speculative: fix MTP draft crash on vision inputs",
        "labels": [],
        "created_at": "2026-06-29T19:15:29Z",
        "updated_at": "2026-07-08T15:38:10Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25144",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "correctness_testing"
        ]
      },
      {
        "number": 25150,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: preset names changed and some presets ignored",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-06-29T22:42:15Z",
        "updated_at": "2026-08-11T10:54:28Z",
        "closed_at": null,
        "comments": 10,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25150",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25152,
        "kind": "pr",
        "state": "open",
        "title": "common, server : preserve HF file for cached models",
        "labels": [
          "server"
        ],
        "created_at": "2026-06-29T23:44:43Z",
        "updated_at": "2026-06-29T23:45:21Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25152",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing"
        ]
      },
      {
        "number": 25153,
        "kind": "pr",
        "state": "open",
        "title": "ggml: imatrix-aware NVFP4 quantization (scale search) + wire NVFP4 ftype",
        "labels": [
          "examples",
          "ggml"
        ],
        "created_at": "2026-06-30T00:59:37Z",
        "updated_at": "2026-08-05T17:18:08Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25153",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "speculative_generation",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 25161,
        "kind": "pr",
        "state": "open",
        "title": "ggml : fix wrong transpose function for int16 data",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-06-30T07:13:18Z",
        "updated_at": "2026-07-01T09:29:46Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25161",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25166,
        "kind": "pr",
        "state": "open",
        "title": "hexagon: allow dflash lm-head offload experiment",
        "labels": [
          "model",
          "examples",
          "ggml",
          "Hexagon"
        ],
        "created_at": "2026-06-30T09:43:59Z",
        "updated_at": "2026-07-06T06:17:19Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25166",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25171,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Deepseek V4 Flash forgets everything",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-06-30T13:06:24Z",
        "updated_at": "2026-08-05T01:10:22Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25171",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25187,
        "kind": "issue",
        "state": "open",
        "title": "Research: FR-Spec-style draft-vocab trimming for native MTP speculative decoding",
        "labels": [],
        "created_at": "2026-07-01T00:33:54Z",
        "updated_at": "2026-08-15T22:33:46Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25187",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 25189,
        "kind": "pr",
        "state": "open",
        "title": "openvino: fix SWA mask detection for long prompts",
        "labels": [
          "ggml",
          "OpenVINO"
        ],
        "created_at": "2026-07-01T02:56:16Z",
        "updated_at": "2026-07-01T02:56:42Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25189",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25201,
        "kind": "pr",
        "state": "open",
        "title": "llama-completion: respect chat_template_kwargs and reasoning_format params",
        "labels": [
          "examples"
        ],
        "created_at": "2026-07-01T12:41:57Z",
        "updated_at": "2026-08-26T20:05:38Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25201",
        "categories": [
          "models_architectures",
          "api_ui_docs"
        ]
      },
      {
        "number": 25203,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: [SYCL] Q8_0 reorder degrades 42% prefill performance (not visible in the standard llama-bench)",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-07-01T13:18:23Z",
        "updated_at": "2026-08-13T01:08:29Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25203",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25206,
        "kind": "pr",
        "state": "open",
        "title": "llama: optimize RWKV7 inference by fusing some graph operators",
        "labels": [
          "model",
          "testing",
          "Vulkan",
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-07-01T16:41:15Z",
        "updated_at": "2026-08-26T15:49:56Z",
        "closed_at": null,
        "comments": 9,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25206",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25207,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Unexpected massive performace drop with Vulkan Flash Attention",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-01T19:35:32Z",
        "updated_at": "2026-08-24T06:41:28Z",
        "closed_at": null,
        "comments": 18,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25207",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 25210,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: (OpenAI compat) unsupported \"dimensions\" parameter in embeddings endpoint",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-01T20:32:22Z",
        "updated_at": "2026-08-07T01:39:18Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25210",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 25214,
        "kind": "pr",
        "state": "open",
        "title": "server: add --no-sleep flag for GPU heartbeat on headless GPUs",
        "labels": [
          "Vulkan",
          "server",
          "ggml",
          "SYCL",
          "CUDA"
        ],
        "created_at": "2026-07-01T21:02:10Z",
        "updated_at": "2026-07-14T03:46:34Z",
        "closed_at": null,
        "comments": 15,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25214",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "training_adapters"
        ]
      },
      {
        "number": 25226,
        "kind": "pr",
        "state": "open",
        "title": "server : don't list cached models when a preset is used",
        "labels": [
          "server"
        ],
        "created_at": "2026-07-02T07:38:10Z",
        "updated_at": "2026-07-02T10:23:28Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25226",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving"
        ]
      },
      {
        "number": 25227,
        "kind": "issue",
        "state": "open",
        "title": "webui: model selector — org-less models visually attach to the previous org group",
        "labels": [],
        "created_at": "2026-07-02T08:10:54Z",
        "updated_at": "2026-08-16T17:58:54Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25227",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "api_ui_docs"
        ]
      },
      {
        "number": 25232,
        "kind": "pr",
        "state": "open",
        "title": "llama : clear error when MTP draft shares KV cache across backends",
        "labels": [],
        "created_at": "2026-07-02T09:51:37Z",
        "updated_at": "2026-08-13T03:59:23Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25232",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25235,
        "kind": "pr",
        "state": "open",
        "title": "common,server : fix custom preset dedup against cached models",
        "labels": [
          "server"
        ],
        "created_at": "2026-07-02T10:21:39Z",
        "updated_at": "2026-07-03T11:39:37Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25235",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving"
        ]
      },
      {
        "number": 25259,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: DeepSeek V4 Flash forgot context when hit cache or use checkpoint in long context",
        "labels": [
          "bug"
        ],
        "created_at": "2026-07-03T06:14:52Z",
        "updated_at": "2026-07-07T15:51:56Z",
        "closed_at": null,
        "comments": 14,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25259",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25261,
        "kind": "pr",
        "state": "open",
        "title": "feat: add --threads-all option to llama-bench",
        "labels": [
          "examples"
        ],
        "created_at": "2026-07-03T06:48:38Z",
        "updated_at": "2026-07-03T06:48:59Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25261",
        "categories": [
          "models_architectures",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25266,
        "kind": "pr",
        "state": "open",
        "title": "[SYC:] support op get_rows_back, only support fp32/fp16",
        "labels": [
          "documentation",
          "ggml",
          "SYCL"
        ],
        "created_at": "2026-07-03T11:47:09Z",
        "updated_at": "2026-07-03T11:47:26Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25266",
        "categories": [
          "native_backends",
          "api_ui_docs"
        ]
      },
      {
        "number": 25271,
        "kind": "pr",
        "state": "open",
        "title": "ggml, server: add ggml_backend_dev_reset() for sleep mode",
        "labels": [
          "Vulkan",
          "server",
          "ggml",
          "SYCL",
          "Apple Metal",
          "Ascend NPU",
          "OpenCL",
          "IBM zDNN",
          "Hexagon",
          "CUDA",
          "AMD ZenDNN",
          "OpenVINO",
          "WebGPU"
        ],
        "created_at": "2026-07-03T15:41:45Z",
        "updated_at": "2026-07-05T16:32:46Z",
        "closed_at": null,
        "comments": 14,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25271",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing"
        ]
      },
      {
        "number": 25273,
        "kind": "pr",
        "state": "open",
        "title": "CANN: Refactor `#ifdef` blocks to avoid unreachable code after `return`",
        "labels": [
          "ggml",
          "Ascend NPU"
        ],
        "created_at": "2026-07-03T16:35:53Z",
        "updated_at": "2026-08-12T14:33:45Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25273",
        "categories": [
          "native_backends",
          "models_architectures",
          "build_platform_ops"
        ]
      },
      {
        "number": 25277,
        "kind": "pr",
        "state": "open",
        "title": "common : resolve non-positive --threads to the number of math cores",
        "labels": [
          "documentation",
          "testing",
          "server"
        ],
        "created_at": "2026-07-03T17:45:02Z",
        "updated_at": "2026-07-03T17:49:03Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25277",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25279,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Releases are not created for UI only PRs",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-03T18:03:14Z",
        "updated_at": "2026-07-26T21:36:25Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25279",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25285,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: failed to tokenize error if contents of /props appears in prompt.",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-03T19:27:35Z",
        "updated_at": "2026-08-17T01:07:23Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25285",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25293,
        "kind": "issue",
        "state": "open",
        "title": "[Docs] --ubatch-size behavior in --embeddings mode (pooling) is undocumented — causes silent 500 errors",
        "labels": [
          "stale"
        ],
        "created_at": "2026-07-04T04:37:09Z",
        "updated_at": "2026-08-17T01:07:20Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25293",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25294,
        "kind": "pr",
        "state": "open",
        "title": "llama : stream MoE routed experts from disk ",
        "labels": [],
        "created_at": "2026-07-04T04:51:20Z",
        "updated_at": "2026-08-25T21:12:17Z",
        "closed_at": null,
        "comments": 23,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25294",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25297,
        "kind": "pr",
        "state": "open",
        "title": "kv-cache : fix SWA state save/load round-trip past n_swa",
        "labels": [],
        "created_at": "2026-07-04T09:32:55Z",
        "updated_at": "2026-07-04T09:32:56Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25297",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 25299,
        "kind": "issue",
        "state": "open",
        "title": "Remote unauthenticated NULL-pointer dereference in ggml-rpc graph_compute() via a node id of 0 (ggml_graph_plan/ggml_is_empty)",
        "labels": [
          "stale"
        ],
        "created_at": "2026-07-04T13:07:10Z",
        "updated_at": "2026-08-14T01:08:23Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25299",
        "categories": [
          "native_backends",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 25304,
        "kind": "issue",
        "state": "open",
        "title": "[CUDA] cublasCreate_v2 resource allocation failure on first inference — regression between b9553 and b9870",
        "labels": [],
        "created_at": "2026-07-04T16:01:22Z",
        "updated_at": "2026-08-26T18:50:08Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25304",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25308,
        "kind": "pr",
        "state": "open",
        "title": "json-schema-to-grammar: harden _visit_pattern against malformed regex patterns",
        "labels": [],
        "created_at": "2026-07-04T19:09:31Z",
        "updated_at": "2026-08-03T21:16:35Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25308",
        "categories": [
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25309,
        "kind": "pr",
        "state": "open",
        "title": "loader: map sparse mmap tensor ranges",
        "labels": [],
        "created_at": "2026-07-04T20:59:43Z",
        "updated_at": "2026-07-04T20:59:43Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25309",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "build_platform_ops"
        ]
      },
      {
        "number": 25318,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: RTX 5070 CUDA drivers crash with MTP",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-05T12:01:56Z",
        "updated_at": "2026-08-18T02:05:09Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25318",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25319,
        "kind": "pr",
        "state": "open",
        "title": "ggml: asynchronous scheduler memory copies",
        "labels": [
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-07-05T12:37:31Z",
        "updated_at": "2026-07-28T09:48:29Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25319",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 25323,
        "kind": "pr",
        "state": "open",
        "title": "ggml : use kernel type for conv_1d im2col to support f32",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-07-05T15:10:45Z",
        "updated_at": "2026-07-05T15:14:48Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25323",
        "categories": [
          "native_backends",
          "models_architectures"
        ]
      },
      {
        "number": 25332,
        "kind": "pr",
        "state": "open",
        "title": "chat : allow gpt-oss non-streaming final without channel markup (#25321)",
        "labels": [
          "testing"
        ],
        "created_at": "2026-07-05T18:54:03Z",
        "updated_at": "2026-07-06T23:05:38Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25332",
        "categories": [
          "models_architectures",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25333,
        "kind": "issue",
        "state": "open",
        "title": "[bug] Built-in tools do not work when llama-server is running in router mode (--models-preset)",
        "labels": [],
        "created_at": "2026-07-05T19:27:33Z",
        "updated_at": "2026-08-22T14:15:53Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25333",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25334,
        "kind": "pr",
        "state": "open",
        "title": "fix: load qwen35 Ollama GGUF exports",
        "labels": [
          "model"
        ],
        "created_at": "2026-07-05T22:24:11Z",
        "updated_at": "2026-07-05T22:24:25Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25334",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "speculative_generation",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25335,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: No \"Final estimate PPL\" when running llama-perplexity with stride option",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-05T23:36:22Z",
        "updated_at": "2026-08-02T23:50:17Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25335",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25336,
        "kind": "pr",
        "state": "open",
        "title": "ggml : add E4M3 (fp8) CPU quantization type",
        "labels": [
          "testing",
          "examples",
          "ggml",
          "Apple Metal",
          "conversion"
        ],
        "created_at": "2026-07-05T23:59:35Z",
        "updated_at": "2026-07-16T11:29:05Z",
        "closed_at": null,
        "comments": 9,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25336",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25341,
        "kind": "pr",
        "state": "open",
        "title": "cann: upgrade flash attention to V5 and add SWIGLU_OAI/ADD_ID ops",
        "labels": [
          "devops",
          "ggml",
          "Ascend NPU"
        ],
        "created_at": "2026-07-06T07:12:09Z",
        "updated_at": "2026-07-06T07:19:15Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25341",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 25342,
        "kind": "pr",
        "state": "open",
        "title": "GigaChat 3.5",
        "labels": [
          "model",
          "testing",
          "conversion"
        ],
        "created_at": "2026-07-06T07:20:04Z",
        "updated_at": "2026-08-22T22:24:19Z",
        "closed_at": null,
        "comments": 10,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25342",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "speculative_generation",
          "correctness_testing"
        ]
      },
      {
        "number": 25343,
        "kind": "pr",
        "state": "open",
        "title": "OpenCL: fix sub_group_shuffle_xor compatibility for Qualcomm Adreno GPUs",
        "labels": [
          "ggml",
          "OpenCL"
        ],
        "created_at": "2026-07-06T07:57:03Z",
        "updated_at": "2026-07-06T07:57:55Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25343",
        "categories": [
          "native_backends",
          "quantization",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25346,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cpu: detect AVX-VNNI in MSVC native builds",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-07-06T09:08:32Z",
        "updated_at": "2026-07-06T10:30:38Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25346",
        "categories": [
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25353,
        "kind": "pr",
        "state": "open",
        "title": "ggml : make quantization bit-reproducible across compilers and architectures (-ffp-contract=off on ggml-quants.c)",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-07-06T12:21:10Z",
        "updated_at": "2026-07-06T21:24:01Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25353",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25356,
        "kind": "issue",
        "state": "open",
        "title": "Vulkan: batched decode throughput cliff at n_tokens=9 on many-expert MoE (fixed 8-token thresholds in MMV dispatch)",
        "labels": [],
        "created_at": "2026-07-06T14:24:26Z",
        "updated_at": "2026-08-25T01:07:27Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25356",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25362,
        "kind": "pr",
        "state": "open",
        "title": "vulkan: add NonTemporal hints to weight loads for AMD L2 cache optimization",
        "labels": [
          "Vulkan",
          "ggml"
        ],
        "created_at": "2026-07-06T16:32:54Z",
        "updated_at": "2026-07-06T16:49:06Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25362",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25365,
        "kind": "pr",
        "state": "open",
        "title": "vulkan: add UMA dual-access mode for zero-copy CPU access on iGPUs",
        "labels": [
          "Vulkan",
          "ggml"
        ],
        "created_at": "2026-07-06T16:52:44Z",
        "updated_at": "2026-07-06T16:56:49Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25365",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "build_platform_ops"
        ]
      },
      {
        "number": 25369,
        "kind": "pr",
        "state": "open",
        "title": "[RFC] autoparser unable to accommodate non-ASCII reasoning markers, e.g. `◁think▷`",
        "labels": [],
        "created_at": "2026-07-06T20:16:05Z",
        "updated_at": "2026-08-04T09:08:01Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25369",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25375,
        "kind": "pr",
        "state": "open",
        "title": "Add Bodega one to UIs",
        "labels": [
          "documentation"
        ],
        "created_at": "2026-07-06T22:44:09Z",
        "updated_at": "2026-07-06T22:44:26Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25375",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "api_ui_docs"
        ]
      },
      {
        "number": 25377,
        "kind": "pr",
        "state": "open",
        "title": "metal : add 64x8 mul_mm tile variant for small batches (q4_0)",
        "labels": [
          "ggml",
          "Apple Metal"
        ],
        "created_at": "2026-07-07T00:01:41Z",
        "updated_at": "2026-07-29T13:36:31Z",
        "closed_at": null,
        "comments": 10,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25377",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 25384,
        "kind": "pr",
        "state": "open",
        "title": "CUDA: check UMA before cudaMemGetInfo in device_get_memory",
        "labels": [
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-07-07T05:37:59Z",
        "updated_at": "2026-07-14T08:12:29Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25384",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing"
        ]
      },
      {
        "number": 25387,
        "kind": "pr",
        "state": "open",
        "title": "readme : add Lynkr to Infrastructure section",
        "labels": [
          "documentation"
        ],
        "created_at": "2026-07-07T06:15:47Z",
        "updated_at": "2026-07-18T20:50:13Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25387",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "api_ui_docs"
        ]
      },
      {
        "number": 25392,
        "kind": "issue",
        "state": "open",
        "title": "Support for gemini-nano (weights.bin)",
        "labels": [
          "stale"
        ],
        "created_at": "2026-07-07T08:38:40Z",
        "updated_at": "2026-08-16T01:07:29Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25392",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "multimodal_embedding",
          "loader_conversion"
        ]
      },
      {
        "number": 25397,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cpu: Added RVV VLEN=1024 vector dot product (vec_dot) kernels for quantized types.",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-07-07T10:04:10Z",
        "updated_at": "2026-07-13T09:01:39Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25397",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25413,
        "kind": "issue",
        "state": "open",
        "title": "Compile bug: Windows compilation with BLAS support using AOCL as BLAS vendor fails (using OpenBLAS works)",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-07-07T17:36:58Z",
        "updated_at": "2026-08-15T01:22:52Z",
        "closed_at": null,
        "comments": 14,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25413",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25423,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Very long (20+ minutes) load times when using SYCL tensor parallelism",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-07-07T22:35:06Z",
        "updated_at": "2026-08-26T01:07:46Z",
        "closed_at": null,
        "comments": 9,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25423",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 25426,
        "kind": "pr",
        "state": "open",
        "title": "common : validate dataset size before computing datapoint count",
        "labels": [],
        "created_at": "2026-07-08T01:17:09Z",
        "updated_at": "2026-07-13T12:19:14Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25426",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25428,
        "kind": "pr",
        "state": "open",
        "title": "qwen2 : use no-cache attention path for training graphs",
        "labels": [
          "model"
        ],
        "created_at": "2026-07-08T01:27:50Z",
        "updated_at": "2026-07-13T12:19:37Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25428",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 25429,
        "kind": "issue",
        "state": "open",
        "title": "Compile bug:",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-07-08T02:51:15Z",
        "updated_at": "2026-08-19T02:07:27Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25429",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25436,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: DeepSeep V4 garbled output on Strix Halo with ROCm",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-08T07:33:42Z",
        "updated_at": "2026-08-21T07:43:12Z",
        "closed_at": null,
        "comments": 28,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25436",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 25437,
        "kind": "issue",
        "state": "open",
        "title": "prompt_clear() doesn't free checkpoints",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-08T09:06:28Z",
        "updated_at": "2026-08-21T02:13:06Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25437",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "training_adapters"
        ]
      },
      {
        "number": 25438,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: OpenCL op not supported node_25 (SIN)",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-07-08T09:16:09Z",
        "updated_at": "2026-08-12T01:08:14Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25438",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 25439,
        "kind": "pr",
        "state": "open",
        "title": "model : Gemma4 12B model type detection",
        "labels": [
          "model"
        ],
        "created_at": "2026-07-08T09:36:50Z",
        "updated_at": "2026-07-08T09:37:05Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25439",
        "categories": [
          "native_backends",
          "models_architectures"
        ]
      },
      {
        "number": 25443,
        "kind": "issue",
        "state": "open",
        "title": "Compile bug: stale tools/ui/dist fails the whole llama-server build even with LLAMA_BUILD_UI=OFF and LLAMA_USE_PREBUILT_UI=OFF",
        "labels": [
          "stale"
        ],
        "created_at": "2026-07-08T13:37:58Z",
        "updated_at": "2026-08-25T01:07:25Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25443",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25444,
        "kind": "pr",
        "state": "open",
        "title": "model: add NVIDIA Nemotron-3-Puzzle-75B-A9B (NemotronHPuzzle) support",
        "labels": [
          "model",
          "conversion"
        ],
        "created_at": "2026-07-08T13:55:53Z",
        "updated_at": "2026-08-26T19:14:32Z",
        "closed_at": null,
        "comments": 10,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25444",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25452,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug:  DSV4-Flash churned-reuse SWA KV-cache exhaustion (crash + stall)",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-08T15:53:03Z",
        "updated_at": "2026-08-15T11:38:52Z",
        "closed_at": null,
        "comments": 12,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25452",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25453,
        "kind": "pr",
        "state": "open",
        "title": "metal: extend small-batch mat-vec dispatch to 16 rows",
        "labels": [
          "ggml",
          "Apple Metal"
        ],
        "created_at": "2026-07-08T16:09:55Z",
        "updated_at": "2026-07-08T16:14:12Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25453",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25454,
        "kind": "pr",
        "state": "open",
        "title": "common/autoparser : support tools with response_format in autoparser",
        "labels": [
          "testing"
        ],
        "created_at": "2026-07-08T16:24:20Z",
        "updated_at": "2026-07-31T23:13:37Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25454",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25455,
        "kind": "issue",
        "state": "open",
        "title": "SYCL: MUL_MAT_ID prefill path produces wrong results on Intel Arc Pro B70 (Battlemage G31) — causes garbage output on MoE models",
        "labels": [
          "stale"
        ],
        "created_at": "2026-07-08T16:25:20Z",
        "updated_at": "2026-08-15T01:22:48Z",
        "closed_at": null,
        "comments": 13,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25455",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25462,
        "kind": "pr",
        "state": "open",
        "title": "/slots related API fixes",
        "labels": [
          "documentation",
          "server"
        ],
        "created_at": "2026-07-08T18:54:34Z",
        "updated_at": "2026-08-10T16:49:26Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25462",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25479,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cuda : optimize mmvq nwarps for Pascal DP4A",
        "labels": [
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-07-09T10:23:26Z",
        "updated_at": "2026-07-29T18:09:58Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25479",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25483,
        "kind": "pr",
        "state": "open",
        "title": "vulkan: skip unneeded MoE work in mul_mm coopmat1 path",
        "labels": [
          "Vulkan",
          "ggml"
        ],
        "created_at": "2026-07-09T11:42:06Z",
        "updated_at": "2026-07-28T01:54:15Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25483",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25489,
        "kind": "issue",
        "state": "open",
        "title": "MTP feature performance drops since b9935",
        "labels": [
          "bug",
          "regression",
          "speculative"
        ],
        "created_at": "2026-07-09T14:37:07Z",
        "updated_at": "2026-08-20T14:47:00Z",
        "closed_at": null,
        "comments": 13,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25489",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25493,
        "kind": "pr",
        "state": "open",
        "title": "Feat/optimize q8 0",
        "labels": [
          "Vulkan",
          "ggml",
          "CUDA",
          "conversion"
        ],
        "created_at": "2026-07-09T15:51:55Z",
        "updated_at": "2026-08-10T06:40:25Z",
        "closed_at": null,
        "comments": 11,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25493",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing"
        ]
      },
      {
        "number": 25502,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: SYCL on Intel Arc 140V (Lunar Lake): fused kernels (FLASH_ATTN_EXT, GLU) fail with UR_RESULT_ERROR_OUT_OF_RESOURCES; >4GB weight buffer additionally requires UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-07-09T18:57:48Z",
        "updated_at": "2026-08-14T01:08:13Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25502",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25507,
        "kind": "pr",
        "state": "open",
        "title": "Placed MMA gateway over WMMA gateway",
        "labels": [
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-07-09T20:51:02Z",
        "updated_at": "2026-07-21T19:38:04Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25507",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25510,
        "kind": "issue",
        "state": "open",
        "title": "Malformed client-supplied tool call arguments return 500 instead of 4xx",
        "labels": [],
        "created_at": "2026-07-10T00:25:33Z",
        "updated_at": "2026-08-18T02:04:54Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25510",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing"
        ]
      },
      {
        "number": 25511,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: regex misses quoted tokens",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-10T01:19:00Z",
        "updated_at": "2026-07-27T02:11:24Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25511",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing"
        ]
      },
      {
        "number": 25518,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Garbage output for model Qwen2.5-0.5B-Instruct-GGUF when -ngl > 0",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-10T08:04:06Z",
        "updated_at": "2026-08-20T15:18:07Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25518",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25520,
        "kind": "issue",
        "state": "open",
        "title": "Vulkan: model load hangs indefinitely on RDNA4 (gfx1201) with plain Unsloth Q6_K quant, or OOMs host with --direct-io",
        "labels": [
          "stale"
        ],
        "created_at": "2026-07-10T08:28:50Z",
        "updated_at": "2026-08-19T02:07:21Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25520",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 25522,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Gemma 4 crashes with MTP",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-10T09:43:43Z",
        "updated_at": "2026-08-16T11:20:44Z",
        "closed_at": null,
        "comments": 12,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25522",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 25524,
        "kind": "pr",
        "state": "open",
        "title": "mtmd : read LFM2 tiling params from GGUF metadata",
        "labels": [
          "mtmd",
          "conversion"
        ],
        "created_at": "2026-07-10T12:20:44Z",
        "updated_at": "2026-08-17T13:02:08Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25524",
        "categories": [
          "quantization",
          "models_architectures",
          "multimodal_embedding",
          "loader_conversion",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 25533,
        "kind": "pr",
        "state": "open",
        "title": "metal : optimize Mamba-2 SSM scan decode",
        "labels": [
          "testing",
          "ggml",
          "Apple Metal"
        ],
        "created_at": "2026-07-10T17:02:28Z",
        "updated_at": "2026-07-10T17:13:14Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25533",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25542,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cpu: get_rows allows overwrite of RIP (PoC included)",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-07-10T21:37:28Z",
        "updated_at": "2026-07-11T19:18:19Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25542",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25549,
        "kind": "pr",
        "state": "open",
        "title": "spec: support DSpark for Gemma4",
        "labels": [
          "documentation",
          "model",
          "server",
          "conversion"
        ],
        "created_at": "2026-07-11T06:27:35Z",
        "updated_at": "2026-08-18T07:42:53Z",
        "closed_at": null,
        "comments": 20,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25549",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "speculative_generation",
          "loader_conversion",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25556,
        "kind": "pr",
        "state": "open",
        "title": "Metal: Optimize q8_0 KV",
        "labels": [
          "testing",
          "ggml",
          "Apple Metal"
        ],
        "created_at": "2026-07-11T09:14:19Z",
        "updated_at": "2026-08-22T14:53:38Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25556",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing"
        ]
      },
      {
        "number": 25557,
        "kind": "pr",
        "state": "open",
        "title": "quant : extend Q8_0 to use full byte range like Q4_0 (Python+CPU)",
        "labels": [
          "ggml",
          "conversion"
        ],
        "created_at": "2026-07-11T09:27:54Z",
        "updated_at": "2026-08-10T08:36:58Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25557",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25558,
        "kind": "pr",
        "state": "open",
        "title": "readme : add nika to infrastructure section",
        "labels": [
          "documentation"
        ],
        "created_at": "2026-07-11T10:52:56Z",
        "updated_at": "2026-07-11T12:18:04Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25558",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "api_ui_docs"
        ]
      },
      {
        "number": 25562,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: OpenVINO docker cannot use multiple GPUs",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-07-11T14:04:15Z",
        "updated_at": "2026-08-26T05:27:35Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25562",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 25569,
        "kind": "pr",
        "state": "open",
        "title": "ggml : fix SUM for non-contiguous tensors",
        "labels": [
          "testing",
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-07-11T18:40:05Z",
        "updated_at": "2026-08-21T04:23:21Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25569",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25570,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Add an option to terminate idle router workers after --sleep-idle-seconds",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-07-11T18:58:35Z",
        "updated_at": "2026-08-18T12:51:01Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25570",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25573,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: A sycl backended cli can't find some binaries from a latest intel-oneapi-runtime-libs",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-07-11T22:12:40Z",
        "updated_at": "2026-08-14T01:08:11Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25573",
        "categories": [
          "native_backends",
          "models_architectures",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25575,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cuda: use argsort for small multi-row top-k",
        "labels": [
          "testing",
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-07-12T01:26:41Z",
        "updated_at": "2026-08-03T12:40:24Z",
        "closed_at": null,
        "comments": 11,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25575",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25577,
        "kind": "issue",
        "state": "open",
        "title": "Feature request: generic layer-window hook - run layers [il_start, il_end) with an injectable/extractable boundary residual",
        "labels": [],
        "created_at": "2026-07-12T03:33:48Z",
        "updated_at": "2026-08-23T01:07:27Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25577",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25582,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: deepseek4 (DeepSeek-V4-Flash): garbled/degraded output when MoE expert layers run on CUDA — CPU-only experts (--n-cpu-moe) produce correct output",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-07-12T08:06:03Z",
        "updated_at": "2026-08-21T02:12:59Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25582",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25583,
        "kind": "pr",
        "state": "open",
        "title": "[server] Return 400 for malformed tool-call arguments instead of 500",
        "labels": [
          "server"
        ],
        "created_at": "2026-07-12T10:12:35Z",
        "updated_at": "2026-07-12T12:10:03Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25583",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing"
        ]
      },
      {
        "number": 25584,
        "kind": "pr",
        "state": "open",
        "title": "ggml : fix view init skipped after buft max_size split",
        "labels": [
          "testing",
          "ggml"
        ],
        "created_at": "2026-07-12T11:37:36Z",
        "updated_at": "2026-08-21T05:20:43Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25584",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25592,
        "kind": "pr",
        "state": "open",
        "title": "server: fix checkpoint handling for hybrid/recurrent models (#24055)",
        "labels": [
          "server"
        ],
        "created_at": "2026-07-12T16:21:40Z",
        "updated_at": "2026-08-26T05:05:53Z",
        "closed_at": null,
        "comments": 11,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25592",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 25593,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: SM_60 Quality Loss, FP32 math silently done in FP16, fix merged in two forks.",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-12T16:21:46Z",
        "updated_at": "2026-08-26T05:18:21Z",
        "closed_at": null,
        "comments": 18,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25593",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25594,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: split-mode tensor/row broken for mixed CUDA+ROCm setup",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-07-12T17:34:26Z",
        "updated_at": "2026-08-12T01:08:08Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25594",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 25602,
        "kind": "pr",
        "state": "open",
        "title": "grammar : reject inverted repetition range {m,n} (m>n)",
        "labels": [],
        "created_at": "2026-07-12T22:58:25Z",
        "updated_at": "2026-08-11T03:03:25Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25602",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing"
        ]
      },
      {
        "number": 25606,
        "kind": "pr",
        "state": "open",
        "title": "security: restrict server CORS to loopback origins",
        "labels": [
          "server"
        ],
        "created_at": "2026-07-13T01:42:06Z",
        "updated_at": "2026-07-13T01:46:34Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25606",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25610,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Fail to run with Ascend NPU 910B3.",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-07-13T03:40:13Z",
        "updated_at": "2026-08-13T01:08:08Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25610",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25612,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: split-mode causes garbled output on dual Intel dGPU setup.",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-13T04:53:49Z",
        "updated_at": "2026-08-01T08:25:42Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25612",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 25618,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Speculative decoding (draft-mtp / draft-dspark): greedy output diverges from vanilla on quantized targets",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-13T11:55:59Z",
        "updated_at": "2026-08-26T11:57:33Z",
        "closed_at": null,
        "comments": 19,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25618",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25620,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: MMVQ host/device nwarps mismatch causes NaN on gfx11-generic / RDNA3.5",
        "labels": [
          "stale"
        ],
        "created_at": "2026-07-13T12:44:31Z",
        "updated_at": "2026-08-23T01:07:25Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25620",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25635,
        "kind": "pr",
        "state": "open",
        "title": "CUDA: XOR swizzle flash attn  K,V smem fp16 tiles",
        "labels": [
          "testing",
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-07-13T19:23:11Z",
        "updated_at": "2026-08-25T07:11:20Z",
        "closed_at": null,
        "comments": 23,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25635",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25636,
        "kind": "pr",
        "state": "open",
        "title": "llama : fix null pointer dereference by checking for rotation buffer in K-cache shifts",
        "labels": [],
        "created_at": "2026-07-13T19:44:07Z",
        "updated_at": "2026-07-13T20:06:26Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25636",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25644,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: -ot and -ts/dev_layer() placement disagreement crashes RPC decode",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-07-14T01:44:58Z",
        "updated_at": "2026-08-14T01:08:09Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25644",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing"
        ]
      },
      {
        "number": 25645,
        "kind": "pr",
        "state": "open",
        "title": "metal : add F16 support for ADD/SUB/MUL/DIV",
        "labels": [
          "ggml",
          "Apple Metal"
        ],
        "created_at": "2026-07-14T01:53:41Z",
        "updated_at": "2026-07-14T05:39:32Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25645",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25646,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Model weight gets evicted from idle Intel dGPU memory causing perf drop on Vulkan",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-14T03:39:12Z",
        "updated_at": "2026-08-01T16:17:22Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25646",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25648,
        "kind": "pr",
        "state": "open",
        "title": "metal : fix null-pipeline crash for F16 src1 mul_mat/mul_mat_id",
        "labels": [
          "ggml",
          "Apple Metal"
        ],
        "created_at": "2026-07-14T05:33:05Z",
        "updated_at": "2026-07-14T05:39:46Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25648",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25652,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: LightOnOCR-1B produces degenerate output (\"@@@@...\") via mtmd — regression since #16764",
        "labels": [
          "stale"
        ],
        "created_at": "2026-07-14T09:50:00Z",
        "updated_at": "2026-08-19T02:07:16Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25652",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25664,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: vk::DeviceLostError on linux-7.x kernels (RADV_STRIXHALO)",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-14T14:16:56Z",
        "updated_at": "2026-08-17T18:16:25Z",
        "closed_at": null,
        "comments": 23,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25664",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 25665,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: SearXNG and Toggle ON/OFF tab goes off.",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-07-14T14:25:29Z",
        "updated_at": "2026-08-14T01:08:08Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25665",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25666,
        "kind": "pr",
        "state": "open",
        "title": "vulkan: do not enable MMVQ for speculative-decode steps on AMD",
        "labels": [
          "Vulkan",
          "ggml"
        ],
        "created_at": "2026-07-14T14:25:42Z",
        "updated_at": "2026-08-26T20:11:17Z",
        "closed_at": null,
        "comments": 17,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25666",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25668,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Can't run gemma4-style model",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-14T15:31:47Z",
        "updated_at": "2026-08-01T19:00:29Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25668",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing"
        ]
      },
      {
        "number": 25669,
        "kind": "pr",
        "state": "open",
        "title": "CUDA: ssm-scan use element strides instead of bytes",
        "labels": [
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-07-14T16:23:59Z",
        "updated_at": "2026-08-21T01:44:06Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25669",
        "categories": [
          "native_backends",
          "models_architectures",
          "multimodal_embedding",
          "correctness_testing"
        ]
      },
      {
        "number": 25670,
        "kind": "pr",
        "state": "open",
        "title": "rpc : reject invalid top-level graph nodes in graph_compute",
        "labels": [
          "examples",
          "ggml"
        ],
        "created_at": "2026-07-14T16:31:41Z",
        "updated_at": "2026-07-17T07:04:30Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25670",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25681,
        "kind": "pr",
        "state": "open",
        "title": "Add PyTorch dequantization API to gguf-py",
        "labels": [
          "conversion"
        ],
        "created_at": "2026-07-14T19:20:28Z",
        "updated_at": "2026-07-17T05:00:15Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25681",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25688,
        "kind": "pr",
        "state": "open",
        "title": "ggml-backend-meta: fix device and buffer type names",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-07-14T21:38:54Z",
        "updated_at": "2026-07-15T09:13:03Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25688",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25689,
        "kind": "pr",
        "state": "open",
        "title": "ggml-backend-meta: do not scale view offsets internal to a split unit",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-07-14T21:40:16Z",
        "updated_at": "2026-07-15T09:13:03Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25689",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 25692,
        "kind": "issue",
        "state": "open",
        "title": "SYCL: GPU hang (xe ccs engine reset) with flash attention + quantized KV cache under sustained server load on Arc Pro B70",
        "labels": [],
        "created_at": "2026-07-15T01:43:10Z",
        "updated_at": "2026-08-24T14:49:54Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25692",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 25693,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: running Qwen3.5-35B-A3B on Ascend910B4 with Multi-batch，get error inference result",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-07-15T01:55:26Z",
        "updated_at": "2026-08-15T01:22:40Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25693",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 25695,
        "kind": "pr",
        "state": "open",
        "title": "common : recognize 'thinking' as reasoning alias in chat diff analyzer",
        "labels": [
          "testing"
        ],
        "created_at": "2026-07-15T03:08:43Z",
        "updated_at": "2026-07-16T09:37:01Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25695",
        "categories": [
          "models_architectures",
          "structured_generation_tools",
          "correctness_testing"
        ]
      },
      {
        "number": 25700,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Decreased perfomance because of input layers on CPU for AMD Strix Halo",
        "labels": [
          "bug",
          "performance",
          "RoCM"
        ],
        "created_at": "2026-07-15T05:27:33Z",
        "updated_at": "2026-08-10T17:57:52Z",
        "closed_at": null,
        "comments": 30,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25700",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25703,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Bug 1: RPC Graph Recompute Use-After-Free After Backing Buffer Free",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-07-15T07:01:47Z",
        "updated_at": "2026-08-15T01:22:39Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25703",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25704,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Bug 2: RPC Gated Delta Net K op_param Out-of-Bounds State Snapshot Write",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-07-15T07:05:37Z",
        "updated_at": "2026-08-15T01:22:38Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25704",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25709,
        "kind": "pr",
        "state": "open",
        "title": "Thinking sampling parametrization",
        "labels": [
          "documentation",
          "testing",
          "server"
        ],
        "created_at": "2026-07-15T09:30:41Z",
        "updated_at": "2026-08-19T21:02:17Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25709",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25711,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: prompt processing stuck on Intel (SYCL) with split mode tensor",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-07-15T11:15:24Z",
        "updated_at": "2026-08-15T01:22:37Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25711",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 25712,
        "kind": "pr",
        "state": "open",
        "title": "server: add support for `stream: true` and OpenAI fallback",
        "labels": [
          "server"
        ],
        "created_at": "2026-07-15T11:40:34Z",
        "updated_at": "2026-08-12T03:45:47Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25712",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25713,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: MTP decoding crash on pre-Ampere GPUs (with working patch!)",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-15T11:47:35Z",
        "updated_at": "2026-08-19T00:05:59Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25713",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25717,
        "kind": "issue",
        "state": "open",
        "title": "server: CUDA illegal memory access on vision requests with Qwen3.5-35B-A3B + mmproj (M-RoPE) — crash matrix: flash-attn / kv_unified / text-length-before-image",
        "labels": [
          "stale"
        ],
        "created_at": "2026-07-15T12:41:46Z",
        "updated_at": "2026-08-15T01:22:35Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25717",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25721,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: \"ggml-cuda.cu:104: CUDA error\" with Qwen3.6-35B-A3B-MXFP4_MOE.gguf and reasoning on",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-07-15T13:35:19Z",
        "updated_at": "2026-08-15T01:22:34Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25721",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 25722,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: mistral4 empty output on Metal for prompts over ~300 tokens (clean GGUF, with and without flash attention)",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-15T14:49:15Z",
        "updated_at": "2026-08-22T07:32:20Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25722",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25724,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Webui causing high CPU load in Firefox",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-07-15T15:27:44Z",
        "updated_at": "2026-08-26T01:07:30Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25724",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25726,
        "kind": "pr",
        "state": "open",
        "title": "spec: add initial adaptive draft length heuristic (curently mtp only)",
        "labels": [],
        "created_at": "2026-07-15T15:42:55Z",
        "updated_at": "2026-08-17T22:08:57Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25726",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25727,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Ternary Bonsai 27B don't run",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-15T16:26:14Z",
        "updated_at": "2026-08-01T07:54:21Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25727",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25731,
        "kind": "pr",
        "state": "open",
        "title": "Add TML Inkling architecture",
        "labels": [
          "model",
          "testing",
          "ggml",
          "mtmd",
          "CUDA",
          "conversion"
        ],
        "created_at": "2026-07-15T18:23:09Z",
        "updated_at": "2026-08-26T10:38:00Z",
        "closed_at": null,
        "comments": 32,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25731",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25732,
        "kind": "issue",
        "state": "open",
        "title": "docs: double ctx-size with \"split-mode: tensor\" clarification.",
        "labels": [
          "stale"
        ],
        "created_at": "2026-07-15T18:35:15Z",
        "updated_at": "2026-08-15T01:22:32Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25732",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25739,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: official Gemma 4 QAT GGUF (google/gemma-4-E2B-it-qat-q4_0-gguf) aborts at vocab load: GGML_ASSERT(id_to_token.size() == token_to_id.size())",
        "labels": [
          "stale"
        ],
        "created_at": "2026-07-15T21:00:28Z",
        "updated_at": "2026-08-19T02:07:12Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25739",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 25742,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: models attempt/fail unwanted tool calls to huggingface",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-07-16T00:00:49Z",
        "updated_at": "2026-08-20T02:06:09Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25742",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25744,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: DeepSeek-v4 Flash takes 200 seconds for prefill of 10 tokens.",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-07-16T00:17:04Z",
        "updated_at": "2026-08-20T02:06:08Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25744",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25746,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: json-schema-to-grammar — nested string maxLength >= 2000 emits un-parseable GBNF (tool-call grammar, b10034)",
        "labels": [],
        "created_at": "2026-07-16T00:30:06Z",
        "updated_at": "2026-08-02T20:03:16Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25746",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25751,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: SWA on Gemma 4 forgets key details",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-16T03:43:29Z",
        "updated_at": "2026-08-13T04:33:38Z",
        "closed_at": null,
        "comments": 14,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25751",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25755,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: GGML_ASSERT \"tensor write out of bounds\" in server_prompt_cache::load with Qwen3.6-27B (Vulkan)",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-07-16T05:55:35Z",
        "updated_at": "2026-08-19T02:07:11Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25755",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25764,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Bug 3: RPC ACC/SET Negative Stride op_param Bounds-Check Bypass OOB Write",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-07-16T07:41:13Z",
        "updated_at": "2026-08-16T01:07:13Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25764",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25765,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Bug 4: RPC SSM_SCAN Unchecked ids Tensor State Offset OOB Read/Write",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-07-16T07:42:03Z",
        "updated_at": "2026-08-16T01:07:12Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25765",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25767,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Flaky Vulkan unit test results on some Intel Battlemage Linux systems",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-16T08:44:05Z",
        "updated_at": "2026-08-17T08:13:41Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25767",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25773,
        "kind": "pr",
        "state": "open",
        "title": "vulkan: use spec constant for matrix matrix multiplication A-type",
        "labels": [
          "Vulkan",
          "ggml"
        ],
        "created_at": "2026-07-16T12:21:23Z",
        "updated_at": "2026-08-26T05:20:19Z",
        "closed_at": null,
        "comments": 21,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25773",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing"
        ]
      },
      {
        "number": 25774,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Supports MonkeyOCRv2",
        "labels": [
          "enhancement",
          "stale"
        ],
        "created_at": "2026-07-16T13:05:26Z",
        "updated_at": "2026-08-16T01:07:12Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25774",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 25777,
        "kind": "issue",
        "state": "open",
        "title": "GGML_ASSERT(id >= 0 && id < n_expert) with Gemma 4 26B-A4B + --n-cpu-moe on Vulkan",
        "labels": [],
        "created_at": "2026-07-16T14:43:00Z",
        "updated_at": "2026-08-11T15:03:45Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25777",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25788,
        "kind": "pr",
        "state": "open",
        "title": "metal : gated_delta_net cache fusion",
        "labels": [
          "testing",
          "ggml",
          "Apple Metal"
        ],
        "created_at": "2026-07-16T17:40:39Z",
        "updated_at": "2026-08-25T09:19:56Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25788",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing"
        ]
      },
      {
        "number": 25789,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: webui does not display image attachments returned by MCP tools (tool block regression in #25450, inline references never resolved)",
        "labels": [
          "stale"
        ],
        "created_at": "2026-07-16T19:51:41Z",
        "updated_at": "2026-08-26T01:07:26Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25789",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25790,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Instruction injection through the Query parameter (?q=) of the WebUI (security)",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-07-16T20:51:42Z",
        "updated_at": "2026-08-26T01:07:25Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25790",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25791,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: DGX Spark GB10 cluster crashing in RPC",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-07-16T20:59:06Z",
        "updated_at": "2026-08-17T01:07:00Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25791",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25796,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: deepseek v4 flash errors out on tool calls with similar parameter names",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-07-17T00:17:14Z",
        "updated_at": "2026-08-16T01:07:11Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25796",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 25798,
        "kind": "pr",
        "state": "open",
        "title": "docs: add embeddings tutorial",
        "labels": [
          "documentation"
        ],
        "created_at": "2026-07-17T01:45:22Z",
        "updated_at": "2026-07-17T02:05:40Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25798",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25799,
        "kind": "pr",
        "state": "open",
        "title": "docs: add gpt-oss guide",
        "labels": [
          "documentation"
        ],
        "created_at": "2026-07-17T01:58:08Z",
        "updated_at": "2026-07-17T02:05:29Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25799",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25804,
        "kind": "issue",
        "state": "open",
        "title": "Performance issue with multi‑batch decoding",
        "labels": [
          "stale"
        ],
        "created_at": "2026-07-17T04:02:29Z",
        "updated_at": "2026-08-25T01:07:11Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25804",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 25807,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: ROCm-7.14 - > 'error while loading shared libraries: libhipblas.so.3'",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-17T04:48:12Z",
        "updated_at": "2026-08-12T15:32:13Z",
        "closed_at": null,
        "comments": 16,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25807",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25808,
        "kind": "issue",
        "state": "open",
        "title": "Compile bug: GGML_SYCL_DEVICE_ARCH=xe2 gives segfault",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-07-17T05:52:54Z",
        "updated_at": "2026-08-24T01:07:12Z",
        "closed_at": null,
        "comments": 29,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25808",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25809,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: OpenVINO NPU: first model load fails with RoPE shape mismatch, second attempt succeeds",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-17T05:55:52Z",
        "updated_at": "2026-07-29T04:55:27Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25809",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 25812,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: SYCL: UR_RESULT_ERROR_OUT_OF_HOST_MEMORY when offloading GLM-5.2 (glm_moe_dsa) MoE experts to Intel Arc GPUs",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-17T07:42:36Z",
        "updated_at": "2026-08-23T01:07:18Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25812",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25814,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: mtmd llava graph feeds projected [CLS] to the LLM and shifts patch position embeddings",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-07-17T09:50:24Z",
        "updated_at": "2026-08-17T01:06:58Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25814",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25817,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: mtmd llava models run one ViT layer short (subtracted at conversion AND graph build)",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-17T10:03:04Z",
        "updated_at": "2026-08-03T07:21:50Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25817",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 25818,
        "kind": "pr",
        "state": "open",
        "title": "server : remote speculative decoding via ethernet",
        "labels": [
          "server"
        ],
        "created_at": "2026-07-17T10:41:00Z",
        "updated_at": "2026-07-17T10:41:23Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25818",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25819,
        "kind": "pr",
        "state": "open",
        "title": "server : add stuck-loop escape for ngram-mod (WIP)",
        "labels": [
          "server"
        ],
        "created_at": "2026-07-17T11:22:58Z",
        "updated_at": "2026-08-25T21:13:59Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25819",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "correctness_testing"
        ]
      },
      {
        "number": 25820,
        "kind": "pr",
        "state": "open",
        "title": "contrib: add contributing guidelines for new backends",
        "labels": [
          "documentation"
        ],
        "created_at": "2026-07-17T11:35:28Z",
        "updated_at": "2026-07-25T08:26:34Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25820",
        "categories": [
          "native_backends",
          "models_architectures",
          "api_ui_docs"
        ]
      },
      {
        "number": 25833,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: b10063cannot run model https://huggingface.co/prism-ml/Ternary-Bonsai-27B-gguf",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-07-17T16:15:11Z",
        "updated_at": "2026-08-18T02:04:40Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25833",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25834,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cuda: add optional DP2A fallback for DP4A",
        "labels": [
          "documentation",
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-07-17T16:19:46Z",
        "updated_at": "2026-07-17T16:23:58Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25834",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25835,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: VRAM leak with CUDA Graphs(Enable CUDA graphs on Volta+Turing #25749) on V100",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-17T16:28:23Z",
        "updated_at": "2026-08-23T01:07:16Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25835",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25836,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: BF16/F16 models crash with \"ROCm error: invalid argument\" (MUL_MAT) on gfx1201/Windows - second rocBLAS GEMM with different lda fails",
        "labels": [
          "stale"
        ],
        "created_at": "2026-07-17T16:45:24Z",
        "updated_at": "2026-08-17T01:06:57Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25836",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25840,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: server -np 3 decode drops ~4.4x (166 -> 38 t/s) whenever the two busy slots are the non-contiguous pair {0,2} — hybrid/recurrent model (Qwen3.6-35B-A3B), deterministic, b9966",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-07-17T18:17:34Z",
        "updated_at": "2026-08-19T02:07:04Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25840",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25841,
        "kind": "pr",
        "state": "open",
        "title": "docs: warn when --ubatch-size is too small in --embeddings mode",
        "labels": [
          "documentation",
          "server"
        ],
        "created_at": "2026-07-17T18:29:33Z",
        "updated_at": "2026-07-17T18:55:22Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25841",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25844,
        "kind": "pr",
        "state": "open",
        "title": "mtmd : fix llava CLS token ordering (#25814)",
        "labels": [
          "mtmd"
        ],
        "created_at": "2026-07-17T19:46:14Z",
        "updated_at": "2026-07-20T07:56:48Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25844",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25855,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Support internlm/Intern-S2-Preview-397B",
        "labels": [
          "enhancement",
          "stale"
        ],
        "created_at": "2026-07-18T09:00:04Z",
        "updated_at": "2026-08-20T02:06:02Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25855",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25857,
        "kind": "issue",
        "state": "open",
        "title": "qwen35moe Bug Report for llama.cpp",
        "labels": [
          "stale"
        ],
        "created_at": "2026-07-18T09:38:10Z",
        "updated_at": "2026-08-18T02:04:39Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25857",
        "categories": [
          "models_architectures",
          "correctness_testing"
        ]
      },
      {
        "number": 25858,
        "kind": "issue",
        "state": "open",
        "title": "test-quantize-fns: scratch buffers assume <=2 bytes/element, overflow for a wider vec_dot_type",
        "labels": [
          "stale"
        ],
        "created_at": "2026-07-18T10:05:22Z",
        "updated_at": "2026-08-18T02:04:37Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25858",
        "categories": [
          "quantization",
          "memory_kv_context",
          "correctness_testing"
        ]
      },
      {
        "number": 25859,
        "kind": "issue",
        "state": "open",
        "title": "Offloaded-MoE prefill leaves the GPU idle waiting on serial expert H2D copies (single-GPU -ncmoe)",
        "labels": [],
        "created_at": "2026-07-18T10:05:36Z",
        "updated_at": "2026-08-14T03:39:47Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25859",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "build_platform_ops"
        ]
      },
      {
        "number": 25860,
        "kind": "pr",
        "state": "open",
        "title": "Deepseek V4: split-mode tensor",
        "labels": [
          "testing",
          "ggml"
        ],
        "created_at": "2026-07-18T10:39:32Z",
        "updated_at": "2026-07-30T17:49:54Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25860",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 25862,
        "kind": "pr",
        "state": "open",
        "title": "vulkan: use mul_mat_vec_id for high-expert-count MoE decode",
        "labels": [
          "Vulkan",
          "ggml"
        ],
        "created_at": "2026-07-18T13:27:37Z",
        "updated_at": "2026-07-19T09:27:39Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25862",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25863,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cuda: avoid direct ROCm_Host compute on HIP integrated GPUs",
        "labels": [
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-07-18T15:22:33Z",
        "updated_at": "2026-08-24T15:46:56Z",
        "closed_at": null,
        "comments": 35,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25863",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25864,
        "kind": "pr",
        "state": "open",
        "title": "server : fix dry_sequence_breakers missing from GET /props generation settings",
        "labels": [
          "server"
        ],
        "created_at": "2026-07-18T16:03:48Z",
        "updated_at": "2026-07-18T16:07:42Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25864",
        "categories": [
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing"
        ]
      },
      {
        "number": 25866,
        "kind": "issue",
        "state": "open",
        "title": "Metal: MTLDispatchTypeConcurrent dependency-tracking race on GCN/Vega (wave64) discrete GPUs — corrupted output and hard command-buffer failures",
        "labels": [
          "stale"
        ],
        "created_at": "2026-07-18T16:13:58Z",
        "updated_at": "2026-08-18T02:04:36Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25866",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25870,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: OpenCL flash-attn: supports_op returns true for kernels skipped at registration under device workgroup limits -> std::out_of_range (map::at) on Adreno 740",
        "labels": [
          "bug-unconfirmed",
          "stale",
          "OpenCL"
        ],
        "created_at": "2026-07-18T18:54:32Z",
        "updated_at": "2026-08-19T02:07:02Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25870",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25876,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: ggml-hexagon HMX produces garbled output on Snapdragon 8 Elite (SM8850, v81)",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-18T21:21:17Z",
        "updated_at": "2026-07-30T23:22:58Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25876",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25882,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: convert_hf_to_gguf.py fails/crashes on Unsloth-derived Llama 3.2 3B checkpoints: three distinct tokenizer-handling issues",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-07-19T03:36:52Z",
        "updated_at": "2026-08-19T02:07:01Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25882",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 25884,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: On windows/vulkan, split-mode row no longer works on hybrid AMD/Intel GPU config since commit 74976e1",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-07-19T07:47:19Z",
        "updated_at": "2026-08-24T01:07:07Z",
        "closed_at": null,
        "comments": 11,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25884",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25887,
        "kind": "issue",
        "state": "open",
        "title": "Qwen 3.6 27B GHCP VS Code agent interruptions",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-19T09:17:25Z",
        "updated_at": "2026-07-28T16:16:58Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25887",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25889,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: server ignores the real audio clip and makes things up when reasoning mode is turned on (Gemma 4 E2B)",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-07-19T10:29:03Z",
        "updated_at": "2026-08-19T02:07:00Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25889",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 25890,
        "kind": "issue",
        "state": "open",
        "title": "rpc: model load serializes read + hash + dispatch on one host core; 535GB load takes ~15 min while NIC and 95 cores idle",
        "labels": [],
        "created_at": "2026-07-19T10:59:22Z",
        "updated_at": "2026-07-30T17:15:21Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25890",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25892,
        "kind": "pr",
        "state": "open",
        "title": "I have used int32_t for llama_memory_seq_div",
        "labels": [],
        "created_at": "2026-07-19T12:24:29Z",
        "updated_at": "2026-07-19T12:24:29Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25892",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "api_ui_docs"
        ]
      },
      {
        "number": 25894,
        "kind": "issue",
        "state": "open",
        "title": "Research: Fake GQA on full sized kv",
        "labels": [
          "research 🔬"
        ],
        "created_at": "2026-07-19T12:45:55Z",
        "updated_at": "2026-07-19T12:45:55Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25894",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25904,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: OpenVINO NPU: SIGSEGV/WeightlessCacheAttribute on Intel Core Ultra 7 255H (Linux)",
        "labels": [
          "stale"
        ],
        "created_at": "2026-07-19T20:46:13Z",
        "updated_at": "2026-08-21T02:12:46Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25904",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25906,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Vulkan / Adreno 702 / Turnip",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-07-19T21:59:13Z",
        "updated_at": "2026-08-21T02:12:45Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25906",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25908,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: speculative decoding (draft-simple): p_min default 0.00 collapses acceptance to 0.07; even at 0.9 acceptance spec is slower than plain decode (Vulkan/RDNA4)",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-19T23:43:21Z",
        "updated_at": "2026-07-31T14:50:07Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25908",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25913,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: /slots save/restore silently loses all prompt reuse on hybrid/recurrent models — checkpoints are never persisted",
        "labels": [],
        "created_at": "2026-07-20T08:19:48Z",
        "updated_at": "2026-08-04T18:13:42Z",
        "closed_at": null,
        "comments": 10,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25913",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25917,
        "kind": "pr",
        "state": "open",
        "title": "cuda : support for sparse KV indices in MMA flash attention kernel",
        "labels": [
          "model",
          "testing",
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-07-20T10:09:17Z",
        "updated_at": "2026-08-05T14:45:09Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25917",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing"
        ]
      },
      {
        "number": 25919,
        "kind": "pr",
        "state": "open",
        "title": "sampling: allow backend sampling to work with reasoning budget",
        "labels": [],
        "created_at": "2026-07-20T10:11:35Z",
        "updated_at": "2026-07-20T21:26:15Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25919",
        "categories": [
          "native_backends",
          "models_architectures",
          "speculative_generation",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25921,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Support multimodal documents in /v1/rerank endpoint for VL reranker models",
        "labels": [
          "enhancement",
          "stale"
        ],
        "created_at": "2026-07-20T11:08:28Z",
        "updated_at": "2026-08-20T02:05:58Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25921",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25923,
        "kind": "issue",
        "state": "open",
        "title": "json-schema-to-grammar: empty-object schema and large maxLength both produce invalid/rejected GBNF, breaking the whole tool-call grammar",
        "labels": [],
        "created_at": "2026-07-20T12:48:44Z",
        "updated_at": "2026-07-27T04:36:40Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25923",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25938,
        "kind": "issue",
        "state": "open",
        "title": "RPC client: RPC_STATUS_ASSERT in ggml_backend_rpc_buffer_get_tensor() aborts entire process on any RPC failure, instead of returning an error",
        "labels": [],
        "created_at": "2026-07-20T21:27:00Z",
        "updated_at": "2026-08-25T01:07:05Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25938",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25939,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Vulkan: GPU memory never returns to idle after long generations (VRAM ~90C, fan pinned); short generations release normally (Radeon AI PRO R9700 / RADV GFX1201)",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-07-20T21:40:37Z",
        "updated_at": "2026-08-21T02:12:44Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25939",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25940,
        "kind": "pr",
        "state": "open",
        "title": "HIP: RDNA 4 MUL_MAT optimizations",
        "labels": [
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-07-20T21:47:44Z",
        "updated_at": "2026-08-25T05:52:29Z",
        "closed_at": null,
        "comments": 19,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25940",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25944,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: When using AI programs with llama.ccp as backend there is often no rocognition of some ROCm hardware when selecting hip forcing vulkan use",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-20T23:08:16Z",
        "updated_at": "2026-08-05T14:30:06Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25944",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25952,
        "kind": "pr",
        "state": "open",
        "title": "cuda: fuse MoE weighted expert reduction",
        "labels": [
          "testing",
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-07-21T08:08:59Z",
        "updated_at": "2026-08-18T17:36:48Z",
        "closed_at": null,
        "comments": 35,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25952",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 25953,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Very high memory leackage suspicion",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-21T10:53:17Z",
        "updated_at": "2026-08-20T08:14:09Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25953",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25958,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Bug 13: RPC create_node Linear Dependency Chain Stack Overflow",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-07-21T14:54:04Z",
        "updated_at": "2026-08-21T02:12:42Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25958",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25959,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Bug 14: Server chat_parser Cyclic PEG Sequence Stack Overflow",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-07-21T14:54:55Z",
        "updated_at": "2026-08-21T02:12:42Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25959",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25960,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Bug 17: Server %llguidance Grammar Disabled-Build GGML_ABORT DoS",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-07-21T14:57:00Z",
        "updated_at": "2026-08-21T02:12:41Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25960",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25961,
        "kind": "pr",
        "state": "open",
        "title": "reasoning-budget: Implementing thinking-budget mechanism to control thought process",
        "labels": [
          "documentation",
          "testing",
          "server"
        ],
        "created_at": "2026-07-21T14:59:22Z",
        "updated_at": "2026-08-09T08:48:57Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25961",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25964,
        "kind": "pr",
        "state": "open",
        "title": "embeddings: skip logits and lm head during inference",
        "labels": [
          "documentation",
          "model",
          "testing",
          "server"
        ],
        "created_at": "2026-07-21T16:10:52Z",
        "updated_at": "2026-07-21T16:38:10Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25964",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 25967,
        "kind": "issue",
        "state": "open",
        "title": "Duplicate rule definitions in generated GBNF grammar with large tools list (harmony/gpt-oss) → \"failed to parse grammar\"",
        "labels": [
          "stale"
        ],
        "created_at": "2026-07-21T16:53:32Z",
        "updated_at": "2026-08-21T02:12:40Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25967",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 25971,
        "kind": "pr",
        "state": "open",
        "title": "Fix issue #5365: Support for Sparse MoE models like Camelidae and Sparsetral",
        "labels": [
          "model",
          "conversion"
        ],
        "created_at": "2026-07-21T18:52:08Z",
        "updated_at": "2026-07-21T19:05:29Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25971",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 25972,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: OpenVINO backend crashes with MTP speculative decoding (--spec-type draft-mtp) - no working draft configuration",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-21T21:01:02Z",
        "updated_at": "2026-08-26T05:32:08Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25972",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25973,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: SYCL: bad performance on newer oneAPI",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-07-21T21:06:52Z",
        "updated_at": "2026-08-26T01:07:14Z",
        "closed_at": null,
        "comments": 9,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25973",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25976,
        "kind": "issue",
        "state": "open",
        "title": "Arm CPU backend: quantized decode is compute-bound at ~55-60% of memory bandwidth (Neoverse-N2, SVE2+i8mm); perf profile included",
        "labels": [
          "stale"
        ],
        "created_at": "2026-07-21T23:04:31Z",
        "updated_at": "2026-08-21T02:12:39Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25976",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 25977,
        "kind": "issue",
        "state": "open",
        "title": "Arm CPU backend: prefill GEMM flat at ~18 t/s pp512 regardless of thread count or batch size (Neoverse-N2, i8mm+bf16 present)",
        "labels": [
          "stale"
        ],
        "created_at": "2026-07-21T23:48:03Z",
        "updated_at": "2026-08-21T02:12:38Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25977",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 25985,
        "kind": "issue",
        "state": "open",
        "title": "Compile bug: Build failure: CoopMat2 shaders fail with glslc -O / SPIR-V capability 5432",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-07-22T05:30:02Z",
        "updated_at": "2026-08-24T01:07:03Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25985",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 25986,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: gemma4 peg template intermittently unparseable with long multi-line tool-call string arguments",
        "labels": [
          "stale"
        ],
        "created_at": "2026-07-22T06:16:05Z",
        "updated_at": "2026-08-22T01:13:03Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25986",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 25988,
        "kind": "pr",
        "state": "open",
        "title": "cmake(ppc64le): Add clang toolchain",
        "labels": [
          "documentation",
          "build"
        ],
        "created_at": "2026-07-22T08:40:37Z",
        "updated_at": "2026-07-24T07:22:52Z",
        "closed_at": null,
        "comments": 9,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25988",
        "categories": [
          "native_backends",
          "models_architectures",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25990,
        "kind": "pr",
        "state": "open",
        "title": "16393 webui models management",
        "labels": [
          "server/ui"
        ],
        "created_at": "2026-07-22T09:19:46Z",
        "updated_at": "2026-07-29T16:03:55Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25990",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 25991,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cpu : add LoongArch64 LASX Q4_0 repack path",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-07-22T10:03:59Z",
        "updated_at": "2026-07-22T10:05:53Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/25991",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 25992,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: server -np 4 --kv-unified returns other requests' responses verbatim on integrated HIP GPU (gfx1151) — bisected to c7d87229",
        "labels": [],
        "created_at": "2026-07-22T10:15:42Z",
        "updated_at": "2026-08-25T20:26:50Z",
        "closed_at": null,
        "comments": 10,
        "url": "https://github.com/ggml-org/llama.cpp/issues/25992",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26001,
        "kind": "pr",
        "state": "open",
        "title": "CUDA: Support of GDN chunked kernel for prefill",
        "labels": [
          "testing",
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-07-22T14:12:07Z",
        "updated_at": "2026-08-18T00:16:33Z",
        "closed_at": null,
        "comments": 14,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26001",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26003,
        "kind": "pr",
        "state": "open",
        "title": "llama : add --lazy-experts for MoE models larger than RAM",
        "labels": [],
        "created_at": "2026-07-22T14:29:33Z",
        "updated_at": "2026-08-09T00:31:13Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26003",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "build_platform_ops"
        ]
      },
      {
        "number": 26004,
        "kind": "pr",
        "state": "open",
        "title": "server : preserve context checkpoints across slot save/restore",
        "labels": [
          "server"
        ],
        "created_at": "2026-07-22T15:08:16Z",
        "updated_at": "2026-08-16T15:37:53Z",
        "closed_at": null,
        "comments": 13,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26004",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26008,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: server: add option to concatenate prompts on /completion endpoint",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-07-22T16:32:52Z",
        "updated_at": "2026-08-10T11:01:57Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26008",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 26010,
        "kind": "issue",
        "state": "open",
        "title": "[SYCL] Low generation throughput on Intel Arc compared to Vulkan backend",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-07-22T17:27:45Z",
        "updated_at": "2026-08-05T05:19:22Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26010",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26013,
        "kind": "pr",
        "state": "open",
        "title": "OAI Responses API json schema support, Cohere2 MoE template parser json schema support, improvements to responses streaming compatibility ",
        "labels": [
          "testing",
          "server"
        ],
        "created_at": "2026-07-22T19:45:01Z",
        "updated_at": "2026-08-20T01:33:00Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26013",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 26014,
        "kind": "pr",
        "state": "open",
        "title": "Windows unbuffered model load",
        "labels": [
          "help wanted",
          "windows"
        ],
        "created_at": "2026-07-22T19:56:08Z",
        "updated_at": "2026-08-07T11:32:04Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26014",
        "categories": [
          "native_backends",
          "models_architectures",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26017,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: When I run Gemma 4e4b with MTP, it doesn't work and crashes to cmd",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-07-22T21:14:09Z",
        "updated_at": "2026-08-23T01:07:09Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26017",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "speculative_generation",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26019,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Add ubuntu CUDA 13 release binaries (x64/arm64)",
        "labels": [
          "enhancement",
          "stale"
        ],
        "created_at": "2026-07-22T22:37:45Z",
        "updated_at": "2026-08-22T01:13:01Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26019",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26020,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Next generation samplers that should be able to beat top-n sigma and other SOTA samplers",
        "labels": [
          "enhancement",
          "stale"
        ],
        "created_at": "2026-07-22T22:40:54Z",
        "updated_at": "2026-08-22T01:13:00Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26020",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 26027,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: GLM-5.2 (glm_moe_dsa) dense-MLA CUDA path produces subtly corrupted output for ANY real transformer layer offloaded to GPU (partial coherent text mixed with garbage)",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-23T05:49:43Z",
        "updated_at": "2026-08-12T20:07:16Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26027",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 26031,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Qwen3.6-35B-A3B-Q8_0.gguf multiple clients concurrently produce garbled output b9922 above（b9918 is ok）",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-23T07:30:34Z",
        "updated_at": "2026-08-20T06:34:41Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26031",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26033,
        "kind": "pr",
        "state": "open",
        "title": "metal: implement soft max backward operation in Metal backend",
        "labels": [
          "ggml",
          "Apple Metal"
        ],
        "created_at": "2026-07-23T08:45:37Z",
        "updated_at": "2026-07-23T09:16:43Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26033",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26035,
        "kind": "pr",
        "state": "open",
        "title": "Feature: Add p-less sampling",
        "labels": [
          "testing"
        ],
        "created_at": "2026-07-23T10:49:17Z",
        "updated_at": "2026-07-24T12:31:28Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26035",
        "categories": [
          "native_backends",
          "models_architectures",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26038,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug:  Excessive compute buffer reservation in MTP draft context on ROCm HIP unnecessarily reduces fitted context size",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-23T12:14:00Z",
        "updated_at": "2026-08-25T01:57:29Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26038",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing"
        ]
      },
      {
        "number": 26043,
        "kind": "issue",
        "state": "open",
        "title": "Server always reports `chat format: peg-native` at runtime",
        "labels": [
          "stale"
        ],
        "created_at": "2026-07-23T14:06:17Z",
        "updated_at": "2026-08-23T01:07:08Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26043",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26044,
        "kind": "issue",
        "state": "open",
        "title": "CUDA / Volta (sm_70): certain inputs make Qwen3-Embedding-8B return an all-NaN embedding and then permanently wedge the server; CPU output is correct",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-23T14:28:31Z",
        "updated_at": "2026-08-11T13:43:15Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26044",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26047,
        "kind": "issue",
        "state": "open",
        "title": "mtmd: `*_init` allocators can throw std::bad_alloc across the extern \"C\" boundary (unlike their siblings)",
        "labels": [
          "stale"
        ],
        "created_at": "2026-07-23T17:49:09Z",
        "updated_at": "2026-08-23T01:07:08Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26047",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26048,
        "kind": "pr",
        "state": "open",
        "title": "CUDA: Optimize prefil via fuse of w_s scale in epilogue MMQ for nvfp4 checkpoints",
        "labels": [
          "build",
          "testing",
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-07-23T17:56:42Z",
        "updated_at": "2026-08-18T00:17:29Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26048",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26057,
        "kind": "pr",
        "state": "open",
        "title": "metal : add support for GGML_OP_REPEAT_BACK",
        "labels": [
          "ggml",
          "Apple Metal"
        ],
        "created_at": "2026-07-24T02:04:57Z",
        "updated_at": "2026-07-24T02:05:24Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26057",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26058,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: [SYCL] Garbled output on 2+ GPUs",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-24T03:05:25Z",
        "updated_at": "2026-07-28T15:13:11Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26058",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26060,
        "kind": "issue",
        "state": "open",
        "title": "Documented Linux HIP Compile Command Fails with pip-installed ROCm",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-07-24T03:47:15Z",
        "updated_at": "2026-08-24T01:07:00Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26060",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26070,
        "kind": "pr",
        "state": "open",
        "title": "ggml : handle graph buffer reservation failure",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-07-24T09:33:05Z",
        "updated_at": "2026-07-24T09:34:03Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26070",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 26073,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Bonsai 27B ternary (Q2) not running on Metal",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-24T10:30:19Z",
        "updated_at": "2026-07-29T17:04:44Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26073",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing"
        ]
      },
      {
        "number": 26078,
        "kind": "pr",
        "state": "open",
        "title": "kleidiai: Update KleidiAI Documentation",
        "labels": [
          "documentation",
          "ggml"
        ],
        "created_at": "2026-07-24T11:20:34Z",
        "updated_at": "2026-08-26T08:32:48Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26078",
        "categories": [
          "models_architectures",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26084,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Content chat have many '/'",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-07-24T15:41:14Z",
        "updated_at": "2026-08-24T01:06:59Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26084",
        "categories": [
          "native_backends",
          "models_architectures",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26087,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Qwen 3.5 122B hangs and spins at 100% cpu usage using Vulkan on AI Max during model loading",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-07-24T16:59:51Z",
        "updated_at": "2026-08-24T01:06:58Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26087",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26088,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Gemma 4 31B enters `<unused49>` (token ID 62) infinite generation loop after long idle period, reproducible across multiple GGUF quant sources",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-24T17:52:37Z",
        "updated_at": "2026-08-23T01:14:02Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26088",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26094,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: quantized Qwen3.5 `ssm_out` falls back to CPU for` batch>1` on Hexagon",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-24T18:57:48Z",
        "updated_at": "2026-08-05T10:45:30Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26094",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26095,
        "kind": "pr",
        "state": "open",
        "title": "Add LlamaNet to the list of tools in the README.",
        "labels": [
          "documentation"
        ],
        "created_at": "2026-07-24T19:09:13Z",
        "updated_at": "2026-07-26T00:11:04Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26095",
        "categories": [
          "models_architectures",
          "api_ui_docs"
        ]
      },
      {
        "number": 26096,
        "kind": "pr",
        "state": "open",
        "title": "docs: use ROCM_PATH instead of HIP_PATH in linux HIP build command (#26060)",
        "labels": [
          "documentation"
        ],
        "created_at": "2026-07-24T19:15:03Z",
        "updated_at": "2026-08-26T02:20:27Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26096",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26100,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: speculative draft-cache replay path bypasses p_min - inflates repeated-prompt benchmarks ~10x, degrades mixed-traffic serving ~3x",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-24T22:58:53Z",
        "updated_at": "2026-07-27T22:45:08Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26100",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26103,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cpu: skip unsupported ARM ISA variants under GGML_CPU_ALL_VARIANTS",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-07-25T00:37:17Z",
        "updated_at": "2026-07-25T04:50:06Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26103",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26104,
        "kind": "pr",
        "state": "open",
        "title": "cmake : fall back to empty UI on stale assets when provisioning is disabled",
        "labels": [
          "testing",
          "server/ui"
        ],
        "created_at": "2026-07-25T01:31:13Z",
        "updated_at": "2026-08-11T12:59:26Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26104",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26106,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Starting from B9992, long contexts trigger a CUDA error: the function failed to launch on the GPU",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-07-25T04:08:28Z",
        "updated_at": "2026-08-25T01:06:57Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26106",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26108,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Failed to read image when using DFlash (Qwen3.6 27B)",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-25T06:33:32Z",
        "updated_at": "2026-08-22T11:09:07Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26108",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26113,
        "kind": "pr",
        "state": "open",
        "title": "hexagon: enable quantized matmul for multi-sequence inputs",
        "labels": [
          "ggml",
          "Hexagon"
        ],
        "created_at": "2026-07-25T11:11:34Z",
        "updated_at": "2026-07-25T11:11:50Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26113",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 26114,
        "kind": "pr",
        "state": "open",
        "title": "Fix speculative models failing to load when running llama-server",
        "labels": [],
        "created_at": "2026-07-25T12:46:32Z",
        "updated_at": "2026-07-28T15:47:49Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26114",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "build_platform_ops"
        ]
      },
      {
        "number": 26115,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: support Upstage's Solar-Open2",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-07-25T13:20:36Z",
        "updated_at": "2026-07-28T22:52:00Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26115",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 26116,
        "kind": "issue",
        "state": "open",
        "title": "allow `llama serve -hf` to use llama-server in router mode",
        "labels": [],
        "created_at": "2026-07-25T16:09:38Z",
        "updated_at": "2026-08-23T22:42:52Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26116",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "api_ui_docs"
        ]
      },
      {
        "number": 26122,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: --swa-full + -sm tensor crashes on load for Gemma4-31B (ISWA KV cache / ggml_backend_meta_alloc_ctx_tensors_from_buft)",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-25T20:12:02Z",
        "updated_at": "2026-08-02T00:42:59Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26122",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26123,
        "kind": "issue",
        "state": "open",
        "title": "Hexagon: dspqueue_read failed: 0x0000002e during graph compute on Snapdragon 8 Gen 2 (v73), reproducible at 4B+ params, works fine at 1B",
        "labels": [
          "stale"
        ],
        "created_at": "2026-07-25T21:35:07Z",
        "updated_at": "2026-08-25T01:06:57Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26123",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 26125,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Support Motif-3-Beta",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-07-26T00:10:25Z",
        "updated_at": "2026-07-29T23:41:40Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26125",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26127,
        "kind": "issue",
        "state": "open",
        "title": "Windows HIP release omits gfx1152; Radeon 860M fails with invalid device function",
        "labels": [
          "stale"
        ],
        "created_at": "2026-07-26T01:13:37Z",
        "updated_at": "2026-08-26T01:07:08Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26127",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "build_platform_ops"
        ]
      },
      {
        "number": 26128,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: server prompt cache is incompatible with the RPC backend — GGML_ASSERT(tensor->data != NULL) in prompt_save() → state_seq_get_data(), triggered as soon as -np > 1",
        "labels": [],
        "created_at": "2026-07-26T02:14:51Z",
        "updated_at": "2026-07-30T12:37:56Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26128",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26129,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: server: expose per-device memory usage (weights / context / compute)",
        "labels": [],
        "created_at": "2026-07-26T04:03:59Z",
        "updated_at": "2026-08-09T20:09:16Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26129",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26130,
        "kind": "pr",
        "state": "open",
        "title": "server : expose per-device memory usage on /metrics and GET /memory",
        "labels": [
          "documentation",
          "server",
          "mtmd"
        ],
        "created_at": "2026-07-26T04:04:00Z",
        "updated_at": "2026-08-15T19:34:29Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26130",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26132,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cpu : fix conv_transpose_2d for multiple batches",
        "labels": [
          "testing",
          "ggml"
        ],
        "created_at": "2026-07-26T08:47:57Z",
        "updated_at": "2026-07-26T08:48:15Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26132",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26136,
        "kind": "pr",
        "state": "open",
        "title": "feat: QFX16/QFX32 lossless quantization (2.05x compression, zero quality loss)",
        "labels": [
          "examples",
          "ggml",
          "Apple Metal"
        ],
        "created_at": "2026-07-26T13:12:58Z",
        "updated_at": "2026-07-26T18:22:25Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26136",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26139,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Gigatoken Tokenizer",
        "labels": [
          "enhancement",
          "stale"
        ],
        "created_at": "2026-07-26T14:15:45Z",
        "updated_at": "2026-08-26T01:07:07Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26139",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 26142,
        "kind": "issue",
        "state": "open",
        "title": "llama_get_embeddings_ith returns NaN/zeros for output slots 2+ with embd-batch decode on Qwen3-VL",
        "labels": [
          "stale"
        ],
        "created_at": "2026-07-26T16:52:46Z",
        "updated_at": "2026-08-26T01:07:06Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26142",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26143,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: RPC Node Cache Grows without Bound?",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-26T17:46:27Z",
        "updated_at": "2026-08-24T23:15:30Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26143",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 26145,
        "kind": "pr",
        "state": "open",
        "title": "Responses API: Convert text.format JSON schema into the corresponding Chat Completions response_format",
        "labels": [
          "testing",
          "server"
        ],
        "created_at": "2026-07-26T18:47:17Z",
        "updated_at": "2026-08-18T10:13:38Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26145",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 26148,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: ROCm gfx1151, GGML_CUDA_ENABLE_UNIFIED_MEMORY=1 causes silent output corruption with Qwen3-Coder-Next",
        "labels": [
          "bug-unconfirmed",
          "stale"
        ],
        "created_at": "2026-07-26T20:21:10Z",
        "updated_at": "2026-08-26T01:07:05Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26148",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26152,
        "kind": "issue",
        "state": "open",
        "title": "[MoE] Scheduler crash during partial offload (GGML_SCHED_MAX_SPLIT_INPUTS)",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-26T22:38:05Z",
        "updated_at": "2026-08-02T16:11:56Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26152",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26153,
        "kind": "pr",
        "state": "open",
        "title": "skills: add ggml-test skill",
        "labels": [
          "documentation"
        ],
        "created_at": "2026-07-26T23:04:26Z",
        "updated_at": "2026-07-26T23:10:39Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26153",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 26157,
        "kind": "pr",
        "state": "open",
        "title": "ggml: add support for MXFP8 CPU",
        "labels": [
          "ggml",
          "conversion"
        ],
        "created_at": "2026-07-27T02:39:55Z",
        "updated_at": "2026-08-21T08:47:04Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26157",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26159,
        "kind": "pr",
        "state": "open",
        "title": "cuda: compact Blackwell NVFP4 MoE work scheduling (+10% to +15% prefill, NVFP4-only)",
        "labels": [
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-07-27T02:51:28Z",
        "updated_at": "2026-07-27T08:43:54Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26159",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26163,
        "kind": "issue",
        "state": "open",
        "title": "Vulkan: AMD flash-attention tuning gated on `maxComputeSharedMemorySize == 65536` is skipped when driver reports 32768 (Vega/gfx90c, Adrenalin 26.5.2) - ~17% slowdown",
        "labels": [],
        "created_at": "2026-07-27T05:44:36Z",
        "updated_at": "2026-08-18T21:39:09Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26163",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26164,
        "kind": "pr",
        "state": "open",
        "title": "[Model/VLA] Support MiniCPM-RobotManip",
        "labels": [
          "documentation",
          "build",
          "server",
          "mtmd"
        ],
        "created_at": "2026-07-27T07:57:16Z",
        "updated_at": "2026-08-03T06:50:05Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26164",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26167,
        "kind": "pr",
        "state": "open",
        "title": "ggml: add a scheduler sanitizer",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-07-27T09:36:22Z",
        "updated_at": "2026-07-27T17:21:48Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26167",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26169,
        "kind": "issue",
        "state": "open",
        "title": "server: inconsistent `created` timestamp in `/v1/models` endpoint breaks client caching",
        "labels": [],
        "created_at": "2026-07-27T09:53:49Z",
        "updated_at": "2026-07-27T09:53:49Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26169",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26174,
        "kind": "pr",
        "state": "open",
        "title": "Add more benchmarks to llama-eval",
        "labels": [
          "documentation",
          "examples"
        ],
        "created_at": "2026-07-27T12:18:07Z",
        "updated_at": "2026-07-31T14:29:59Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26174",
        "categories": [
          "quantization",
          "models_architectures",
          "api_ui_docs"
        ]
      },
      {
        "number": 26179,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: OpenVINO backend fails on gpt-oss-20b Q4_0/MXFP4 with incompatible reshape on attention weights",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-27T14:11:13Z",
        "updated_at": "2026-07-30T05:52:50Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26179",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26182,
        "kind": "pr",
        "state": "open",
        "title": "feat: Add '--ignore-api-samplers' CLI flag to ignore API sampler settings",
        "labels": [
          "server"
        ],
        "created_at": "2026-07-27T15:37:06Z",
        "updated_at": "2026-07-27T15:37:26Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26182",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26186,
        "kind": "issue",
        "state": "open",
        "title": "Compile bug: ggml-cpu: clang-cl build fails with GGML_AVX_VNNI (alderlake variant of GGML_CPU_ALL_VARIANTS): always_inline requires target feature 'avxvnni'",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-27T18:06:21Z",
        "updated_at": "2026-07-27T18:06:21Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26186",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26187,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cpu: add -mavxvnni for clang-cl when GGML_AVX_VNNI is enabled",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-07-27T18:25:18Z",
        "updated_at": "2026-07-27T18:26:33Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26187",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26188,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: CUDA 5060",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-27T18:47:08Z",
        "updated_at": "2026-07-27T19:18:15Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26188",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26195,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Vulkan flash-attn produces garbled output with quantized V cache on head_dim=256 models (Adreno X1-85)",
        "labels": [],
        "created_at": "2026-07-27T20:39:45Z",
        "updated_at": "2026-07-27T20:39:45Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26195",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 26197,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Vulkan - Using multiple AMD GPUs lead to gibberish output",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-27T21:25:22Z",
        "updated_at": "2026-07-27T21:41:05Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26197",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26200,
        "kind": "issue",
        "state": "open",
        "title": "CPU decode: GGML_OPENMP=ON with mingw-w64 libgomp costs ~40% on small-op-heavy graphs (every graph barrier is a kernel semaphore; the port cannot spin)",
        "labels": [],
        "created_at": "2026-07-27T21:49:36Z",
        "updated_at": "2026-07-27T21:49:36Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26200",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 26201,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: <__media__> is being ignored when embedding with Qwen3-VL-Embedding-8B",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-27T22:07:26Z",
        "updated_at": "2026-07-28T08:24:49Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26201",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26204,
        "kind": "pr",
        "state": "open",
        "title": "server : add /slots endpoint action=clone_to (KV clone between slots)",
        "labels": [
          "documentation",
          "server"
        ],
        "created_at": "2026-07-27T22:49:32Z",
        "updated_at": "2026-07-29T18:10:13Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26204",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 26205,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Silent crash during/before server initialization on RTX 5060 (sm_120) using Andgihat Blackwell build",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-27T23:40:19Z",
        "updated_at": "2026-07-27T23:40:19Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26205",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 26206,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Gemma 4 12B produces garbled output on large prompts, Intel Arc Pro B70 (Xe2)",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-27T23:43:35Z",
        "updated_at": "2026-08-03T09:00:53Z",
        "closed_at": null,
        "comments": 11,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26206",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26207,
        "kind": "issue",
        "state": "open",
        "title": "server: prompt cache is reused across requests with different per-request `lora` — output silently contaminated by the previous adapter",
        "labels": [],
        "created_at": "2026-07-28T00:24:25Z",
        "updated_at": "2026-07-28T00:24:25Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26207",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 26208,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: VRAM does not allocate with ROCm 7.14 on gfx1201",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-28T01:07:35Z",
        "updated_at": "2026-08-25T18:15:53Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26208",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26209,
        "kind": "issue",
        "state": "open",
        "title": "Regression after c7d8722: llama-server generates infinite \"/\" tokens in agent mode",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-28T01:28:18Z",
        "updated_at": "2026-08-21T01:52:51Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26209",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 26213,
        "kind": "issue",
        "state": "open",
        "title": "Bug: Vulkan on Raspberry Pi 5 (V3D) — WG=512 / 16KB SMEM breaks inference",
        "labels": [],
        "created_at": "2026-07-28T03:45:11Z",
        "updated_at": "2026-07-31T01:29:56Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26213",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26215,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Trapped into garbled \"/\" loop with a specific conversation",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-28T04:02:19Z",
        "updated_at": "2026-08-21T17:15:41Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26215",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 26219,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Intel UHD 770 graphics crashes on Vulkan TOP_K unit test",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-28T08:02:59Z",
        "updated_at": "2026-08-17T05:45:16Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26219",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26220,
        "kind": "issue",
        "state": "open",
        "title": "Bug: Native MMA FA kernel regresses prompt processing up to 2x at depth on RDNA4 (gfx1201) after rocWMMA removal",
        "labels": [],
        "created_at": "2026-07-28T09:26:37Z",
        "updated_at": "2026-08-25T22:20:44Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26220",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26223,
        "kind": "pr",
        "state": "open",
        "title": "metal: fix NaN in mul_mm_id when activations exceed f16 range",
        "labels": [
          "testing",
          "ggml",
          "Apple Metal"
        ],
        "created_at": "2026-07-28T10:32:37Z",
        "updated_at": "2026-08-22T07:32:03Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26223",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26224,
        "kind": "issue",
        "state": "open",
        "title": "core: support output vocab size distinct from embedding vocab size",
        "labels": [],
        "created_at": "2026-07-28T10:35:17Z",
        "updated_at": "2026-08-07T12:42:14Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26224",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26226,
        "kind": "pr",
        "state": "open",
        "title": "core : support output vocab size distinct from embedding vocab size",
        "labels": [
          "conversion"
        ],
        "created_at": "2026-07-28T11:25:12Z",
        "updated_at": "2026-08-11T15:21:39Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26226",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26236,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: llama-bench should move to next combination on error",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-07-28T17:24:48Z",
        "updated_at": "2026-07-28T17:24:48Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26236",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26237,
        "kind": "pr",
        "state": "open",
        "title": "mtmd: bound InternVL preproc_max_tiles read from GGUF",
        "labels": [
          "mtmd"
        ],
        "created_at": "2026-07-28T17:49:26Z",
        "updated_at": "2026-08-10T21:01:04Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26237",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion"
        ]
      },
      {
        "number": 26238,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Hy3 Performance is very poor",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-28T18:21:47Z",
        "updated_at": "2026-08-21T12:46:09Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26238",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26239,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Gemma4 <unused49> infinite generation / unstable output on gfx1151 (HIP, Windows) with long prompts",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-28T18:45:58Z",
        "updated_at": "2026-08-23T01:27:49Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26239",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 26243,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: vulkan: command-buffer debug labels for GPU profilers",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-07-28T20:06:21Z",
        "updated_at": "2026-07-28T20:06:21Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26243",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26245,
        "kind": "pr",
        "state": "open",
        "title": "ggml: fix integer overflow in tensor size computation (ggml_new_tensor_impl)",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-07-28T20:54:23Z",
        "updated_at": "2026-08-06T20:40:22Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26245",
        "categories": [
          "quantization"
        ]
      },
      {
        "number": 26246,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Ollama experiencing CUDA crash during a Gemma4b12 load",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-28T21:14:20Z",
        "updated_at": "2026-07-28T21:14:20Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26246",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26249,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Huggingface remove old models on model updates",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-07-28T22:00:17Z",
        "updated_at": "2026-07-28T22:00:17Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26249",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 26255,
        "kind": "pr",
        "state": "open",
        "title": "quant : do not require imatrix when the tensor keeps its type",
        "labels": [],
        "created_at": "2026-07-28T23:25:42Z",
        "updated_at": "2026-07-29T00:06:50Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26255",
        "categories": [
          "quantization",
          "models_architectures",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26257,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Qwen3.6-27B garbled output on dual-GPU CUDA (RTX 5060 Ti + RTX 3060), single GPU works",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-29T01:37:49Z",
        "updated_at": "2026-08-22T04:35:25Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26257",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 26258,
        "kind": "pr",
        "state": "open",
        "title": "ci: try using gpu-webgpu-intel in CI",
        "labels": [
          "devops"
        ],
        "created_at": "2026-07-29T02:41:56Z",
        "updated_at": "2026-08-11T13:44:43Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26258",
        "categories": [
          "native_backends",
          "models_architectures",
          "build_platform_ops"
        ]
      },
      {
        "number": 26266,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Thinking process triggered resulting in doubled generation time",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-29T08:05:51Z",
        "updated_at": "2026-07-29T08:50:37Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26266",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26268,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: llama_params_fit aborts instead of reporting FAILURE for a large context",
        "labels": [],
        "created_at": "2026-07-29T09:48:24Z",
        "updated_at": "2026-07-29T09:48:24Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26268",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 26273,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Add a CLI flag to escape special tokens in user input and tool responses",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-07-29T12:48:07Z",
        "updated_at": "2026-07-29T12:50:49Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26273",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26282,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: llama-server --embedding got Corrupted Data, quality decline",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-29T15:00:30Z",
        "updated_at": "2026-07-30T08:30:11Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26282",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 26283,
        "kind": "pr",
        "state": "open",
        "title": "Suffix decode",
        "labels": [
          "documentation",
          "model",
          "testing",
          "Vulkan",
          "devops",
          "server",
          "ggml",
          "SYCL",
          "Apple Metal",
          "OpenCL",
          "mtmd",
          "CUDA",
          "AMD ZenDNN",
          "WebGPU",
          "server/ui",
          "conversion",
          "vendor"
        ],
        "created_at": "2026-07-29T15:26:20Z",
        "updated_at": "2026-08-20T18:41:40Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26283",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26284,
        "kind": "pr",
        "state": "open",
        "title": "HIP: tune rdna 3 and rdna 4 mmq configs",
        "labels": [
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-07-29T15:55:00Z",
        "updated_at": "2026-08-26T07:30:54Z",
        "closed_at": null,
        "comments": 16,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26284",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26285,
        "kind": "issue",
        "state": "open",
        "title": "ggml-cuda: MMQ incorrectly disabled on RTX 3090 (shared memory check)",
        "labels": [],
        "created_at": "2026-07-29T16:05:28Z",
        "updated_at": "2026-08-16T10:01:25Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26285",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26286,
        "kind": "pr",
        "state": "open",
        "title": "qwen3 : add sliding-window attention pattern support",
        "labels": [
          "model",
          "examples",
          "conversion"
        ],
        "created_at": "2026-07-29T16:32:14Z",
        "updated_at": "2026-07-31T10:27:42Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26286",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26288,
        "kind": "pr",
        "state": "open",
        "title": "Feature/ngram map tests",
        "labels": [
          "testing"
        ],
        "created_at": "2026-07-29T18:23:03Z",
        "updated_at": "2026-08-01T23:13:11Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26288",
        "categories": [
          "models_architectures",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26289,
        "kind": "pr",
        "state": "open",
        "title": "CUDA: tune fp16 tile FlashAttention configs for head sizes 40-112",
        "labels": [
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-07-29T18:26:47Z",
        "updated_at": "2026-07-29T18:33:52Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26289",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26291,
        "kind": "pr",
        "state": "open",
        "title": "rpc: parallelize cached tensor hashing during model load",
        "labels": [
          "documentation",
          "examples",
          "ggml"
        ],
        "created_at": "2026-07-29T19:00:40Z",
        "updated_at": "2026-08-26T17:30:52Z",
        "closed_at": null,
        "comments": 9,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26291",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 26292,
        "kind": "pr",
        "state": "open",
        "title": "fix: return fit failure when scheduler allocation fails",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-07-29T19:05:39Z",
        "updated_at": "2026-07-31T15:45:46Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26292",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving"
        ]
      },
      {
        "number": 26293,
        "kind": "issue",
        "state": "open",
        "title": "server: prompt cache is populated by rerank tasks that can never read it back (cache_idle_slots missing the SERVER_TASK_TYPE_COMPLETION guard)",
        "labels": [],
        "created_at": "2026-07-29T20:43:32Z",
        "updated_at": "2026-08-21T16:24:30Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26293",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "build_platform_ops"
        ]
      },
      {
        "number": 26294,
        "kind": "pr",
        "state": "open",
        "title": "CUDA: fix duplicate expert id compaction in mul_mat_id (#24591)",
        "labels": [
          "testing",
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-07-29T21:05:48Z",
        "updated_at": "2026-08-24T05:02:05Z",
        "closed_at": null,
        "comments": 9,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26294",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26298,
        "kind": "pr",
        "state": "open",
        "title": "model: Add support for Motif 3 Beta ",
        "labels": [
          "model",
          "testing",
          "ggml",
          "CUDA",
          "conversion"
        ],
        "created_at": "2026-07-29T23:38:28Z",
        "updated_at": "2026-08-26T15:07:56Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26298",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 26300,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Support Apertus (a family of models from ETH Zurich)",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-07-30T02:03:52Z",
        "updated_at": "2026-08-07T12:40:48Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26300",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 26301,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cuda: add dequant-float matvec (mmvdq) for Q4_K/Q5_K/Q6_K",
        "labels": [
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-07-30T03:28:19Z",
        "updated_at": "2026-08-12T19:14:19Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26301",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 26305,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Add support for Escha-W2 (2-bit) quantization format",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-07-30T06:05:26Z",
        "updated_at": "2026-07-30T06:10:09Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26305",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 26306,
        "kind": "issue",
        "state": "open",
        "title": "Update vulkan SDK in Linux vulkan builders to latest 1.4.357 release..",
        "labels": [],
        "created_at": "2026-07-30T07:05:16Z",
        "updated_at": "2026-07-30T07:10:24Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26306",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26308,
        "kind": "pr",
        "state": "open",
        "title": "cuda: support row-contiguous SUM_ROWS",
        "labels": [
          "testing",
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-07-30T08:28:14Z",
        "updated_at": "2026-08-22T21:57:05Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26308",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26309,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: server: post /tokenize \"parse_special\": false is not being respected (b10147)",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-30T08:34:27Z",
        "updated_at": "2026-07-30T08:34:27Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26309",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "speculative_generation",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 26310,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cpu/repack: size mul_mat_id row list by slots x tokens",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-07-30T09:17:48Z",
        "updated_at": "2026-07-30T09:31:22Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26310",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26311,
        "kind": "pr",
        "state": "open",
        "title": "CUDA: linear weight repack + MMVQ for NVFP4",
        "labels": [
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-07-30T09:41:18Z",
        "updated_at": "2026-07-30T13:46:34Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26311",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26312,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Hy3 no thinking",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-30T09:45:33Z",
        "updated_at": "2026-07-30T20:19:48Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26312",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26314,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: gibberish output while inferencing on gpu",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-30T11:05:25Z",
        "updated_at": "2026-07-30T11:05:25Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26314",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26315,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: server: /slots?action=save follows a symlink out of --slot-save-path and overwrites the target",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-30T11:19:33Z",
        "updated_at": "2026-07-30T11:27:06Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26315",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26316,
        "kind": "pr",
        "state": "open",
        "title": "server : contain slot save/restore paths within --slot-save-path (#26315)",
        "labels": [
          "server"
        ],
        "created_at": "2026-07-30T11:37:53Z",
        "updated_at": "2026-07-30T12:27:21Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26316",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26318,
        "kind": "pr",
        "state": "open",
        "title": "ggml: fix size wrap in allocator and gguf alignment read",
        "labels": [
          "testing",
          "ggml"
        ],
        "created_at": "2026-07-30T12:19:18Z",
        "updated_at": "2026-07-31T02:24:18Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26318",
        "categories": [
          "quantization",
          "models_architectures",
          "correctness_testing"
        ]
      },
      {
        "number": 26321,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Reasoning button is gone",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-30T13:59:19Z",
        "updated_at": "2026-08-05T17:24:42Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26321",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26322,
        "kind": "pr",
        "state": "open",
        "title": "CUDA: add MOE_FFN",
        "labels": [
          "testing",
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-07-30T14:08:28Z",
        "updated_at": "2026-08-01T11:45:18Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26322",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26323,
        "kind": "pr",
        "state": "open",
        "title": "et-backend: Q4_0 matrix-engine GEMM rewrite for prefill",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-07-30T14:11:29Z",
        "updated_at": "2026-08-19T20:58:08Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26323",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26325,
        "kind": "issue",
        "state": "open",
        "title": "llama-server.exe (Windows MSVC) returns 500 on CJK output — invalid UTF-8 in JSON response",
        "labels": [],
        "created_at": "2026-07-30T15:19:22Z",
        "updated_at": "2026-07-30T15:19:22Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26325",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26326,
        "kind": "pr",
        "state": "open",
        "title": "et-backend: Q8_0 matrix-engine GEMM kernel for prefill",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-07-30T15:30:51Z",
        "updated_at": "2026-08-11T15:03:59Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26326",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26327,
        "kind": "pr",
        "state": "open",
        "title": "et-backend: F16 vecdot GEMV + matrix-engine GEMM",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-07-30T15:30:58Z",
        "updated_at": "2026-08-11T15:01:27Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26327",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26328,
        "kind": "pr",
        "state": "open",
        "title": "et-backend: F32 vecdot GEMV + matrix-engine GEMM",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-07-30T15:31:06Z",
        "updated_at": "2026-08-11T15:22:01Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26328",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26330,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: test-quant-type-selection --generate returns HTTP 401",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-30T15:54:15Z",
        "updated_at": "2026-07-30T15:54:15Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26330",
        "categories": [
          "quantization",
          "models_architectures",
          "loader_conversion",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 26331,
        "kind": "pr",
        "state": "open",
        "title": "opencl: add Adreno xmem SDPA path",
        "labels": [
          "ggml",
          "OpenCL"
        ],
        "created_at": "2026-07-30T16:01:26Z",
        "updated_at": "2026-08-25T14:43:29Z",
        "closed_at": null,
        "comments": 11,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26331",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26335,
        "kind": "issue",
        "state": "open",
        "title": "[feature request] moe optimizations on ssd",
        "labels": [],
        "created_at": "2026-07-30T17:14:59Z",
        "updated_at": "2026-07-30T17:14:59Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26335",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving"
        ]
      },
      {
        "number": 26337,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Ternary-Bonsai-27B fails to load when using its Dspark draft model",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-30T17:49:08Z",
        "updated_at": "2026-08-10T05:42:04Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26337",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "loader_conversion",
          "correctness_testing"
        ]
      },
      {
        "number": 26339,
        "kind": "issue",
        "state": "open",
        "title": "Dspark model fails to load with -sm tensor on MI50",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-30T18:28:33Z",
        "updated_at": "2026-07-30T18:28:33Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26339",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26340,
        "kind": "issue",
        "state": "open",
        "title": "Compile bug: Problem with no matching function for call to ‘std::vector<nlohmann::json_abi..",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-30T19:11:26Z",
        "updated_at": "2026-07-30T20:27:45Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26340",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26346,
        "kind": "issue",
        "state": "open",
        "title": "Vulkan: ~6-8% MoE prompt-processing regression on 64-CU RDNA4 (RX 9070 XT) since #25005 (flops-based submission heuristic)",
        "labels": [],
        "created_at": "2026-07-30T21:32:49Z",
        "updated_at": "2026-08-02T17:37:36Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26346",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26348,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cpu: add x86 VNNI Q2_0 dot product -- 3x speed improvement for VNNI-compatible CPUs",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-07-30T22:10:42Z",
        "updated_at": "2026-08-20T03:59:41Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26348",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26351,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Voxtral audio: a single clip is encoded twice by mtmd",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-31T02:22:53Z",
        "updated_at": "2026-07-31T02:22:53Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26351",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26355,
        "kind": "issue",
        "state": "open",
        "title": "Request for a private security reporting channel - no vulnerability details",
        "labels": [],
        "created_at": "2026-07-31T05:40:45Z",
        "updated_at": "2026-07-31T05:40:45Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26355",
        "categories": [
          "native_backends",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26356,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Tool Call Performance degradation over sustained load on unified KV cache",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-31T05:56:11Z",
        "updated_at": "2026-07-31T06:05:27Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26356",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26358,
        "kind": "pr",
        "state": "open",
        "title": "vulkan : disable FA GQA packing for multi-token queries",
        "labels": [
          "testing",
          "Vulkan",
          "ggml"
        ],
        "created_at": "2026-07-31T07:11:16Z",
        "updated_at": "2026-08-26T06:18:10Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26358",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26359,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: llama-server returns a tool-call id that its own chat template then rejects (Ministral-8B, --jinja)",
        "labels": [],
        "created_at": "2026-07-31T07:25:37Z",
        "updated_at": "2026-07-31T07:25:37Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26359",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 26360,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: vulkan - multi-token GQA FA packing like CUDA ncols1 x ncols2",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-07-31T08:01:36Z",
        "updated_at": "2026-07-31T09:09:18Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26360",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 26363,
        "kind": "issue",
        "state": "open",
        "title": "Quantized T5 (Aya-101/mT5-XXL) produces garbage output on Vulkan backend, works correctly on CPU",
        "labels": [],
        "created_at": "2026-07-31T09:28:53Z",
        "updated_at": "2026-07-31T09:28:53Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26363",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26365,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Enable split-mode row/tensor for kimi-k3-fullsize-vision branch",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-07-31T10:42:49Z",
        "updated_at": "2026-07-31T10:42:49Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26365",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26366,
        "kind": "issue",
        "state": "open",
        "title": "gguf: SIGFPE on x86-64 from a zero tensor dimension (valid GGUF, accepted by gguf-py)",
        "labels": [],
        "created_at": "2026-07-31T11:22:44Z",
        "updated_at": "2026-07-31T11:22:44Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26366",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "multimodal_embedding",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 26367,
        "kind": "issue",
        "state": "open",
        "title": "ggml-backend-meta: `axis <= GGML_MAX_DIMS` off-by-one (inconsistent with ~21 other sites)",
        "labels": [],
        "created_at": "2026-07-31T11:22:58Z",
        "updated_at": "2026-08-23T07:55:05Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26367",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26369,
        "kind": "issue",
        "state": "open",
        "title": "Deepseek4",
        "labels": [],
        "created_at": "2026-07-31T12:33:39Z",
        "updated_at": "2026-08-04T19:19:25Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26369",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "speculative_generation",
          "structured_generation_tools"
        ]
      },
      {
        "number": 26381,
        "kind": "issue",
        "state": "open",
        "title": "Generation bug: The model produced output that does not match the expected peg-native format",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-31T18:33:12Z",
        "updated_at": "2026-08-01T00:30:34Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26381",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26382,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: same K and V cache type enforced for models with no V cache",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-31T18:56:26Z",
        "updated_at": "2026-08-07T13:29:13Z",
        "closed_at": null,
        "comments": 10,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26382",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing"
        ]
      },
      {
        "number": 26384,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: deepseek_v4 over RPC (2-node Metal): empty assistant turns degenerate output to \"====\", then a worker-side graph failure kills the whole server (RPC has no error recovery)",
        "labels": [],
        "created_at": "2026-07-31T20:07:50Z",
        "updated_at": "2026-07-31T20:07:50Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26384",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 26390,
        "kind": "pr",
        "state": "open",
        "title": "CUDA: use batched argsort for multi row top-k",
        "labels": [
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-07-31T21:34:45Z",
        "updated_at": "2026-08-12T13:35:31Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26390",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26391,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: --batch-size and --ubatch-size ignored when set to 1024/1024",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-07-31T21:39:07Z",
        "updated_at": "2026-08-25T07:49:47Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26391",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26393,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: windows OpenVINO cannot use the GPU device",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-01T03:19:18Z",
        "updated_at": "2026-08-07T08:50:04Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26393",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 26394,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: chat responses mode need change conver logic",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-01T04:17:51Z",
        "updated_at": "2026-08-19T21:17:03Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26394",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 26395,
        "kind": "pr",
        "state": "open",
        "title": "ggml-backend-meta: fix axis bounds check off-by-one",
        "labels": [
          "testing",
          "ggml"
        ],
        "created_at": "2026-08-01T06:06:21Z",
        "updated_at": "2026-08-02T06:43:28Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26395",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26399,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: GGML_OP_TOP_K falls back to CPU on HIP/ROCm above ~3–4K context — 6.4× token-generation loss on DeepSeek-V4-Flash",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-01T11:24:51Z",
        "updated_at": "2026-08-04T19:19:17Z",
        "closed_at": null,
        "comments": 19,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26399",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26401,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: mmproj memory margin is added twice on resume from sleep, causing CPU offload",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-01T13:09:48Z",
        "updated_at": "2026-08-26T17:18:39Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26401",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing"
        ]
      },
      {
        "number": 26402,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: llama-cpp loading model to dGPU VRAM, but using iGPU for inference when compressing KV-cache",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-01T14:05:35Z",
        "updated_at": "2026-08-06T12:08:53Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26402",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 26404,
        "kind": "pr",
        "state": "open",
        "title": "CUDA: FA support for head size 192/128 with GQA ratios that are not a multiples of 8.",
        "labels": [
          "testing",
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-08-01T15:25:17Z",
        "updated_at": "2026-08-12T18:36:14Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26404",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26409,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: SYCL `--split-mode tensor` is 3x slower than a single GPU, and hangs with a quantized KV cache",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-01T18:30:28Z",
        "updated_at": "2026-08-17T12:47:12Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26409",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26414,
        "kind": "pr",
        "state": "open",
        "title": "llama-hot-experts: pin hottest MoE experts in RAM via --pin-hot-experts",
        "labels": [],
        "created_at": "2026-08-01T20:38:21Z",
        "updated_at": "2026-08-09T14:12:46Z",
        "closed_at": null,
        "comments": 14,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26414",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26416,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Windows: common_params_parse discards the caller's argv",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-01T21:16:55Z",
        "updated_at": "2026-08-01T21:17:23Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26416",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26417,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Incoherent output for BF16 gemma-4 MoE models on devices without VK_KHR_shader_bfloat16",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-01T21:47:40Z",
        "updated_at": "2026-08-02T16:42:12Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26417",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 26419,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cuda: enable MMA FlashAttention for head dim 256 on AMD RDNA",
        "labels": [
          "testing",
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-08-01T22:43:52Z",
        "updated_at": "2026-08-25T22:10:07Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26419",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26425,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: MTP retains inter-request state causing non-deterministic output and model degradation (Qwen3.6-35B-A3B-MTP)",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-02T02:18:19Z",
        "updated_at": "2026-08-25T04:23:56Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26425",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26426,
        "kind": "pr",
        "state": "open",
        "title": "sampling : fix heap corruption in llama_sampler_init_dry_testing",
        "labels": [
          "testing"
        ],
        "created_at": "2026-08-02T02:23:30Z",
        "updated_at": "2026-08-05T06:26:56Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26426",
        "categories": [
          "models_architectures",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26429,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Support Inline Image Rendering in Chat Messages (including MCP tool outputs)",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-08-02T02:43:49Z",
        "updated_at": "2026-08-02T02:43:49Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26429",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26430,
        "kind": "pr",
        "state": "open",
        "title": "opencl: extend, speed up and widen the MQ/cluster decode flash-attention (head sizes 64-512, GQA 2-16)",
        "labels": [
          "testing",
          "ggml",
          "Apple Metal",
          "OpenCL"
        ],
        "created_at": "2026-08-02T02:55:53Z",
        "updated_at": "2026-08-22T20:47:19Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26430",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26432,
        "kind": "issue",
        "state": "open",
        "title": "Silent GTT fallback when context + MTP exceeds VRAM — no error at load, massive throughput collapse at runtime",
        "labels": [],
        "created_at": "2026-08-02T03:24:57Z",
        "updated_at": "2026-08-25T07:49:47Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26432",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "loader_conversion"
        ]
      },
      {
        "number": 26435,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: HIP/ROCm: Regression in prompt processing MI210",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-02T03:57:47Z",
        "updated_at": "2026-08-22T21:03:55Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26435",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing"
        ]
      },
      {
        "number": 26436,
        "kind": "pr",
        "state": "open",
        "title": "opencl: build one program per kernel where the compiler cannot hold several (for Adreno A6x)",
        "labels": [
          "ggml",
          "OpenCL"
        ],
        "created_at": "2026-08-02T04:08:54Z",
        "updated_at": "2026-08-22T20:47:19Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26436",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26437,
        "kind": "pr",
        "state": "open",
        "title": "opencl: keep the GPU driver's submission threads off the little cores for Adreno A6x and A7x",
        "labels": [
          "ggml",
          "OpenCL"
        ],
        "created_at": "2026-08-02T04:30:18Z",
        "updated_at": "2026-08-22T20:47:19Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26437",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26438,
        "kind": "pr",
        "state": "open",
        "title": "opencl: tune the quant paths for Intel Xe-LP GPUs to improve its TG and PP performance",
        "labels": [
          "ggml",
          "OpenCL"
        ],
        "created_at": "2026-08-02T04:49:31Z",
        "updated_at": "2026-08-22T20:47:19Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26438",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26443,
        "kind": "issue",
        "state": "open",
        "title": "Regression: Cannot spawn subprocess in non-ASCII path on Windows",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-02T08:00:44Z",
        "updated_at": "2026-08-02T08:07:56Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26443",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26444,
        "kind": "pr",
        "state": "open",
        "title": "server: Fix UTF8 subprocess path encoding issue on Windows (#26443)",
        "labels": [
          "server"
        ],
        "created_at": "2026-08-02T08:31:59Z",
        "updated_at": "2026-08-02T11:10:40Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26444",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 26445,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: GLM-5.2 terrible performance on ROCm/HIP after Indexer PR #25407: prefill ~6x slower, load time ~40x longer",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-02T08:49:59Z",
        "updated_at": "2026-08-09T13:18:14Z",
        "closed_at": null,
        "comments": 9,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26445",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26446,
        "kind": "pr",
        "state": "open",
        "title": "Batch of ET backend updates",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-08-02T09:14:59Z",
        "updated_at": "2026-08-02T14:15:14Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26446",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving"
        ]
      },
      {
        "number": 26447,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Running large models on Vega 8 iGPU with Vulkan errors `vk::Queue::submit: ErrorDeviceLost` after ~ 50K context size",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-02T10:07:28Z",
        "updated_at": "2026-08-20T06:33:09Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26447",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26448,
        "kind": "issue",
        "state": "open",
        "title": "Feature request: run MoE expert weights from host RAM via PCIe DMA (no H2D copy) — 23GB MoE on 1.6GB VRAM, 81GB Qwen3-235B on 8GB cards",
        "labels": [],
        "created_at": "2026-08-02T10:13:43Z",
        "updated_at": "2026-08-10T17:54:42Z",
        "closed_at": null,
        "comments": 10,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26448",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "build_platform_ops"
        ]
      },
      {
        "number": 26449,
        "kind": "pr",
        "state": "open",
        "title": "tests : size quantize scratch buffers by row size",
        "labels": [
          "testing"
        ],
        "created_at": "2026-08-02T11:10:44Z",
        "updated_at": "2026-08-02T11:14:49Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26449",
        "categories": [
          "quantization",
          "correctness_testing"
        ]
      },
      {
        "number": 26451,
        "kind": "issue",
        "state": "open",
        "title": "Compile bug: RHEL8 / too low glibc cannot compile with default flags",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-02T13:17:00Z",
        "updated_at": "2026-08-05T21:34:36Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26451",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26453,
        "kind": "pr",
        "state": "open",
        "title": "docs : clarify tensor split context sizing",
        "labels": [
          "documentation"
        ],
        "created_at": "2026-08-02T14:47:17Z",
        "updated_at": "2026-08-02T14:51:39Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26453",
        "categories": [
          "memory_kv_context",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26454,
        "kind": "pr",
        "state": "open",
        "title": "ggml: add gfx90c HIP support",
        "labels": [
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-08-02T14:55:32Z",
        "updated_at": "2026-08-26T14:51:22Z",
        "closed_at": null,
        "comments": 18,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26454",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26455,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: ggml model cache and kv cache in remote rpc server",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-08-02T15:11:44Z",
        "updated_at": "2026-08-04T16:02:52Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26455",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 26456,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Cannot load models in Termux (on device built)",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-02T15:21:07Z",
        "updated_at": "2026-08-02T15:25:12Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26456",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26460,
        "kind": "pr",
        "state": "open",
        "title": "llama: re-create the KV cache when flash attention resolves to disabled (performance and layout bug)",
        "labels": [],
        "created_at": "2026-08-02T18:05:54Z",
        "updated_at": "2026-08-22T20:47:19Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26460",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26462,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: SIGSEGV running DS4F on Intel B70 (Missing SYCL f16 Conversion)",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-02T18:11:42Z",
        "updated_at": "2026-08-04T05:08:00Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26462",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 26466,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cuda: HIP replace __shfl_xor_sync with dpp instructions",
        "labels": [
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-08-02T18:54:06Z",
        "updated_at": "2026-08-25T21:37:08Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26466",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26467,
        "kind": "pr",
        "state": "open",
        "title": "Instella moe",
        "labels": [
          "model",
          "conversion"
        ],
        "created_at": "2026-08-02T19:17:11Z",
        "updated_at": "2026-08-21T16:06:17Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26467",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26468,
        "kind": "pr",
        "state": "open",
        "title": "ggml : fuse soft_max sweeps into fewer passes",
        "labels": [
          "testing",
          "ggml"
        ],
        "created_at": "2026-08-02T19:17:43Z",
        "updated_at": "2026-08-10T23:18:36Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26468",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26470,
        "kind": "issue",
        "state": "open",
        "title": "Metal: Gemma 4 E4B Q8_0 decode regression (~13%) between b9730 and b10219; qwen models unaffected",
        "labels": [],
        "created_at": "2026-08-02T20:28:01Z",
        "updated_at": "2026-08-05T17:18:06Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26470",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26472,
        "kind": "pr",
        "state": "open",
        "title": "Fix issue where tool arguments can appear in any JSON-valid order",
        "labels": [
          "testing"
        ],
        "created_at": "2026-08-02T20:40:51Z",
        "updated_at": "2026-08-25T23:47:41Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26472",
        "categories": [
          "models_architectures",
          "structured_generation_tools",
          "correctness_testing"
        ]
      },
      {
        "number": 26473,
        "kind": "pr",
        "state": "open",
        "title": "Preserve omitted assistant content during OpenAI-compatible message round-tripping",
        "labels": [
          "testing"
        ],
        "created_at": "2026-08-02T20:45:38Z",
        "updated_at": "2026-08-02T20:49:39Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26473",
        "categories": [
          "models_architectures",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing"
        ]
      },
      {
        "number": 26475,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: using -devd CUDA0 on draft-dspark causes model to crash",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-02T20:50:12Z",
        "updated_at": "2026-08-11T16:24:35Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26475",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "speculative_generation",
          "correctness_testing"
        ]
      },
      {
        "number": 26477,
        "kind": "pr",
        "state": "open",
        "title": "opencl: quant lm_head / decode GEMV and medium-batch GEMM optimizations (speculative decoding/MTP)",
        "labels": [
          "ggml",
          "OpenCL"
        ],
        "created_at": "2026-08-02T21:30:49Z",
        "updated_at": "2026-08-22T20:47:19Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26477",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26478,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: lllama-spec failure at 16k boundary due to non-consecutive KV cache position tracking (Y != X + 1)",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-02T21:43:51Z",
        "updated_at": "2026-08-10T06:19:27Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26478",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing"
        ]
      },
      {
        "number": 26479,
        "kind": "pr",
        "state": "open",
        "title": "Add Vulkan SDK setup and caching to ubuntu release workflow",
        "labels": [
          "devops"
        ],
        "created_at": "2026-08-02T22:45:30Z",
        "updated_at": "2026-08-04T08:44:51Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26479",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 26481,
        "kind": "issue",
        "state": "open",
        "title": "Qwen3-Coder-Next (Gated Delta Net) load crash on RTX 5090 CUDA: shared object initialization failed - fixed by downgrading llama.cpp CUDA build",
        "labels": [],
        "created_at": "2026-08-03T00:52:01Z",
        "updated_at": "2026-08-03T00:52:01Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26481",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26482,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Add native HarmonyOS NPU (HiAI) support",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-08-03T01:17:55Z",
        "updated_at": "2026-08-03T01:17:55Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26482",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 26483,
        "kind": "pr",
        "state": "open",
        "title": "docs(server): note OpenAI client base_url for multi-model gateways",
        "labels": [
          "documentation",
          "server"
        ],
        "created_at": "2026-08-03T01:28:57Z",
        "updated_at": "2026-08-03T01:33:39Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26483",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 26484,
        "kind": "issue",
        "state": "open",
        "title": "Arm CPU backend: Effective decode bandwidth stays near 10 GB/s across quantizations on Pi 5",
        "labels": [],
        "created_at": "2026-08-03T01:59:33Z",
        "updated_at": "2026-08-22T17:11:57Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26484",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 26486,
        "kind": "pr",
        "state": "open",
        "title": "ggml : vectorize rms_norm reduce and fuse the scale write",
        "labels": [
          "testing",
          "ggml"
        ],
        "created_at": "2026-08-03T03:01:55Z",
        "updated_at": "2026-08-03T03:14:24Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26486",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26487,
        "kind": "pr",
        "state": "open",
        "title": "cuda: add GGML_CUDA_BLOCKING_SYNC to eliminate 100% CPU busy-wait on full GPU offload",
        "labels": [
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-08-03T03:08:23Z",
        "updated_at": "2026-08-03T04:23:07Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26487",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 26495,
        "kind": "pr",
        "state": "open",
        "title": "llama-bench : continue after setup errors",
        "labels": [
          "examples"
        ],
        "created_at": "2026-08-03T06:07:42Z",
        "updated_at": "2026-08-03T06:11:56Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26495",
        "categories": [
          "models_architectures",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26497,
        "kind": "issue",
        "state": "open",
        "title": "UI Bug: Configured MCP Servers don't show anymore",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-03T06:26:09Z",
        "updated_at": "2026-08-20T16:06:46Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26497",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 26499,
        "kind": "pr",
        "state": "open",
        "title": "common : add env to override n_rs_seq",
        "labels": [],
        "created_at": "2026-08-03T06:40:50Z",
        "updated_at": "2026-08-04T13:05:36Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26499",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 26500,
        "kind": "pr",
        "state": "open",
        "title": "rpc: avoid serializing buffers from other servers",
        "labels": [
          "testing",
          "examples",
          "ggml"
        ],
        "created_at": "2026-08-03T07:01:28Z",
        "updated_at": "2026-08-21T16:34:43Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26500",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 26501,
        "kind": "pr",
        "state": "open",
        "title": "hexagon: support for multi-NPU devices (IQ9, IQ10) and fully asynchronous backend",
        "labels": [
          "documentation",
          "testing",
          "ggml",
          "merge ready",
          "Hexagon"
        ],
        "created_at": "2026-08-03T07:05:26Z",
        "updated_at": "2026-08-26T16:05:31Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26501",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26504,
        "kind": "pr",
        "state": "open",
        "title": "CUDA: add support for non-contiguous tensors in CEIL op",
        "labels": [
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-08-03T07:35:33Z",
        "updated_at": "2026-08-24T03:46:42Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26504",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing"
        ]
      },
      {
        "number": 26505,
        "kind": "pr",
        "state": "open",
        "title": "Add Opt.Gear Model support",
        "labels": [
          "model",
          "conversion"
        ],
        "created_at": "2026-08-03T08:59:19Z",
        "updated_at": "2026-08-03T08:59:49Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26505",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26516,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: server: expose speculative decoding counters in /metrics endpoint",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-08-03T13:50:48Z",
        "updated_at": "2026-08-16T10:01:32Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26516",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 26527,
        "kind": "pr",
        "state": "open",
        "title": "metal : add GGML_OP_CROSS_ENTROPY_LOSS",
        "labels": [
          "testing",
          "ggml",
          "Apple Metal"
        ],
        "created_at": "2026-08-03T16:50:49Z",
        "updated_at": "2026-08-15T15:31:33Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26527",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing"
        ]
      },
      {
        "number": 26529,
        "kind": "issue",
        "state": "open",
        "title": "server: prompt-cache save aborts with RPC backend (rpc_buffer_get_tensor abort); --cache-ram 0 works around it",
        "labels": [],
        "created_at": "2026-08-03T17:28:47Z",
        "updated_at": "2026-08-03T17:28:47Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26529",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26530,
        "kind": "issue",
        "state": "open",
        "title": "Qwen3-Coder template: model fails to trigger XML tool calls with large prompts — request JSON format option",
        "labels": [],
        "created_at": "2026-08-03T17:32:56Z",
        "updated_at": "2026-08-03T19:24:28Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26530",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26535,
        "kind": "issue",
        "state": "open",
        "title": "llama-server object-form tool_choice silent fallback",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-03T20:01:20Z",
        "updated_at": "2026-08-03T20:01:20Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26535",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26539,
        "kind": "pr",
        "state": "open",
        "title": "hexagon: support for HMX flash-attention head_dim not multiple of 64",
        "labels": [
          "testing",
          "ggml",
          "Hexagon"
        ],
        "created_at": "2026-08-03T21:07:36Z",
        "updated_at": "2026-08-11T01:26:51Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26539",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26540,
        "kind": "issue",
        "state": "open",
        "title": "Streaming behavior of the WebUI",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-08-03T21:41:20Z",
        "updated_at": "2026-08-03T21:43:52Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26540",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26541,
        "kind": "issue",
        "state": "open",
        "title": "Unbuffered positional read on Windows",
        "labels": [],
        "created_at": "2026-08-03T22:11:08Z",
        "updated_at": "2026-08-03T22:11:08Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26541",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "build_platform_ops"
        ]
      },
      {
        "number": 26542,
        "kind": "pr",
        "state": "open",
        "title": "llama : add unbuffered positional read on Windows",
        "labels": [
          "testing"
        ],
        "created_at": "2026-08-03T22:12:13Z",
        "updated_at": "2026-08-09T12:51:56Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26542",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26546,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: CUDA: Out-of-bounds read in gated_delta_net_cuda on Pascal (SM61) reported by Compute Sanitizer",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-04T00:02:42Z",
        "updated_at": "2026-08-04T16:45:19Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26546",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26547,
        "kind": "issue",
        "state": "open",
        "title": "KleidiAI: SME2 dispatch is silently thread-gated on Apple Silicon (banner still says \"SME2\" even when the kernel never runs); SVE path is architecturally unreachable on 128-bit-SVE2 cores",
        "labels": [],
        "created_at": "2026-08-04T00:26:31Z",
        "updated_at": "2026-08-07T11:47:10Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26547",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 26549,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Missing Lightning Indexer and HC Fused Ops with DS4F on Intel B70 (SYCL)",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-04T04:14:21Z",
        "updated_at": "2026-08-22T23:43:11Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26549",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 26550,
        "kind": "pr",
        "state": "open",
        "title": "common/json-schema-to-grammar: tolerate unanchored patterns and large repetition bounds",
        "labels": [],
        "created_at": "2026-08-04T04:53:27Z",
        "updated_at": "2026-08-04T04:57:56Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26550",
        "categories": [
          "native_backends",
          "structured_generation_tools",
          "loader_conversion",
          "build_platform_ops"
        ]
      },
      {
        "number": 26551,
        "kind": "pr",
        "state": "open",
        "title": "Feature - Deterministic Draft Filter with pluggable SDK",
        "labels": [
          "documentation",
          "build",
          "testing",
          "server"
        ],
        "created_at": "2026-08-04T06:04:41Z",
        "updated_at": "2026-08-16T08:44:52Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26551",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 26553,
        "kind": "pr",
        "state": "open",
        "title": "dont (immedietly) promote fa-auto -> fa-on",
        "labels": [],
        "created_at": "2026-08-04T06:10:07Z",
        "updated_at": "2026-08-04T17:47:16Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26553",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing"
        ]
      },
      {
        "number": 26556,
        "kind": "pr",
        "state": "open",
        "title": "quantization: Add NVFP4 quantization to llama-quantize",
        "labels": [
          "examples"
        ],
        "created_at": "2026-08-04T07:45:31Z",
        "updated_at": "2026-08-20T07:52:52Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26556",
        "categories": [
          "quantization",
          "models_architectures",
          "api_ui_docs"
        ]
      },
      {
        "number": 26558,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: llama-server hard crash (cublasSgemm INVALID_VALUE) with --spec-type draft-mtp under KV-cache saturation",
        "labels": [],
        "created_at": "2026-08-04T08:00:11Z",
        "updated_at": "2026-08-24T10:45:25Z",
        "closed_at": null,
        "comments": 10,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26558",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 26559,
        "kind": "issue",
        "state": "open",
        "title": "`-fa auto` costs prefill throughput on plain F16 KV cache on Arm CPU (Graviton)",
        "labels": [],
        "created_at": "2026-08-04T08:32:49Z",
        "updated_at": "2026-08-04T08:32:49Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26559",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26565,
        "kind": "issue",
        "state": "open",
        "title": "`llama-server` / `llama-cli` never call `llama_encode()` for encoder-decoder (T5/mT5) models: crashes or produces garbage on every T5-family GGUF",
        "labels": [],
        "created_at": "2026-08-04T10:27:04Z",
        "updated_at": "2026-08-04T10:27:04Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26565",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 26575,
        "kind": "pr",
        "state": "open",
        "title": "spec: respect safe draft caps before block decode",
        "labels": [],
        "created_at": "2026-08-04T13:12:54Z",
        "updated_at": "2026-08-04T15:29:11Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26575",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "build_platform_ops"
        ]
      },
      {
        "number": 26578,
        "kind": "pr",
        "state": "open",
        "title": "vulkan: add DeepSeek-V4 hyper-connection fused ops (DSV4_HC_COMB/PRE/POST)",
        "labels": [
          "testing",
          "Vulkan",
          "ggml"
        ],
        "created_at": "2026-08-04T14:01:15Z",
        "updated_at": "2026-08-21T15:50:41Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26578",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26580,
        "kind": "pr",
        "state": "open",
        "title": "server: use base64::decode with lenient wrapper",
        "labels": [
          "server"
        ],
        "created_at": "2026-08-04T14:37:24Z",
        "updated_at": "2026-08-04T14:42:33Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26580",
        "categories": [
          "scheduling_serving"
        ]
      },
      {
        "number": 26581,
        "kind": "issue",
        "state": "open",
        "title": "Performance: decode attention is memory-latency-bound on Intel Xe2 (Arc Pro B70) — constant ~21-25 ns per KV position per layer, identical on Vulkan and SYCL",
        "labels": [],
        "created_at": "2026-08-04T14:38:47Z",
        "updated_at": "2026-08-20T03:33:19Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26581",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 26583,
        "kind": "issue",
        "state": "open",
        "title": "RPC: GLM-5.2 crashes on multi-node CUDA RPC - invalid data ptr / graph_compute failed",
        "labels": [],
        "created_at": "2026-08-04T14:54:56Z",
        "updated_at": "2026-08-21T12:54:05Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26583",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26584,
        "kind": "pr",
        "state": "open",
        "title": "quant : allow quantization of ffn_gate_inp tensors",
        "labels": [],
        "created_at": "2026-08-04T15:14:28Z",
        "updated_at": "2026-08-04T16:34:46Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26584",
        "categories": [
          "quantization",
          "models_architectures"
        ]
      },
      {
        "number": 26587,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: convrot for int8 and fp4",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-08-04T16:06:08Z",
        "updated_at": "2026-08-04T16:06:08Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26587",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 26591,
        "kind": "issue",
        "state": "open",
        "title": "Compile bug: Error building with with CUDA 11.8: cudaGridDependencySynchronize undefined in common.cuh",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-04T17:10:35Z",
        "updated_at": "2026-08-04T17:17:33Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26591",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26592,
        "kind": "pr",
        "state": "open",
        "title": "CUDA: enable the CUB path on HIP via hipCUB",
        "labels": [
          "testing",
          "devops",
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-08-04T17:12:57Z",
        "updated_at": "2026-08-26T08:37:20Z",
        "closed_at": null,
        "comments": 49,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26592",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26595,
        "kind": "pr",
        "state": "open",
        "title": "vulkan: keep p021/nc mat-vec paths for large row counts (chunked dispatch), add GQA sharing to nc",
        "labels": [
          "Vulkan",
          "ggml"
        ],
        "created_at": "2026-08-04T17:55:22Z",
        "updated_at": "2026-08-17T12:49:45Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26595",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26599,
        "kind": "pr",
        "state": "open",
        "title": "CUDA: add f16 support for OUT_PROD",
        "labels": [
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-08-04T18:42:05Z",
        "updated_at": "2026-08-15T08:10:58Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26599",
        "categories": [
          "native_backends",
          "quantization",
          "loader_conversion",
          "correctness_testing"
        ]
      },
      {
        "number": 26600,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Issue draft — \"llama-server: unauthenticated POST /completion with deeply nested grammar crashes the process (SIGSEGV)",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-04T19:15:47Z",
        "updated_at": "2026-08-04T19:15:47Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26600",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26602,
        "kind": "pr",
        "state": "open",
        "title": "grammar : cap parser nesting and rule-chain depth to prevent stack exhaustion",
        "labels": [
          "testing"
        ],
        "created_at": "2026-08-04T19:23:52Z",
        "updated_at": "2026-08-04T19:28:54Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26602",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 26603,
        "kind": "pr",
        "state": "open",
        "title": "server : add POST /tts endpoint for TTS models",
        "labels": [
          "documentation",
          "server",
          "mtmd"
        ],
        "created_at": "2026-08-04T20:18:05Z",
        "updated_at": "2026-08-18T22:33:25Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26603",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26609,
        "kind": "issue",
        "state": "open",
        "title": "CUDA illegal memory access in cudaStreamSynchronize (flash-attn path) with Qwen3.6-35B MoE + partial expert offload — deterministic, cross-build (b10107, b10243); disappears with -fa off",
        "labels": [],
        "created_at": "2026-08-05T00:47:22Z",
        "updated_at": "2026-08-19T03:59:20Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26609",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 26610,
        "kind": "pr",
        "state": "open",
        "title": "RPC: add `-sm tensor`",
        "labels": [
          "documentation",
          "model",
          "build",
          "testing",
          "Vulkan",
          "examples",
          "devops",
          "server",
          "ggml",
          "SYCL",
          "Apple Metal",
          "Ascend NPU",
          "OpenCL",
          "IBM zDNN",
          "Hexagon",
          "mtmd",
          "CUDA",
          "AMD ZenDNN",
          "OpenVINO",
          "WebGPU",
          "server/ui",
          "conversion",
          "vendor"
        ],
        "created_at": "2026-08-05T04:59:29Z",
        "updated_at": "2026-08-20T17:35:10Z",
        "closed_at": null,
        "comments": 31,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26610",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26616,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: An opt-in flag on `llama-quantize` to throw error on \"required fallback quantization\" vs logged warning",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-08-05T06:25:29Z",
        "updated_at": "2026-08-05T06:25:29Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26616",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26617,
        "kind": "pr",
        "state": "open",
        "title": "llama: add forward-telescoping rollback engine for recurrent models (prep for #22746)",
        "labels": [
          "testing"
        ],
        "created_at": "2026-08-05T07:16:01Z",
        "updated_at": "2026-08-15T14:32:40Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26617",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26621,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cuda: pad matmul weight rows to avoid cache-set aliasing on RDNA3.5",
        "labels": [
          "testing",
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-08-05T08:26:58Z",
        "updated_at": "2026-08-08T05:34:51Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26621",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 26622,
        "kind": "pr",
        "state": "open",
        "title": "llama : add --n-cpu-ffn option",
        "labels": [],
        "created_at": "2026-08-05T08:27:21Z",
        "updated_at": "2026-08-26T15:38:00Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26622",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26625,
        "kind": "pr",
        "state": "open",
        "title": "llama-context : report graph inputs and input tensors during sched reserve",
        "labels": [],
        "created_at": "2026-08-05T09:03:58Z",
        "updated_at": "2026-08-13T13:57:59Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26625",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26628,
        "kind": "pr",
        "state": "open",
        "title": "Uses a fake simulated GPU that only outputs zero tensors",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-08-05T09:38:28Z",
        "updated_at": "2026-08-06T01:21:58Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26628",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26630,
        "kind": "issue",
        "state": "open",
        "title": "Build: `-DGGML_CPU_KLEIDIAI=ON`'s documented build line silently compiles zero `kai_run_matmul` kernels on gcc 13.3 + Cortex-X925 (banner still reports `KLEIDIAI = 1`); cost grows with model size",
        "labels": [],
        "created_at": "2026-08-05T10:44:11Z",
        "updated_at": "2026-08-05T10:44:11Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26630",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26631,
        "kind": "pr",
        "state": "open",
        "title": "ggml : allow MUL_MAT_ID -1 index to skip the computation",
        "labels": [
          "testing",
          "Vulkan",
          "ggml",
          "SYCL",
          "Apple Metal",
          "OpenCL",
          "Hexagon",
          "CUDA",
          "AMD ZenDNN",
          "WebGPU"
        ],
        "created_at": "2026-08-05T11:05:41Z",
        "updated_at": "2026-08-26T16:47:50Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26631",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26636,
        "kind": "pr",
        "state": "open",
        "title": "llama : add backends of the other model to the context",
        "labels": [],
        "created_at": "2026-08-05T13:26:41Z",
        "updated_at": "2026-08-18T06:19:10Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26636",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26637,
        "kind": "pr",
        "state": "open",
        "title": "fix: persist resolved ctx size to params_base on main model load",
        "labels": [],
        "created_at": "2026-08-05T13:50:33Z",
        "updated_at": "2026-08-05T13:50:34Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26637",
        "categories": [
          "models_architectures",
          "speculative_generation",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 26642,
        "kind": "pr",
        "state": "open",
        "title": "Support i32, i16, and bf16 for GGML_OP_REPEAT on CUDA",
        "labels": [
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-08-05T14:53:30Z",
        "updated_at": "2026-08-07T16:54:52Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26642",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26648,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: llama-sampler.cpp:1100: Assertion 'found' failed with --spec-type draft-mtp on long context",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-05T20:31:19Z",
        "updated_at": "2026-08-05T22:22:50Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26648",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26651,
        "kind": "pr",
        "state": "open",
        "title": "llama : fix dist sampler crash on degenerate logits",
        "labels": [
          "testing"
        ],
        "created_at": "2026-08-05T21:23:25Z",
        "updated_at": "2026-08-26T13:10:11Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26651",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26654,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: DeepSeek-V4 (deepseek4) fails to start on CUDA with the default -np -1 (auto) — dsv4 KV cache forces per-sequence streams while kv_unified leaves n_ctx undivided",
        "labels": [],
        "created_at": "2026-08-05T22:39:00Z",
        "updated_at": "2026-08-07T02:05:36Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26654",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26657,
        "kind": "pr",
        "state": "open",
        "title": "test: add duplicated index case for ggml_mul_mat_id",
        "labels": [
          "testing"
        ],
        "created_at": "2026-08-05T23:17:03Z",
        "updated_at": "2026-08-05T23:24:44Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26657",
        "categories": [
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26659,
        "kind": "pr",
        "state": "open",
        "title": "llama : wire up ggml_backend_register_host_buffer for mmap'd weights",
        "labels": [],
        "created_at": "2026-08-06T00:45:58Z",
        "updated_at": "2026-08-06T00:50:48Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26659",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26661,
        "kind": "pr",
        "state": "open",
        "title": "Two more minor supply-chain hardening improvements",
        "labels": [
          "devops",
          "server/ui"
        ],
        "created_at": "2026-08-06T02:06:54Z",
        "updated_at": "2026-08-09T13:39:09Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26661",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26663,
        "kind": "issue",
        "state": "open",
        "title": "[Vulkan] Pathological token-generation slowdown on RX 9070 XT (gfx1201) for models with hidden_size >= 4096 — 5-7x slower than HIP backend, ~100 GB/s effective bandwidth",
        "labels": [],
        "created_at": "2026-08-06T04:52:47Z",
        "updated_at": "2026-08-07T08:09:47Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26663",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 26669,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Support for Laguna S 2.1 DFlash",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-08-06T08:16:12Z",
        "updated_at": "2026-08-14T08:38:49Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26669",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 26670,
        "kind": "pr",
        "state": "open",
        "title": "use standard base64::decode in server-common..",
        "labels": [
          "server"
        ],
        "created_at": "2026-08-06T08:26:08Z",
        "updated_at": "2026-08-10T00:21:03Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26670",
        "categories": [
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 26671,
        "kind": "pr",
        "state": "open",
        "title": "ggml-hip : add target-derived build RPATH",
        "labels": [
          "documentation",
          "ggml"
        ],
        "created_at": "2026-08-06T08:29:21Z",
        "updated_at": "2026-08-10T06:19:17Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26671",
        "categories": [
          "native_backends",
          "models_architectures",
          "multimodal_embedding",
          "loader_conversion",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26674,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Gemma 4 tg128 performance on RTX 5060 Ti (Blackwell) appears abnormally low compared to other architectures. Is this expected?",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-06T10:59:28Z",
        "updated_at": "2026-08-14T07:46:52Z",
        "closed_at": null,
        "comments": 14,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26674",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26675,
        "kind": "pr",
        "state": "open",
        "title": "ggml : update ggml_prec specification",
        "labels": [
          "model",
          "testing",
          "ggml",
          "mtmd"
        ],
        "created_at": "2026-08-06T11:42:12Z",
        "updated_at": "2026-08-26T12:16:02Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26675",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26676,
        "kind": "issue",
        "state": "open",
        "title": "llama-server slot KV state restore is a no-op (restore reads file, slot stays empty, cache_n=0)",
        "labels": [],
        "created_at": "2026-08-06T11:45:03Z",
        "updated_at": "2026-08-07T09:36:24Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26676",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "build_platform_ops"
        ]
      },
      {
        "number": 26680,
        "kind": "issue",
        "state": "open",
        "title": "rpc-server dies silently (no log past startup banner) on ANY relayed connection (portproxy or SSH tunnel), even a zero-byte TCP probe — works fine on direct loopback",
        "labels": [],
        "created_at": "2026-08-06T15:06:33Z",
        "updated_at": "2026-08-06T15:06:33Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26680",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26682,
        "kind": "issue",
        "state": "open",
        "title": "[rocm] Eval bug: qwen3-coder-next (unsloth UD-Q6_K) at 2100 t/s by repeating tokens after first 256 times",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-06T16:41:50Z",
        "updated_at": "2026-08-06T16:41:50Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26682",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26685,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: DeepSeek V4 garbled output with RPC Vulkan.",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-06T18:09:02Z",
        "updated_at": "2026-08-09T21:04:42Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26685",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26686,
        "kind": "pr",
        "state": "open",
        "title": "Vulkan: add hoisting support for row IDs and expert count in shaders",
        "labels": [
          "Vulkan",
          "ggml"
        ],
        "created_at": "2026-08-06T18:33:54Z",
        "updated_at": "2026-08-26T19:12:27Z",
        "closed_at": null,
        "comments": 11,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26686",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "api_ui_docs"
        ]
      },
      {
        "number": 26689,
        "kind": "pr",
        "state": "open",
        "title": "SYCL: TILE for quantized KV decode",
        "labels": [
          "documentation",
          "ggml",
          "SYCL"
        ],
        "created_at": "2026-08-06T19:52:28Z",
        "updated_at": "2026-08-25T22:43:58Z",
        "closed_at": null,
        "comments": 21,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26689",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26694,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: DeepSeek-V4-Flash degenerates into repetition and leaks special tokens in long agentic chats (Metal, b10289)",
        "labels": [],
        "created_at": "2026-08-07T03:19:47Z",
        "updated_at": "2026-08-14T04:23:01Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26694",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26702,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: ROCm gfx1031 build report (why ROCM is better than vulkan)",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-07T07:36:17Z",
        "updated_at": "2026-08-26T00:42:33Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26702",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26704,
        "kind": "pr",
        "state": "open",
        "title": "CUDA: Add experimental SM120 CUTLASS MoE prefill for MXFP4 and NVFP4",
        "labels": [
          "documentation",
          "model",
          "testing",
          "ggml",
          "CUDA",
          "conversion"
        ],
        "created_at": "2026-08-07T08:08:11Z",
        "updated_at": "2026-08-24T01:20:34Z",
        "closed_at": null,
        "comments": 25,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26704",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26705,
        "kind": "pr",
        "state": "open",
        "title": "CUDA: Update Q4_K and Q5_K to use branchless computation, which stops the scale unpack being re-executed for every column in mmvq, improving perf at batch sizes > 1",
        "labels": [
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-08-07T09:07:47Z",
        "updated_at": "2026-08-26T08:49:16Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26705",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26719,
        "kind": "pr",
        "state": "open",
        "title": "mmap: Reduce windows FormatMessageA code duplication",
        "labels": [],
        "created_at": "2026-08-07T11:56:00Z",
        "updated_at": "2026-08-07T11:58:24Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26719",
        "categories": [
          "models_architectures",
          "build_platform_ops"
        ]
      },
      {
        "number": 26721,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Full width chat layout option in the WebUI",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-08-07T12:10:32Z",
        "updated_at": "2026-08-07T12:21:37Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26721",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 26722,
        "kind": "pr",
        "state": "open",
        "title": "Expose memory breakdown entries through C API",
        "labels": [],
        "created_at": "2026-08-07T12:11:23Z",
        "updated_at": "2026-08-10T03:17:46Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26722",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26723,
        "kind": "pr",
        "state": "open",
        "title": "ui: add dynamic switchable full-width chat layout option (#26721)",
        "labels": [
          "server/ui"
        ],
        "created_at": "2026-08-07T12:19:15Z",
        "updated_at": "2026-08-07T12:23:48Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26723",
        "categories": [
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 26724,
        "kind": "pr",
        "state": "open",
        "title": "rpc : do not abort the process when the remote server fails",
        "labels": [
          "testing",
          "ggml"
        ],
        "created_at": "2026-08-07T13:43:45Z",
        "updated_at": "2026-08-08T16:58:32Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26724",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26730,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Qwen3-VL-235B-A22B (qwen3vlmoe) crashes on RPC during warmup graph_compute (ggml_backend_rpc_buffer_get_tensor abort)",
        "labels": [],
        "created_at": "2026-08-07T15:01:52Z",
        "updated_at": "2026-08-07T15:02:08Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26730",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26738,
        "kind": "issue",
        "state": "open",
        "title": "Page fault at depth with -fa 0 and MoE expert offload on HIP",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-07T19:12:52Z",
        "updated_at": "2026-08-07T19:12:52Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26738",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26740,
        "kind": "pr",
        "state": "open",
        "title": "Consistify polling frequency of llama-bench with args",
        "labels": [
          "examples"
        ],
        "created_at": "2026-08-07T20:01:08Z",
        "updated_at": "2026-08-07T20:03:01Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26740",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26742,
        "kind": "pr",
        "state": "open",
        "title": "model-saver : emit moe latent size for nemotron-h",
        "labels": [
          "testing"
        ],
        "created_at": "2026-08-07T20:18:51Z",
        "updated_at": "2026-08-07T22:14:29Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26742",
        "categories": [
          "quantization",
          "models_architectures",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26744,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: vulkan flash attention lets stale K/V in freed cells influence the output",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-07T20:57:41Z",
        "updated_at": "2026-08-10T03:45:21Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26744",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 26745,
        "kind": "pr",
        "state": "open",
        "title": "docs: add RDNA4 (gfx1200/gfx1201) to HIP build support",
        "labels": [
          "documentation"
        ],
        "created_at": "2026-08-07T21:12:41Z",
        "updated_at": "2026-08-13T00:54:13Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26745",
        "categories": [
          "native_backends",
          "models_architectures",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26746,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: ROCm gfx1151 RPC worker crashes in GGML_OP_TOP_K during DeepSeek V4 prefill after 4096 tokens",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-07T21:41:06Z",
        "updated_at": "2026-08-18T17:48:21Z",
        "closed_at": null,
        "comments": 9,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26746",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26747,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: SYCL: use less  VRAM",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-08-07T22:04:55Z",
        "updated_at": "2026-08-12T06:40:52Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26747",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 26749,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: qwen3 asr outputs `asr_text` which clients do not expect and cannot handle",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-08T03:07:01Z",
        "updated_at": "2026-08-08T03:07:01Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26749",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26750,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: draft-mtp acceptance rate collapses on CUDA (40.7%) vs Vulkan (~92%) — same GGUF, same build, parameter-invariant",
        "labels": [],
        "created_at": "2026-08-08T03:23:13Z",
        "updated_at": "2026-08-21T16:24:29Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26750",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26752,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Improve SYCL Host-to-Device Memory Access",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-08-08T04:57:14Z",
        "updated_at": "2026-08-09T04:26:55Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26752",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 26753,
        "kind": "pr",
        "state": "open",
        "title": "Optimized HIP Q2_0 dot-product path for gfx1201 by replacing HIP's byte permutation with native permutation instructions. ",
        "labels": [
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-08-08T05:17:51Z",
        "updated_at": "2026-08-26T08:42:59Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26753",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26755,
        "kind": "pr",
        "state": "open",
        "title": "introduce fs_validate_path_in_directory for robust path sandboxing..",
        "labels": [
          "testing",
          "server"
        ],
        "created_at": "2026-08-08T06:07:09Z",
        "updated_at": "2026-08-08T06:11:29Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26755",
        "categories": [
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 26757,
        "kind": "pr",
        "state": "open",
        "title": "model: Add support for A.X K2",
        "labels": [
          "model",
          "testing",
          "conversion"
        ],
        "created_at": "2026-08-08T07:12:19Z",
        "updated_at": "2026-08-22T18:04:17Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26757",
        "categories": [
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26758,
        "kind": "pr",
        "state": "open",
        "title": "server : honor id_slot on the Anthropic /v1/messages route",
        "labels": [
          "server"
        ],
        "created_at": "2026-08-08T07:40:47Z",
        "updated_at": "2026-08-10T00:18:25Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26758",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26759,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: (ggml-hexagon) FLASH_ATTN_EXT produces nondeterministic wrong results on the HMX path (v75/SM8650)",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-08T07:43:34Z",
        "updated_at": "2026-08-08T07:49:55Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26759",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26761,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: DSpark/DFlash drafter with no_vocab tokenizer ignores tokenizer.ggml.mask_token_id → \"invalid token = -1\" / llama_decode fails",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-08T08:32:31Z",
        "updated_at": "2026-08-08T08:32:31Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26761",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26765,
        "kind": "issue",
        "state": "open",
        "title": "qwen35: TENSOR_SKIP on MTP block triggers \"unused tensor\" warnings for regular attention/FFN weights",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-08T10:54:56Z",
        "updated_at": "2026-08-08T15:25:13Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26765",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26766,
        "kind": "issue",
        "state": "open",
        "title": "Add support for Maple architecture (20B-A1B ternary MoE)",
        "labels": [],
        "created_at": "2026-08-08T11:18:45Z",
        "updated_at": "2026-08-08T11:18:45Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26766",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26771,
        "kind": "pr",
        "state": "open",
        "title": "CUDA: Recapture graph after memory pool flush",
        "labels": [
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-08-08T16:31:36Z",
        "updated_at": "2026-08-09T10:10:37Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26771",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26776,
        "kind": "issue",
        "state": "open",
        "title": "Huge Q6_K Performance Boost on Non-AVX CPUs with SSE4.1",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-08-08T19:13:45Z",
        "updated_at": "2026-08-08T19:19:25Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26776",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26777,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: DSV4 state_read() clears the compressed KV cache of all sequences, silently desyncing other slots",
        "labels": [],
        "created_at": "2026-08-08T19:32:40Z",
        "updated_at": "2026-08-09T12:23:29Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26777",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26781,
        "kind": "issue",
        "state": "open",
        "title": "chat: enable_thinking forced true in jinja/caps.cpp capability probe leaks into non-DeepSeek templates (regression from #26398)",
        "labels": [],
        "created_at": "2026-08-09T01:47:28Z",
        "updated_at": "2026-08-22T17:14:15Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26781",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26782,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Running gemma 4 12b with draft-mtp causes memory access fault",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-09T02:19:29Z",
        "updated_at": "2026-08-09T02:21:08Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26782",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26786,
        "kind": "pr",
        "state": "open",
        "title": "sycl: fix conv2d with a non-contiguous input",
        "labels": [
          "ggml",
          "SYCL"
        ],
        "created_at": "2026-08-09T03:39:22Z",
        "updated_at": "2026-08-11T02:18:12Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26786",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26795,
        "kind": "issue",
        "state": "open",
        "title": "Vulkan: Qwen3.6-35B-A3B (GATED_DELTA_NET) decode collapses to ~4 t/s on RDNA4 (gfx1201/RADV) while prefill runs at ~3,200 t/s",
        "labels": [],
        "created_at": "2026-08-09T08:52:10Z",
        "updated_at": "2026-08-11T00:10:19Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26795",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26809,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Downloading a new model should be possible even if it would exceed models-max",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-09T14:11:11Z",
        "updated_at": "2026-08-24T04:15:11Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26809",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26812,
        "kind": "pr",
        "state": "open",
        "title": "cuda : split argmax over multiple tiles for large rows",
        "labels": [
          "testing",
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-08-09T14:52:04Z",
        "updated_at": "2026-08-25T15:23:16Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26812",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26817,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: temp-0 tool-calling is nondeterministic for hybrid Qwen3.6-A3B on Vulkan — flips with prompt-cache mode and across restarts",
        "labels": [],
        "created_at": "2026-08-09T20:34:31Z",
        "updated_at": "2026-08-09T22:34:42Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26817",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26822,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: second llama serve silently waits on an occupied port instead of failing",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-10T01:14:05Z",
        "updated_at": "2026-08-11T12:04:42Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26822",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26825,
        "kind": "issue",
        "state": "open",
        "title": "[RPC] GET_ROWS can read past an input tensor buffer in release builds",
        "labels": [],
        "created_at": "2026-08-10T03:24:31Z",
        "updated_at": "2026-08-10T03:25:42Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26825",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26827,
        "kind": "pr",
        "state": "open",
        "title": "fix(mtp): serialize multi-ubatch decode execution",
        "labels": [
          "testing"
        ],
        "created_at": "2026-08-10T04:49:39Z",
        "updated_at": "2026-08-26T14:56:41Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26827",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26829,
        "kind": "pr",
        "state": "open",
        "title": "vulkan: Intel vulkan kernel optimizations for LLM inference.",
        "labels": [
          "testing",
          "Vulkan",
          "ggml"
        ],
        "created_at": "2026-08-10T05:22:47Z",
        "updated_at": "2026-08-10T14:57:29Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26829",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "speculative_generation",
          "correctness_testing"
        ]
      },
      {
        "number": 26830,
        "kind": "pr",
        "state": "open",
        "title": "common : assert tensor existence when loading control vectors",
        "labels": [],
        "created_at": "2026-08-10T05:50:36Z",
        "updated_at": "2026-08-10T05:55:02Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26830",
        "categories": [
          "native_backends",
          "quantization",
          "loader_conversion",
          "correctness_testing"
        ]
      },
      {
        "number": 26833,
        "kind": "pr",
        "state": "open",
        "title": "fix qwen tool parser anyOf object args",
        "labels": [
          "testing"
        ],
        "created_at": "2026-08-10T06:55:41Z",
        "updated_at": "2026-08-11T13:11:53Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26833",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26834,
        "kind": "pr",
        "state": "open",
        "title": "Update build.md: new Vulkan MSYS2 instruction step and warnings",
        "labels": [
          "documentation"
        ],
        "created_at": "2026-08-10T08:26:24Z",
        "updated_at": "2026-08-10T08:28:56Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26834",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26837,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: 3 GPU with tensor crashes",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-10T08:52:43Z",
        "updated_at": "2026-08-22T17:16:52Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26837",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26845,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: SYCL garbage on the second prompt.",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-10T12:15:06Z",
        "updated_at": "2026-08-22T16:08:23Z",
        "closed_at": null,
        "comments": 10,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26845",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 26850,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Add dependency cooldown for pip packages",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-08-10T15:44:45Z",
        "updated_at": "2026-08-10T15:44:45Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26850",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26851,
        "kind": "pr",
        "state": "open",
        "title": "ci: set pip dependency cooldown to protect against supply-chain attacks",
        "labels": [
          "devops"
        ],
        "created_at": "2026-08-10T15:57:28Z",
        "updated_at": "2026-08-10T15:57:50Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26851",
        "categories": [
          "native_backends",
          "models_architectures",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26853,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: [Vulkan] GGML_ASSERT failed in GET_ROWS with view offset (Qwen3-TTS)",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-10T16:25:04Z",
        "updated_at": "2026-08-10T16:25:04Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26853",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26854,
        "kind": "pr",
        "state": "open",
        "title": "vulkan: fall back to CPU for GET_ROWS with misaligned offsets",
        "labels": [
          "testing",
          "Vulkan",
          "ggml"
        ],
        "created_at": "2026-08-10T16:44:59Z",
        "updated_at": "2026-08-19T03:34:52Z",
        "closed_at": null,
        "comments": 9,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26854",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26855,
        "kind": "pr",
        "state": "open",
        "title": "docs: document CUDA 12.4/13.1 runtime mismatch crash in mul_mat_q",
        "labels": [
          "documentation"
        ],
        "created_at": "2026-08-10T16:54:24Z",
        "updated_at": "2026-08-10T16:59:22Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26855",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26857,
        "kind": "pr",
        "state": "open",
        "title": "metal: fix SIGABRT in ggml_metal_rsets_free during process exit",
        "labels": [
          "ggml",
          "Apple Metal"
        ],
        "created_at": "2026-08-10T17:54:25Z",
        "updated_at": "2026-08-11T05:05:15Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26857",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26862,
        "kind": "issue",
        "state": "open",
        "title": "Restore MCP resources and prompts with faster UI response in web UI?",
        "labels": [
          "bug-unconfirmed",
          "server/ui"
        ],
        "created_at": "2026-08-10T19:38:12Z",
        "updated_at": "2026-08-11T17:17:11Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26862",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26863,
        "kind": "pr",
        "state": "open",
        "title": "SYCL: Refactor GGML_SYCL_ENABLE_MKL_FA to global var",
        "labels": [
          "documentation",
          "ggml",
          "SYCL"
        ],
        "created_at": "2026-08-10T19:40:00Z",
        "updated_at": "2026-08-21T05:20:21Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26863",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 26869,
        "kind": "pr",
        "state": "open",
        "title": "quantization: complete MXFP4 and NVFP4",
        "labels": [
          "examples"
        ],
        "created_at": "2026-08-10T21:46:01Z",
        "updated_at": "2026-08-26T21:12:13Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26869",
        "categories": [
          "quantization",
          "models_architectures",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26873,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Glimmer consumes more memory and prefill slows down after first mmproj use",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-10T23:10:32Z",
        "updated_at": "2026-08-15T11:11:04Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26873",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26875,
        "kind": "issue",
        "state": "open",
        "title": "Compile bug: KleidiAI .S kernels fail to assemble on Windows ARM64 with llvm-mingw (COFF target) - ELF-only .type/.size in asm preamble",
        "labels": [],
        "created_at": "2026-08-11T00:36:23Z",
        "updated_at": "2026-08-11T00:36:23Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26875",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26876,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Partial MoE offload (-ncmoe) on integrated GPU is ~2x slower than CPU-only (Arc 140T / Vulkan / UMA)",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-11T00:41:20Z",
        "updated_at": "2026-08-11T00:41:20Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26876",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26877,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Formal release of Windows Arm Hexagon NPU builds",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-08-11T01:02:59Z",
        "updated_at": "2026-08-11T04:01:13Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26877",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26883,
        "kind": "pr",
        "state": "open",
        "title": "speculative-simple : pass ctx_other to the draft context and fail gracefully",
        "labels": [
          "examples"
        ],
        "created_at": "2026-08-11T04:05:11Z",
        "updated_at": "2026-08-11T05:36:53Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26883",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 26884,
        "kind": "issue",
        "state": "open",
        "title": "speculative-simple: dflash drafters draft without target features (silent low acceptance)",
        "labels": [],
        "created_at": "2026-08-11T04:05:28Z",
        "updated_at": "2026-08-11T04:05:28Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26884",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 26885,
        "kind": "pr",
        "state": "open",
        "title": "[Grammars] Optimization: Single lookup, remove extra copies - 1.2x-1.3x speedup",
        "labels": [],
        "created_at": "2026-08-11T05:40:23Z",
        "updated_at": "2026-08-13T16:46:35Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26885",
        "categories": [
          "native_backends",
          "models_architectures",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26887,
        "kind": "pr",
        "state": "open",
        "title": "ci : disable OpenVINO builds",
        "labels": [
          "devops"
        ],
        "created_at": "2026-08-11T05:51:13Z",
        "updated_at": "2026-08-26T06:05:31Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26887",
        "categories": [
          "native_backends",
          "models_architectures",
          "build_platform_ops"
        ]
      },
      {
        "number": 26888,
        "kind": "pr",
        "state": "open",
        "title": "Openvino Backend Muse-Glimmer Support",
        "labels": [
          "ggml",
          "OpenVINO"
        ],
        "created_at": "2026-08-11T06:23:08Z",
        "updated_at": "2026-08-19T16:46:26Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26888",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "build_platform_ops"
        ]
      },
      {
        "number": 26893,
        "kind": "pr",
        "state": "open",
        "title": "server : avoid caching non-completion idle slots in RAM",
        "labels": [
          "server"
        ],
        "created_at": "2026-08-11T09:32:14Z",
        "updated_at": "2026-08-11T10:07:38Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26893",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing"
        ]
      },
      {
        "number": 26894,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: DFlash drafter fails to bind (vector::_M_range_check) when target GGUF encodes attention.sliding_window_pattern as an array — Muse-Glimmer-30B official GGUF",
        "labels": [],
        "created_at": "2026-08-11T09:38:14Z",
        "updated_at": "2026-08-13T15:09:17Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26894",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26895,
        "kind": "pr",
        "state": "open",
        "title": "CUDA: add POOL_1D",
        "labels": [
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-08-11T09:45:36Z",
        "updated_at": "2026-08-11T10:01:26Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26895",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing"
        ]
      },
      {
        "number": 26898,
        "kind": "pr",
        "state": "open",
        "title": "llama : add KleidiAI weight cache",
        "labels": [
          "documentation",
          "build",
          "testing",
          "examples",
          "devops",
          "ggml",
          "vendor"
        ],
        "created_at": "2026-08-11T10:38:42Z",
        "updated_at": "2026-08-17T11:13:14Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26898",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26899,
        "kind": "pr",
        "state": "open",
        "title": "Revert \"common : dedup preset and cached model entries in /v1/models (#25131)",
        "labels": [],
        "created_at": "2026-08-11T10:54:15Z",
        "updated_at": "2026-08-11T10:54:16Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26899",
        "categories": [
          "models_architectures",
          "memory_kv_context"
        ]
      },
      {
        "number": 26901,
        "kind": "issue",
        "state": "open",
        "title": "Blackwell discrete GPUs (RTX 5080/5070Ti) misclassified as iGPU — breaks multi-GPU tensor split on Linux",
        "labels": [],
        "created_at": "2026-08-11T11:40:35Z",
        "updated_at": "2026-08-11T11:40:35Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26901",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26902,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Glimmer Q8_0 on 4 x Tesla T10 tensor split: ggml-backend-meta.cpp:537: GGML_ASSERT(ret.axis != GGML_BACKEND_SPLIT_AXIS_UNKNOWN) failed",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-11T12:24:09Z",
        "updated_at": "2026-08-18T03:57:56Z",
        "closed_at": null,
        "comments": 9,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26902",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26906,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: model stop answering after some hours of work",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-11T14:51:59Z",
        "updated_at": "2026-08-11T14:51:59Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26906",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26907,
        "kind": "issue",
        "state": "open",
        "title": "Compile bug: llama-ui-assets.dir `tools/ui/ui-src/ui-src/ui-src/ui-src/...` infinite COPY loop",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-11T16:02:21Z",
        "updated_at": "2026-08-14T18:18:43Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26907",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26912,
        "kind": "issue",
        "state": "open",
        "title": "[RPC] SET_ROWS can write past an output tensor buffer in release builds",
        "labels": [],
        "created_at": "2026-08-11T16:11:36Z",
        "updated_at": "2026-08-13T12:22:12Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26912",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26913,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: EAGLE3 draft model fails to load in speculative pairing with 'invalid vector subscript' on Windows (standalone load works)",
        "labels": [],
        "created_at": "2026-08-11T16:14:40Z",
        "updated_at": "2026-08-11T16:14:40Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26913",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "speculative_generation",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26914,
        "kind": "pr",
        "state": "open",
        "title": "ggml : add native INT8 convrot support",
        "labels": [
          "testing",
          "Vulkan",
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-08-11T16:20:11Z",
        "updated_at": "2026-08-25T18:30:17Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26914",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 26916,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Qwen3.5-Hybrid model (qwen3_5, SSM+Attention) fails to load — \"tensor 'blk.32.attn_norm.weight' not found\"",
        "labels": [],
        "created_at": "2026-08-11T17:01:41Z",
        "updated_at": "2026-08-12T07:12:28Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26916",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26921,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Vulkan (Mali-G925 / Immortalis MC12): Qwen3.5-0.8B multimodal prefill returns all-NaN logits, while the same build on CPU devices works",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-11T17:33:06Z",
        "updated_at": "2026-08-23T09:13:45Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26921",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26928,
        "kind": "pr",
        "state": "open",
        "title": "web UI: Add download button for generated files in the chat",
        "labels": [
          "server/ui"
        ],
        "created_at": "2026-08-11T21:23:08Z",
        "updated_at": "2026-08-20T20:14:58Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26928",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 26932,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cuda: fix UMA memory detection for AMD iGPUs via sysfs VRAM",
        "labels": [
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-08-11T21:51:39Z",
        "updated_at": "2026-08-14T20:01:47Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26932",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 26933,
        "kind": "pr",
        "state": "open",
        "title": "ggml-rpc : fix out-of-bounds write in SET_ROWS graph execution",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-08-11T22:30:14Z",
        "updated_at": "2026-08-26T07:20:46Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26933",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving"
        ]
      },
      {
        "number": 26935,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug:  Big list of MCP Tools cause timeout at UI",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-11T23:39:11Z",
        "updated_at": "2026-08-13T06:43:18Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26935",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26937,
        "kind": "issue",
        "state": "open",
        "title": "Profile prefil(tokenizer firstly)",
        "labels": [],
        "created_at": "2026-08-12T00:01:51Z",
        "updated_at": "2026-08-13T18:07:11Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26937",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "api_ui_docs"
        ]
      },
      {
        "number": 26945,
        "kind": "issue",
        "state": "open",
        "title": "Vulkan backend crash (STATUS_STACK_BUFFER_OVERRUN) when offloading 2+ repeating layers to GPU with qwen35moe (Gated Delta Net / Lightning Indexer / DeepSeek V4 HC)",
        "labels": [],
        "created_at": "2026-08-12T05:36:49Z",
        "updated_at": "2026-08-12T05:38:32Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26945",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26948,
        "kind": "pr",
        "state": "open",
        "title": "CPU flash-attn: support quantized K/V in the tiled prefill kernel",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-08-12T06:39:00Z",
        "updated_at": "2026-08-14T05:29:34Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26948",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26949,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Metal: GGML_ASSERT(buf_src) failed in ggml_metal_buffer_set_tensor on Intel Mac with AMD discrete GPU (newBufferWithBytesNoCopy page-alignment requirement)",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-12T06:48:10Z",
        "updated_at": "2026-08-12T10:10:14Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26949",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26954,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cpu/amx: request XFD XTILEDATA permission per thread before AMX instructions",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-08-12T09:34:51Z",
        "updated_at": "2026-08-12T09:39:48Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26954",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26957,
        "kind": "issue",
        "state": "open",
        "title": "Performance regressions on Android arm64 CPU (Snapdragon 8 Gen 2): text generation ~8% slower since b10333, prompt processing ~13% slower since b10369",
        "labels": [],
        "created_at": "2026-08-12T11:12:11Z",
        "updated_at": "2026-08-13T08:28:55Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26957",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26961,
        "kind": "pr",
        "state": "open",
        "title": "ui: persistent memory for the model, stored in IndexedDB",
        "labels": [
          "server/ui"
        ],
        "created_at": "2026-08-12T12:34:15Z",
        "updated_at": "2026-08-23T07:39:01Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26961",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26963,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Pre-built ROCm Windows binary crashes with \"cudaMemGetInfo failed\"",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-12T13:01:18Z",
        "updated_at": "2026-08-17T06:21:45Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26963",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26964,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Latest Windows ROCM not using GPU",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-12T13:02:02Z",
        "updated_at": "2026-08-15T14:54:23Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26964",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 26965,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: DeepSeek V4 Flash tokenizer blows its stack on long tool output",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-12T13:04:17Z",
        "updated_at": "2026-08-12T13:12:50Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26965",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26966,
        "kind": "pr",
        "state": "open",
        "title": "unicode: stop shared pre-tokenizer regex from blowing the stack",
        "labels": [
          "testing"
        ],
        "created_at": "2026-08-12T13:09:06Z",
        "updated_at": "2026-08-13T02:39:24Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26966",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing"
        ]
      },
      {
        "number": 26967,
        "kind": "issue",
        "state": "open",
        "title": "DFlash: corrupted predicted_ms on some Q4/Metal requests",
        "labels": [],
        "created_at": "2026-08-12T13:18:47Z",
        "updated_at": "2026-08-12T13:18:47Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26967",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "loader_conversion",
          "build_platform_ops"
        ]
      },
      {
        "number": 26968,
        "kind": "pr",
        "state": "open",
        "title": "common : load the draft model from the draft model path",
        "labels": [],
        "created_at": "2026-08-12T13:20:04Z",
        "updated_at": "2026-08-13T10:40:23Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26968",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "loader_conversion"
        ]
      },
      {
        "number": 26973,
        "kind": "pr",
        "state": "open",
        "title": "ci : bundle HIP runtime DLLs with Windows ROCm release",
        "labels": [
          "need feedback",
          "devops"
        ],
        "created_at": "2026-08-12T14:02:59Z",
        "updated_at": "2026-08-26T20:21:44Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26973",
        "categories": [
          "native_backends",
          "models_architectures",
          "build_platform_ops"
        ]
      },
      {
        "number": 26975,
        "kind": "pr",
        "state": "open",
        "title": "mtmd: add JPEG XL decode via libjxl",
        "labels": [
          "mtmd",
          "server/ui"
        ],
        "created_at": "2026-08-12T14:52:54Z",
        "updated_at": "2026-08-14T18:04:13Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26975",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26977,
        "kind": "pr",
        "state": "open",
        "title": "CUDA: Improve selection of routed MoE MMQ tile widths",
        "labels": [
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-08-12T17:06:04Z",
        "updated_at": "2026-08-12T17:56:53Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26977",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 26978,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: GGUF loader accepts a tensor size that wraps to 0 after padding",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-12T17:20:25Z",
        "updated_at": "2026-08-12T17:20:25Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26978",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26979,
        "kind": "pr",
        "state": "open",
        "title": "gguf : reject tensor size that wraps after padding",
        "labels": [
          "testing",
          "ggml"
        ],
        "created_at": "2026-08-12T17:33:10Z",
        "updated_at": "2026-08-17T12:26:41Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26979",
        "categories": [
          "quantization",
          "models_architectures",
          "loader_conversion",
          "correctness_testing"
        ]
      },
      {
        "number": 26981,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: gemma4uv mmproj + CUDA → SIGABRT in mtmd_helper_decode_image_chunk → llama_context::decode (workarounds from #24251 and #24314 do not help)",
        "labels": [],
        "created_at": "2026-08-12T19:12:31Z",
        "updated_at": "2026-08-12T19:12:31Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26981",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26982,
        "kind": "issue",
        "state": "open",
        "title": "HIP: check why the pertubation from -funsafe-math-optimizations is so large",
        "labels": [],
        "created_at": "2026-08-12T19:30:28Z",
        "updated_at": "2026-08-12T19:30:32Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26982",
        "categories": [
          "native_backends",
          "models_architectures"
        ]
      },
      {
        "number": 26985,
        "kind": "issue",
        "state": "open",
        "title": "Bump ggml-org/ccache-action to upstream's v1.2.23",
        "labels": [],
        "created_at": "2026-08-12T21:02:59Z",
        "updated_at": "2026-08-13T12:12:43Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26985",
        "categories": [
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26987,
        "kind": "issue",
        "state": "open",
        "title": "Qwen3-Coder parser: lazy tool-call trigger never fires when model skips both <tool_call> and <function=",
        "labels": [],
        "created_at": "2026-08-12T21:25:16Z",
        "updated_at": "2026-08-17T03:10:39Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26987",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 26988,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: --cors-origins does not follow spec",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-13T00:48:25Z",
        "updated_at": "2026-08-13T00:48:25Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26988",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26989,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Add NVFP4 quantization support for KV cache (--cache-type-k / --cache-type-v)",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-08-13T02:48:00Z",
        "updated_at": "2026-08-13T02:48:00Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26989",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 26992,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug:",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-13T05:38:43Z",
        "updated_at": "2026-08-13T05:38:43Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26992",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26994,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Image prompt cache never hits for hybrid SSM models (Qwen3.6) when only the text suffix changes",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-13T06:27:24Z",
        "updated_at": "2026-08-13T06:37:35Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26994",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 26995,
        "kind": "pr",
        "state": "open",
        "title": "api: expose selected layer input pooling",
        "labels": [],
        "created_at": "2026-08-13T06:46:35Z",
        "updated_at": "2026-08-13T07:04:03Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/26995",
        "categories": [
          "native_backends",
          "models_architectures",
          "multimodal_embedding",
          "loader_conversion",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26996,
        "kind": "issue",
        "state": "open",
        "title": "win-rocm-7.14 Windows release missing hipblas.dll — GPU not detected, `--list-devices` returns empty",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-13T07:29:10Z",
        "updated_at": "2026-08-18T01:45:18Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26996",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 26999,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Multi-line prompt in terminal",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-08-13T09:13:22Z",
        "updated_at": "2026-08-13T12:46:55Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/26999",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 27000,
        "kind": "pr",
        "state": "open",
        "title": "llama: add Maple 20B-A1B ternary MoE architecture (CPU)",
        "labels": [
          "model",
          "testing",
          "conversion"
        ],
        "created_at": "2026-08-13T09:28:35Z",
        "updated_at": "2026-08-16T11:56:58Z",
        "closed_at": null,
        "comments": 19,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27000",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "loader_conversion",
          "correctness_testing"
        ]
      },
      {
        "number": 27004,
        "kind": "pr",
        "state": "open",
        "title": "server: add disaggregated prefill via HTTP state handoff",
        "labels": [
          "server"
        ],
        "created_at": "2026-08-13T10:49:13Z",
        "updated_at": "2026-08-14T08:55:40Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27004",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "build_platform_ops"
        ]
      },
      {
        "number": 27007,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Gemma 4 26B A4B (QAT) full-GPU offload corrupts output on Vulkan (Radeon 890M gfx1150) - isolated to fused MMVQ kernel path",
        "labels": [],
        "created_at": "2026-08-13T11:11:04Z",
        "updated_at": "2026-08-15T15:55:39Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27007",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27009,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: CUDA graphs for multi-slot decode via shape-stable padded u",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-08-13T11:23:28Z",
        "updated_at": "2026-08-22T06:07:57Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27009",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27011,
        "kind": "issue",
        "state": "open",
        "title": "server : add Prometheus histogram for request context sizes",
        "labels": [],
        "created_at": "2026-08-13T11:49:10Z",
        "updated_at": "2026-08-13T11:49:10Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27011",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27012,
        "kind": "pr",
        "state": "open",
        "title": "server : add Prometheus histogram for request context sizes",
        "labels": [
          "documentation",
          "server"
        ],
        "created_at": "2026-08-13T11:49:49Z",
        "updated_at": "2026-08-15T07:23:40Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27012",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 27016,
        "kind": "pr",
        "state": "open",
        "title": "fix/cuda-f16-activation-scaling",
        "labels": [
          "testing",
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-08-13T13:09:37Z",
        "updated_at": "2026-08-14T01:19:34Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27016",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 27019,
        "kind": "issue",
        "state": "open",
        "title": "convert_hf_to_gguf: Qwen3.5 (qwen3_5) hybrid linear-attention tensors fail - ssm_conv1d kernel dim + in_proj_a/b expansion not handled",
        "labels": [],
        "created_at": "2026-08-13T13:49:17Z",
        "updated_at": "2026-08-15T13:20:42Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27019",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 27020,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Dense Llama 4 model crashes",
        "labels": [],
        "created_at": "2026-08-13T14:11:44Z",
        "updated_at": "2026-08-15T11:43:59Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27020",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27021,
        "kind": "issue",
        "state": "open",
        "title": "ROCm: TOP_K crashes with \"invalid configuration argument\" when ncols > 1024 (bitonic kernel block-size overflow; blocks DeepSeek V4 ctx > 128K)",
        "labels": [],
        "created_at": "2026-08-13T14:19:48Z",
        "updated_at": "2026-08-18T18:31:45Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27021",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27022,
        "kind": "issue",
        "state": "open",
        "title": "Vulkan: qwen3next MoE with IQ1_M router weights produces all-zero/garbage output (works on CPU)",
        "labels": [],
        "created_at": "2026-08-13T14:19:54Z",
        "updated_at": "2026-08-13T14:19:54Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27022",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27023,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: reasoning_effort seems broken",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-13T14:26:20Z",
        "updated_at": "2026-08-15T16:15:48Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27023",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27024,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cpu/amx: fix block_q8_K VNNI quantization and enable VNNI path",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-08-13T14:29:20Z",
        "updated_at": "2026-08-13T14:29:40Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27024",
        "categories": [
          "quantization",
          "models_architectures",
          "correctness_testing"
        ]
      },
      {
        "number": 27027,
        "kind": "pr",
        "state": "open",
        "title": "server: (router) require API key for non-GET /models requests",
        "labels": [
          "server"
        ],
        "created_at": "2026-08-13T14:51:23Z",
        "updated_at": "2026-08-14T12:59:07Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27027",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27031,
        "kind": "pr",
        "state": "open",
        "title": "fix: correct default threads in llama-bench",
        "labels": [
          "examples"
        ],
        "created_at": "2026-08-13T17:26:43Z",
        "updated_at": "2026-08-15T08:00:16Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27031",
        "categories": [
          "models_architectures",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 27032,
        "kind": "pr",
        "state": "open",
        "title": "CPU Optimizations - Prefill, Tokenization, and Token Generation",
        "labels": [
          "testing",
          "ggml"
        ],
        "created_at": "2026-08-13T18:04:30Z",
        "updated_at": "2026-08-19T15:44:52Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27032",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27033,
        "kind": "pr",
        "state": "open",
        "title": "feat(win32): optimize thread scheduling and core affinity on hybrid CPUs",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-08-13T18:07:00Z",
        "updated_at": "2026-08-15T06:51:59Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27033",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27035,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Remote server bug upper/lower case with args gpt-oss-20b-default",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-13T19:48:37Z",
        "updated_at": "2026-08-13T19:48:37Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27035",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27038,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: SYCL new host-pinned memory high CPU utilization for large allocations",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-13T20:24:28Z",
        "updated_at": "2026-08-25T08:34:19Z",
        "closed_at": null,
        "comments": 14,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27038",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 27039,
        "kind": "issue",
        "state": "open",
        "title": "Refactor: share thread pools across drafter and main model contexts",
        "labels": [],
        "created_at": "2026-08-13T21:14:24Z",
        "updated_at": "2026-08-19T15:47:02Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27039",
        "categories": [
          "models_architectures",
          "memory_kv_context"
        ]
      },
      {
        "number": 27044,
        "kind": "pr",
        "state": "open",
        "title": "CUDA: size MMQ ids-path tail padding from the flattened row count, not ne11",
        "labels": [
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-08-13T22:42:02Z",
        "updated_at": "2026-08-16T03:33:03Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27044",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27045,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Near 0% CUDA Utilization With 75% CPU Offload and 25% GPU Layers",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-14T01:26:49Z",
        "updated_at": "2026-08-14T01:28:15Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27045",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27046,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: SIGSEGV (null-ptr jump) on GPU offload — resolve_fused_ops false-positives on Intel Lunar Lake iGPU (Arc 140V), reproduces on unrelated architectures (gemma4, qwen2)",
        "labels": [],
        "created_at": "2026-08-14T01:36:25Z",
        "updated_at": "2026-08-17T10:03:35Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27046",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27050,
        "kind": "issue",
        "state": "open",
        "title": "Server: backend sampling (-bs) gives +48% throughput at 32 slots (706 -> 1046 tok/s) — should it be the default for multi-slot serving?",
        "labels": [],
        "created_at": "2026-08-14T02:30:27Z",
        "updated_at": "2026-08-14T02:43:17Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27050",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27053,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: ggml_cont on transposed block-quantized tensor silently corrupts memory on CPU (should abort if unsupported)",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-14T07:16:54Z",
        "updated_at": "2026-08-14T07:16:54Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27053",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27055,
        "kind": "issue",
        "state": "open",
        "title": "mp4 file input not possible",
        "labels": [],
        "created_at": "2026-08-14T08:35:49Z",
        "updated_at": "2026-08-14T08:35:49Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27055",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving"
        ]
      },
      {
        "number": 27058,
        "kind": "pr",
        "state": "open",
        "title": "server: add disaggregated prompt prefill over RPC",
        "labels": [
          "documentation",
          "server"
        ],
        "created_at": "2026-08-14T10:32:36Z",
        "updated_at": "2026-08-14T11:41:46Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27058",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27062,
        "kind": "pr",
        "state": "open",
        "title": "sycl: reduce redundant work in Q4_K multi-column MMVQ",
        "labels": [
          "testing",
          "ggml",
          "merge ready",
          "SYCL"
        ],
        "created_at": "2026-08-14T12:09:05Z",
        "updated_at": "2026-08-17T10:48:29Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27062",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27063,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: SYCL completely broken on A770",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-14T12:24:06Z",
        "updated_at": "2026-08-22T15:35:37Z",
        "closed_at": null,
        "comments": 17,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27063",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27065,
        "kind": "issue",
        "state": "open",
        "title": "Double-free crash in  console::history_t::~history_t()  on Termux (Android aarch64 bionic libc)",
        "labels": [],
        "created_at": "2026-08-14T12:59:09Z",
        "updated_at": "2026-08-14T12:59:09Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27065",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27066,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Adaptive P is broken on muse glimmer.",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-14T13:13:01Z",
        "updated_at": "2026-08-14T13:13:01Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27066",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 27068,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: failed slot restore leaves corrupted K/V data that breaks subsequent inference",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-14T13:59:20Z",
        "updated_at": "2026-08-14T14:01:53Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27068",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27069,
        "kind": "pr",
        "state": "open",
        "title": "ggml-webgpu: no dawn native features on wasi",
        "labels": [
          "ggml",
          "WebGPU"
        ],
        "created_at": "2026-08-14T15:32:26Z",
        "updated_at": "2026-08-16T23:39:12Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27069",
        "categories": [
          "native_backends",
          "models_architectures"
        ]
      },
      {
        "number": 27072,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: llamacpp crashes on load on Linux using Vulkan.",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-14T15:45:02Z",
        "updated_at": "2026-08-16T22:07:35Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27072",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27076,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: ggml_vulkan: device lost on Vulkan0",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-14T16:31:40Z",
        "updated_at": "2026-08-18T06:15:49Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27076",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27078,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: KleidiAI dotprod GEMV ~3x slower than CPU repack for token generation on Apple M1 Max (Q4_0), while winning on Neoverse N1/N2 — no runtime escape when both compiled in",
        "labels": [],
        "created_at": "2026-08-14T17:01:34Z",
        "updated_at": "2026-08-14T17:01:34Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27078",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27079,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Server Error dry_penalty_last_n using Vulkan & ROCm images",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-14T17:24:52Z",
        "updated_at": "2026-08-19T15:28:08Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27079",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27086,
        "kind": "issue",
        "state": "open",
        "title": "Flash attention default (-fa auto) costs ~2x prefill throughput on Arm Neoverse V-series CPUs",
        "labels": [],
        "created_at": "2026-08-14T20:33:16Z",
        "updated_at": "2026-08-14T21:27:51Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27086",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 27089,
        "kind": "issue",
        "state": "open",
        "title": "Library (in-process) hosts cannot use speculative decoding / DSpark — no core C-API for draft/spec params",
        "labels": [],
        "created_at": "2026-08-14T21:42:00Z",
        "updated_at": "2026-08-14T21:42:00Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27089",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27090,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Qwen3.8-27B (qwen35 hybrid) llama-server crashes silently at ~520K prefill tokens with YaRN rope-scale 4 (just under 2x yarn-orig-ctx), b10434 CUDA",
        "labels": [],
        "created_at": "2026-08-14T21:55:30Z",
        "updated_at": "2026-08-14T21:55:30Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27090",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 27093,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: very tiny LoRA causes severe performance degradation",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-14T23:43:33Z",
        "updated_at": "2026-08-14T23:43:33Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27093",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 27094,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: llama-server cannot connect To RPC node even though llama-cli works",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-15T00:01:12Z",
        "updated_at": "2026-08-16T23:43:52Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27094",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27096,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cpu : fix soft_max_back wrong output when dst aliases src1",
        "labels": [
          "testing",
          "ggml"
        ],
        "created_at": "2026-08-15T01:04:10Z",
        "updated_at": "2026-08-24T15:10:11Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27096",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing"
        ]
      },
      {
        "number": 27097,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Vulkan - slow token generation with AMD DGPU Resizable BAR disabled",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-15T01:39:16Z",
        "updated_at": "2026-08-22T19:30:03Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27097",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27098,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request:",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-08-15T01:40:49Z",
        "updated_at": "2026-08-17T16:55:54Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27098",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 27102,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: CUDA kernel stall during model execution, killed by watchdog",
        "labels": [
          "bug",
          "help wanted",
          "Nvidia GPU"
        ],
        "created_at": "2026-08-15T02:40:06Z",
        "updated_at": "2026-08-26T09:20:15Z",
        "closed_at": null,
        "comments": 30,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27102",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27105,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: MTP (--spec-type draft-mtp) causes GPU MMU page fault (Xid 31) on Maxwell (sm_52) — Tesla M40",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-15T03:24:55Z",
        "updated_at": "2026-08-15T03:24:55Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27105",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27108,
        "kind": "pr",
        "state": "open",
        "title": "server: fix case-sensitive HF file lookup breaking --gpt-oss-20b-default",
        "labels": [],
        "created_at": "2026-08-15T04:14:56Z",
        "updated_at": "2026-08-15T04:27:51Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27108",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving"
        ]
      },
      {
        "number": 27109,
        "kind": "issue",
        "state": "open",
        "title": "CUDA: 4-bit KV cache (q4_1/q4_0) collapses prefill to ~34 t/s on qwen35 hybrid (RTX 3090, MMQ guard passes)",
        "labels": [],
        "created_at": "2026-08-15T04:30:09Z",
        "updated_at": "2026-08-17T08:04:48Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27109",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "build_platform_ops"
        ]
      },
      {
        "number": 27110,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug:reproduces with both internal AllReduce and NCCL",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-15T05:01:18Z",
        "updated_at": "2026-08-15T12:22:06Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27110",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27112,
        "kind": "issue",
        "state": "open",
        "title": "Proposal: public API for hidden-state extraction (residual stream) + tools",
        "labels": [],
        "created_at": "2026-08-15T07:49:09Z",
        "updated_at": "2026-08-15T07:49:09Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27112",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "api_ui_docs"
        ]
      },
      {
        "number": 27113,
        "kind": "pr",
        "state": "open",
        "title": "llama: relax `equal_seqs` assert in `llm_graph_input_attn_cross::set_input`",
        "labels": [],
        "created_at": "2026-08-15T07:50:21Z",
        "updated_at": "2026-08-16T02:57:29Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27113",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27116,
        "kind": "issue",
        "state": "open",
        "title": "`GGML_ASSERT(ret.axis != GGML_BACKEND_SPLIT_AXIS_UNKNOWN) failed` on startup with `--split-mode tensor` and `iq4_nl` kv-cache",
        "labels": [],
        "created_at": "2026-08-15T09:12:32Z",
        "updated_at": "2026-08-17T09:30:13Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27116",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27117,
        "kind": "issue",
        "state": "open",
        "title": "speculative: draft-dflash draft acceptance collapses under concurrent sequences (-np 16 pathological, -np 4 healthy, --spec-draft-n-max 1 recovers)",
        "labels": [],
        "created_at": "2026-08-15T09:14:17Z",
        "updated_at": "2026-08-23T16:52:33Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27117",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27118,
        "kind": "issue",
        "state": "open",
        "title": "Proposal: Have 2 reasoning settings in the webui, one for effort/strength and another for the token limit",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-08-15T09:31:02Z",
        "updated_at": "2026-08-15T10:43:01Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27118",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 27119,
        "kind": "pr",
        "state": "open",
        "title": "grammar : support legacy ac-XX syntax and regex class shorthands",
        "labels": [],
        "created_at": "2026-08-15T09:56:06Z",
        "updated_at": "2026-08-15T10:11:13Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27119",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 27122,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: MTP triggers reproducible CUDA lockups with Qwen3.8-27B while --split-mode tensor",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-15T10:40:00Z",
        "updated_at": "2026-08-26T14:55:29Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27122",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 27123,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: WebUI - Permit sending different sampling params depending on reasoning state.",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-08-15T10:46:07Z",
        "updated_at": "2026-08-16T10:14:05Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27123",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 27125,
        "kind": "pr",
        "state": "open",
        "title": "llama : use int32_t for llama_sampler_chain_n return type",
        "labels": [],
        "created_at": "2026-08-15T11:49:10Z",
        "updated_at": "2026-08-16T02:15:05Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27125",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27127,
        "kind": "issue",
        "state": "open",
        "title": "ggml-hip: dp4a-based vec_dot for ternary/binary quants on RDNA4 (+37% measured Q2_0) — intent to upstream",
        "labels": [],
        "created_at": "2026-08-15T12:12:41Z",
        "updated_at": "2026-08-15T20:37:33Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27127",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 27128,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: llama-b10441-bin-win-sycl-x64 run error",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-15T12:22:07Z",
        "updated_at": "2026-08-22T15:55:27Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27128",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27129,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: server silently drops the tools array when the chat template has no tool support (--jinja)",
        "labels": [],
        "created_at": "2026-08-15T12:47:58Z",
        "updated_at": "2026-08-15T12:47:58Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27129",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27131,
        "kind": "pr",
        "state": "open",
        "title": "Fix Jinja strict template parsing for system messages",
        "labels": [
          "server"
        ],
        "created_at": "2026-08-15T13:13:09Z",
        "updated_at": "2026-08-15T13:18:00Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27131",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 27132,
        "kind": "pr",
        "state": "open",
        "title": "fix(convert): qwen3_5 hybrid linear-attention tensors - ssm_conv1d kernel dim + in_proj_a/b layout",
        "labels": [
          "conversion"
        ],
        "created_at": "2026-08-15T13:32:06Z",
        "updated_at": "2026-08-15T18:52:52Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27132",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "speculative_generation",
          "loader_conversion",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 27134,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: regression — assistant reply misfiled into reasoning_content (content empty) for chat templates whose generation prompt ends with a closing </think> (works on fe2adf0/b10252, broken on b10360)",
        "labels": [],
        "created_at": "2026-08-15T14:37:10Z",
        "updated_at": "2026-08-15T14:37:10Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27134",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27136,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Audio not working with ggml-org/Qwen3-Omni-30B-A3B-Instruct-GGUF",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-15T15:01:43Z",
        "updated_at": "2026-08-15T15:01:43Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27136",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27137,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: 2.3x performance regression from flash attention auto-enabling.",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-15T15:03:15Z",
        "updated_at": "2026-08-15T15:14:54Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27137",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27139,
        "kind": "issue",
        "state": "open",
        "title": "Qwen3.8 Codex error resolved by using the Qwen3.6 chat template file.",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-15T15:55:52Z",
        "updated_at": "2026-08-15T16:20:19Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27139",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27140,
        "kind": "pr",
        "state": "open",
        "title": "Slow prefill on small KV quants fixed",
        "labels": [
          "documentation",
          "examples",
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-08-15T16:26:13Z",
        "updated_at": "2026-08-19T08:58:05Z",
        "closed_at": null,
        "comments": 12,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27140",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27141,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: nemotron_h_moe aborts in ggml_ssm_scan during context reservation — GGML_ASSERT(K >= 1) failed (b10360)",
        "labels": [],
        "created_at": "2026-08-15T16:29:47Z",
        "updated_at": "2026-08-18T20:40:58Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27141",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27143,
        "kind": "pr",
        "state": "open",
        "title": "examples: share thread pools between target and drafter",
        "labels": [
          "examples"
        ],
        "created_at": "2026-08-15T17:01:18Z",
        "updated_at": "2026-08-18T18:37:07Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27143",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "speculative_generation",
          "api_ui_docs"
        ]
      },
      {
        "number": 27145,
        "kind": "pr",
        "state": "open",
        "title": "Update ROCm to 7.14.0 for Docker builds",
        "labels": [
          "documentation",
          "devops"
        ],
        "created_at": "2026-08-15T17:51:22Z",
        "updated_at": "2026-08-26T14:39:09Z",
        "closed_at": null,
        "comments": 15,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27145",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27146,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: mmproj/mtmd models balloon GTT allocations to ~33 GB total-vm at load on AMD iGPU (Vulkan) -> system-wide OOM",
        "labels": [],
        "created_at": "2026-08-15T18:06:15Z",
        "updated_at": "2026-08-15T18:06:15Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27146",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 27147,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: first image request on multimodal models is much slower than subsequent ones, because mtmd warmup doesn't perform an encode",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-15T21:12:36Z",
        "updated_at": "2026-08-15T21:12:36Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27147",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27148,
        "kind": "issue",
        "state": "open",
        "title": "server: RAM-backed prompt cache (--cache-ram/--cache-idle-slots) restores content from an unrelated conversation into a fresh slot under concurrent load",
        "labels": [],
        "created_at": "2026-08-15T21:16:21Z",
        "updated_at": "2026-08-22T16:16:41Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27148",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 27151,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: MTP draft acceptance 1 of 633",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-15T22:11:25Z",
        "updated_at": "2026-08-25T07:49:54Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27151",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27152,
        "kind": "pr",
        "state": "open",
        "title": "mtmd : trigger one-time inits for vision encoder at warmup",
        "labels": [
          "mtmd"
        ],
        "created_at": "2026-08-15T23:07:15Z",
        "updated_at": "2026-08-25T13:33:42Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27152",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27155,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: VRAM leak with DeepSeek V4 Flash + DSpark: Draft KV cache grows ~10MB per PP+TG cycle until OOM",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-16T02:22:23Z",
        "updated_at": "2026-08-18T06:00:46Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27155",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27156,
        "kind": "pr",
        "state": "open",
        "title": "vulkan: add runtime cache and Termux compatibility fixes",
        "labels": [
          "Vulkan",
          "ggml"
        ],
        "created_at": "2026-08-16T02:35:45Z",
        "updated_at": "2026-08-16T03:39:27Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27156",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 27158,
        "kind": "issue",
        "state": "open",
        "title": "support MiniMax-Music3",
        "labels": [],
        "created_at": "2026-08-16T02:59:54Z",
        "updated_at": "2026-08-16T02:59:54Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27158",
        "categories": [
          "quantization",
          "models_architectures",
          "multimodal_embedding"
        ]
      },
      {
        "number": 27161,
        "kind": "pr",
        "state": "open",
        "title": "convert : support mixed FP8 and NVFP4 compressed-tensors",
        "labels": [
          "testing",
          "conversion"
        ],
        "created_at": "2026-08-16T04:14:08Z",
        "updated_at": "2026-08-16T05:19:06Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27161",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27162,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: OOM being displayed as \"\"\"layer 0 is assigned to device CPU but fused Gated Delta Net (chunked) is assigned to device CUDA0 (usually due to missing support)\"\"\" instead of correctly erroring out.",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-16T06:22:06Z",
        "updated_at": "2026-08-16T06:22:06Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27162",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27163,
        "kind": "pr",
        "state": "open",
        "title": "vulkan: optimize ARM Mali subgroup warptiles",
        "labels": [
          "documentation",
          "Vulkan",
          "ggml"
        ],
        "created_at": "2026-08-16T06:23:35Z",
        "updated_at": "2026-08-16T07:33:35Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27163",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27166,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: stack buffer overflow in clip_encode MiniCPM-V position buckets via malicious mmproj (bucket_coords_h[1024] written for i < pos_h)",
        "labels": [],
        "created_at": "2026-08-16T07:16:21Z",
        "updated_at": "2026-08-16T07:16:21Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27166",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27168,
        "kind": "issue",
        "state": "open",
        "title": "SYCL/OpenCL: ur_die 'Experimental P2P feature is not implemented for OpenCL adapter' in dev2dev_memcpy on multi-GPU Arc",
        "labels": [],
        "created_at": "2026-08-16T07:36:30Z",
        "updated_at": "2026-08-22T15:42:45Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27168",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 27169,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: SIGFPE (integer divide-by-zero) in common_params_fit_impl with -ngl 0 on multi-GPU",
        "labels": [],
        "created_at": "2026-08-16T08:24:01Z",
        "updated_at": "2026-08-16T09:57:58Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27169",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27170,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Can you disable kv cache?",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-08-16T08:39:25Z",
        "updated_at": "2026-08-16T10:01:48Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27170",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 27171,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: regression with Qwen3.6-35B-A3B-Q4_K_M and --fit-target",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-16T08:44:49Z",
        "updated_at": "2026-08-16T22:20:26Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27171",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27172,
        "kind": "pr",
        "state": "open",
        "title": "cuda : add DeviceBatchedTopK ggml_top_k() implementation",
        "labels": [
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-08-16T09:04:16Z",
        "updated_at": "2026-08-17T14:28:51Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27172",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27173,
        "kind": "pr",
        "state": "open",
        "title": "speculative : draft performance improvement (+10% t/s with deeper draft depth possible with this) + token rollback bugfix",
        "labels": [
          "model",
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-08-16T09:42:31Z",
        "updated_at": "2026-08-25T15:38:38Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27173",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27174,
        "kind": "issue",
        "state": "open",
        "title": "Completions endpoint: logprobs returned for generated tokens only — no prompt/echo logprobs, silently breaks all loglikelihood evals (lm-eval etc.)",
        "labels": [],
        "created_at": "2026-08-16T10:39:55Z",
        "updated_at": "2026-08-17T19:16:59Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27174",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27175,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: GGML_ASSERT(ggml_can_mul_mat) in build_pooling when n_batch/n_ubatch not divisible by n_seq_max (MEAN pooling)",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-16T11:01:17Z",
        "updated_at": "2026-08-16T14:47:55Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27175",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27177,
        "kind": "pr",
        "state": "open",
        "title": "grammar : accept repetitions of exactly MAX_REPETITION_THRESHOLD",
        "labels": [
          "testing"
        ],
        "created_at": "2026-08-16T11:56:08Z",
        "updated_at": "2026-08-16T12:00:52Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27177",
        "categories": [
          "models_architectures",
          "structured_generation_tools",
          "correctness_testing",
          "training_adapters"
        ]
      },
      {
        "number": 27178,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Muse Glimmer NVFP4 fails to load, then emits one repeated token",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-16T12:28:10Z",
        "updated_at": "2026-08-19T17:35:57Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27178",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 27180,
        "kind": "issue",
        "state": "open",
        "title": "ggml-openvino: frontend compiled-model cache import loads but cannot compute — exported port names embed pointer-derived hash suffix (map::at on first inference)",
        "labels": [],
        "created_at": "2026-08-16T13:11:38Z",
        "updated_at": "2026-08-20T03:47:06Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27180",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27181,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: MMQ prefill speed on gfx1201 not proportional to quantization size Q2_K and Q6_K use unoptimized path",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-08-16T13:27:35Z",
        "updated_at": "2026-08-16T15:40:15Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27181",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 27183,
        "kind": "pr",
        "state": "open",
        "title": "vulkan : return GGML_STATUS_FAILED on device loss instead of aborting",
        "labels": [
          "Vulkan",
          "ggml"
        ],
        "created_at": "2026-08-16T14:22:33Z",
        "updated_at": "2026-08-26T06:14:49Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27183",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27185,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: HIP/ROCm slot admitted but never dispatched to a prefill batch (n_prompt_tokens_processed stuck at 0) while a concurrent slot processes normally",
        "labels": [],
        "created_at": "2026-08-16T14:27:40Z",
        "updated_at": "2026-08-17T03:10:36Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27185",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27186,
        "kind": "pr",
        "state": "open",
        "title": "llama: fix mul_mat assert when reserving graph",
        "labels": [
          "testing"
        ],
        "created_at": "2026-08-16T14:45:11Z",
        "updated_at": "2026-08-16T14:46:00Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27186",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27187,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: llama-server crashes with unhandled std::bad_function_call",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-16T15:03:29Z",
        "updated_at": "2026-08-22T22:09:47Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27187",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 27189,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Instant crash (segfault) when loading models with AMD Strix Halo and Vulkan",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-16T16:17:25Z",
        "updated_at": "2026-08-17T10:47:41Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27189",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 27190,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: SmolVLM-Instruct fails with `invalid token[?] = -1`",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-16T16:34:08Z",
        "updated_at": "2026-08-16T16:34:08Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27190",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 27191,
        "kind": "issue",
        "state": "open",
        "title": "Adreno/Turnip: GPU clock throttling between submissions hurts tg128 (KGSL power-constraint, likely affects any KGSL-kernel Snapdragon Linux, not just Android)",
        "labels": [],
        "created_at": "2026-08-16T16:34:14Z",
        "updated_at": "2026-08-16T16:34:14Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27191",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 27192,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cuda: hipBLAS 7.x compat — fix enum-type mismatch in cublasGemmEx",
        "labels": [
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-08-16T17:19:55Z",
        "updated_at": "2026-08-16T17:24:37Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27192",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27193,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: vulkan: fuse GATED_DELTA_NET state write into recurrent cache (skip following CPY)",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-08-16T17:21:50Z",
        "updated_at": "2026-08-16T17:21:50Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27193",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 27194,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: vulkan: fuse UNARY(SIGMOID|SILU) + MUL (CUDA parity)",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-08-16T17:24:34Z",
        "updated_at": "2026-08-16T17:24:34Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27194",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 27195,
        "kind": "pr",
        "state": "open",
        "title": "fix(mtmd): use string templates for Idefics3 overview image tokens",
        "labels": [
          "mtmd"
        ],
        "created_at": "2026-08-16T17:51:42Z",
        "updated_at": "2026-08-16T17:51:58Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27195",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27196,
        "kind": "pr",
        "state": "open",
        "title": "server: Support logprobs with speculative decoding",
        "labels": [
          "server"
        ],
        "created_at": "2026-08-16T18:01:23Z",
        "updated_at": "2026-08-25T15:22:02Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27196",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 27198,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: [SYCL] --split-mode tensor crashes in dev2dev_memcpy (DEVICE_LOST) on dual Arc Pro B70, despite working P2P",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-16T18:18:43Z",
        "updated_at": "2026-08-26T19:35:53Z",
        "closed_at": null,
        "comments": 17,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27198",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27199,
        "kind": "pr",
        "state": "open",
        "title": "finetune: fix no KV cache",
        "labels": [
          "documentation",
          "examples"
        ],
        "created_at": "2026-08-16T19:11:12Z",
        "updated_at": "2026-08-16T19:11:29Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27199",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 27200,
        "kind": "pr",
        "state": "open",
        "title": "ggml : fix integer overflow in block-quantized ggml_nbytes (heap OOB read)",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-08-16T19:55:54Z",
        "updated_at": "2026-08-25T15:48:03Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27200",
        "categories": [
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27201,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Quantization suffix in -hf-repo does not work as expected",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-16T20:00:27Z",
        "updated_at": "2026-08-21T16:05:24Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27201",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27202,
        "kind": "pr",
        "state": "open",
        "title": "mtmd: validate scalar mmproj metadata types and bound layer count",
        "labels": [
          "testing",
          "mtmd"
        ],
        "created_at": "2026-08-16T20:17:29Z",
        "updated_at": "2026-08-16T21:38:37Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27202",
        "categories": [
          "quantization",
          "models_architectures",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27203,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Overflow bypasing tensor dimension check",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-16T20:22:17Z",
        "updated_at": "2026-08-25T12:51:03Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27203",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27205,
        "kind": "issue",
        "state": "open",
        "title": "ggml-openvino: RESHAPE result that crosses a graph-split boundary is bound with the source tensor's shape → GPU plugin rejects input (\"tensor size is not equal to model\") on hybrid MoE models (qwen35moe 35B)",
        "labels": [],
        "created_at": "2026-08-16T21:19:28Z",
        "updated_at": "2026-08-16T21:19:28Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27205",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27206,
        "kind": "issue",
        "state": "open",
        "title": "ggml-openvino: hybrid-attention MoE models (qwen35moe, e.g. Ornith 1.0 35B) fail to compute — three defects in split-boundary tensor handling (reshape view-src extra sharing, boundary view/reshape outputs never emitted, shape-blind compiled-model cache)",
        "labels": [],
        "created_at": "2026-08-16T21:19:33Z",
        "updated_at": "2026-08-16T21:22:21Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27206",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27207,
        "kind": "pr",
        "state": "open",
        "title": "Fix wrong update in pull req #25574",
        "labels": [],
        "created_at": "2026-08-16T22:18:42Z",
        "updated_at": "2026-08-20T08:11:04Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27207",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "speculative_generation",
          "correctness_testing"
        ]
      },
      {
        "number": 27209,
        "kind": "pr",
        "state": "open",
        "title": "model: Bias tensor splitting by device fractions, rather than rotating",
        "labels": [],
        "created_at": "2026-08-17T00:36:12Z",
        "updated_at": "2026-08-17T00:36:12Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27209",
        "categories": [
          "models_architectures",
          "build_platform_ops"
        ]
      },
      {
        "number": 27210,
        "kind": "pr",
        "state": "open",
        "title": "spec : add adaptive MTP draft depth (draft-mtp-adaptive)",
        "labels": [
          "model",
          "testing",
          "server"
        ],
        "created_at": "2026-08-17T01:38:49Z",
        "updated_at": "2026-08-26T11:34:15Z",
        "closed_at": null,
        "comments": 61,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27210",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27211,
        "kind": "issue",
        "state": "open",
        "title": "Opt-in codec for recurrent-state context checkpoints (2x less host RAM, off by default)",
        "labels": [],
        "created_at": "2026-08-17T01:49:43Z",
        "updated_at": "2026-08-17T02:36:03Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27211",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27212,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: draft-mtp fails on Tesla P100 (sm_60) with Linux Docker",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-17T01:57:59Z",
        "updated_at": "2026-08-17T01:57:59Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27212",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27214,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: -no-cnv has no effect on llama-cli; the process then blocks at \">\" forever",
        "labels": [],
        "created_at": "2026-08-17T02:08:26Z",
        "updated_at": "2026-08-22T12:24:08Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27214",
        "categories": [
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27217,
        "kind": "issue",
        "state": "open",
        "title": "server: tool_choice: \"required\" accepted but not enforced on templates with supports_preserve_reasoning=true",
        "labels": [],
        "created_at": "2026-08-17T03:09:16Z",
        "updated_at": "2026-08-19T07:54:55Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27217",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 27218,
        "kind": "pr",
        "state": "open",
        "title": "ggml-hrx: add AMD ROCm HRX native ggml backend",
        "labels": [
          "testing",
          "ggml"
        ],
        "created_at": "2026-08-17T03:15:04Z",
        "updated_at": "2026-08-21T09:50:52Z",
        "closed_at": null,
        "comments": 7,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27218",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27220,
        "kind": "pr",
        "state": "open",
        "title": "vulkan : fuse UNARY(GELU|SIGMOID|SILU|SOFTPLUS) + MUL",
        "labels": [
          "testing",
          "Vulkan",
          "ggml"
        ],
        "created_at": "2026-08-17T05:20:33Z",
        "updated_at": "2026-08-26T15:16:20Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27220",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27221,
        "kind": "pr",
        "state": "open",
        "title": "chat: allow reasoning_effort to control enable_thinking",
        "labels": [
          "server"
        ],
        "created_at": "2026-08-17T06:01:00Z",
        "updated_at": "2026-08-17T06:01:28Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27221",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing"
        ]
      },
      {
        "number": 27222,
        "kind": "issue",
        "state": "open",
        "title": "[Bug]: Granite4 Vision mmproj with downsample_window_side=0 crashes at load (integer division by zero in warmup graph build)",
        "labels": [],
        "created_at": "2026-08-17T06:42:37Z",
        "updated_at": "2026-08-17T06:42:37Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27222",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27227,
        "kind": "pr",
        "state": "open",
        "title": "Eliminate slashes as end-of-text markers. Ctrl-D ends input now.",
        "labels": [],
        "created_at": "2026-08-17T06:44:20Z",
        "updated_at": "2026-08-17T06:44:20Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27227",
        "categories": [
          "models_architectures",
          "build_platform_ops"
        ]
      },
      {
        "number": 27230,
        "kind": "pr",
        "state": "open",
        "title": "server : check existence of sink callback before calling it",
        "labels": [
          "server"
        ],
        "created_at": "2026-08-17T07:18:01Z",
        "updated_at": "2026-08-17T07:18:18Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27230",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing"
        ]
      },
      {
        "number": 27231,
        "kind": "pr",
        "state": "open",
        "title": "Improve error message when requested quant is not found in repo",
        "labels": [],
        "created_at": "2026-08-17T07:24:41Z",
        "updated_at": "2026-08-18T13:52:23Z",
        "closed_at": null,
        "comments": 9,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27231",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27236,
        "kind": "pr",
        "state": "open",
        "title": "convert: add test_convert.py",
        "labels": [
          "conversion"
        ],
        "created_at": "2026-08-17T09:10:03Z",
        "updated_at": "2026-08-17T09:10:19Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27236",
        "categories": [
          "models_architectures",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27237,
        "kind": "issue",
        "state": "open",
        "title": "[Vulkan] Qwen3.5-27B (qwen35 hybrid DeltaNet) garbage output at batch size 512; OK at 1024/4096",
        "labels": [],
        "created_at": "2026-08-17T09:11:09Z",
        "updated_at": "2026-08-19T17:04:41Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27237",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27243,
        "kind": "pr",
        "state": "open",
        "title": "ui: Models Selector Advanced Options",
        "labels": [
          "server/ui"
        ],
        "created_at": "2026-08-17T09:33:04Z",
        "updated_at": "2026-08-25T13:55:42Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27243",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27244,
        "kind": "pr",
        "state": "open",
        "title": "ui: Import Conversations or Settings by drag-and-drop",
        "labels": [
          "server/ui"
        ],
        "created_at": "2026-08-17T09:35:51Z",
        "updated_at": "2026-08-21T13:38:51Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27244",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27248,
        "kind": "pr",
        "state": "open",
        "title": "Add CUDA support for `q4_1`, `iq4_nl`, `q5_0` and `q5_1` kv-cache quants",
        "labels": [
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-08-17T10:03:04Z",
        "updated_at": "2026-08-17T10:35:55Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27248",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context"
        ]
      },
      {
        "number": 27256,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cpu : fix dup_bytes row size for block-quantized types",
        "labels": [
          "testing",
          "ggml"
        ],
        "created_at": "2026-08-17T12:02:45Z",
        "updated_at": "2026-08-18T07:39:37Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27256",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27257,
        "kind": "issue",
        "state": "open",
        "title": "Compile bug: Level Zero loader or headers not found, Level Zero support disabled",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-17T12:09:48Z",
        "updated_at": "2026-08-22T15:55:06Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27257",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27258,
        "kind": "pr",
        "state": "open",
        "title": "[DRAFT]: Formalizing scheduler and async backend behavior through tests",
        "labels": [
          "testing",
          "Vulkan",
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-08-17T12:23:09Z",
        "updated_at": "2026-08-26T15:10:13Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27258",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27259,
        "kind": "issue",
        "state": "open",
        "title": "gguf : GGML_PAD difference underflows to ~SIZE_MAX for near-SIZE_MAX tensor sizes",
        "labels": [],
        "created_at": "2026-08-17T12:23:16Z",
        "updated_at": "2026-08-17T12:23:16Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27259",
        "categories": [
          "quantization",
          "models_architectures",
          "loader_conversion",
          "training_adapters"
        ]
      },
      {
        "number": 27264,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug:  [Vulkan] -ngl not working as intended. Model loads entirely in VRAM",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-17T13:37:54Z",
        "updated_at": "2026-08-21T21:36:06Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27264",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27265,
        "kind": "pr",
        "state": "open",
        "title": "common : avoid zero divisor in multi-device fit (#27169)",
        "labels": [],
        "created_at": "2026-08-17T14:04:32Z",
        "updated_at": "2026-08-17T17:12:52Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27265",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27266,
        "kind": "pr",
        "state": "open",
        "title": "Server toolcall image output",
        "labels": [
          "testing",
          "server"
        ],
        "created_at": "2026-08-17T14:08:09Z",
        "updated_at": "2026-08-17T14:38:45Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27266",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27267,
        "kind": "pr",
        "state": "open",
        "title": "CUDA : fuse RWKV7 output preparation",
        "labels": [
          "testing",
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-08-17T14:51:38Z",
        "updated_at": "2026-08-26T09:30:13Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27267",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27268,
        "kind": "pr",
        "state": "open",
        "title": "server: do not re-reserve sched on lora scale changes",
        "labels": [
          "server"
        ],
        "created_at": "2026-08-17T14:58:27Z",
        "updated_at": "2026-08-17T19:00:37Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27268",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "training_adapters"
        ]
      },
      {
        "number": 27269,
        "kind": "pr",
        "state": "open",
        "title": "[cuda] enable q8_0 q4_0 FAttn vector kernels",
        "labels": [
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-08-17T15:01:01Z",
        "updated_at": "2026-08-17T15:35:59Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27269",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "build_platform_ops"
        ]
      },
      {
        "number": 27275,
        "kind": "issue",
        "state": "open",
        "title": "ROCm 7.x comgr segfault on gfx1030 (RDNA2) — side-by-side ROCm 6.3.1 workaround restores full MTP throughput",
        "labels": [],
        "created_at": "2026-08-17T16:48:35Z",
        "updated_at": "2026-08-17T16:48:35Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27275",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27276,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: [CPU] ARM i8mm 8x8 blocked GEMM/GEMV paths for TQ1_0 and TQ2_0",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-08-17T16:49:26Z",
        "updated_at": "2026-08-22T18:07:26Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27276",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27279,
        "kind": "issue",
        "state": "open",
        "title": "server: `response_format.json_schema` under `peg-native` — generation stops mid-object, then `common_chat_peg_parse` throws 500",
        "labels": [],
        "created_at": "2026-08-17T17:47:18Z",
        "updated_at": "2026-08-18T04:39:40Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27279",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 27280,
        "kind": "pr",
        "state": "open",
        "title": "server: add optional OTLP/HTTP tracing",
        "labels": [
          "documentation",
          "build",
          "server"
        ],
        "created_at": "2026-08-17T18:20:33Z",
        "updated_at": "2026-08-18T08:10:13Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27280",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27282,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: native MTP reserves a separate CUDA compute arena and OOMs; shared gallocr fixes it",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-17T18:42:09Z",
        "updated_at": "2026-08-25T12:04:22Z",
        "closed_at": null,
        "comments": 10,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27282",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27283,
        "kind": "pr",
        "state": "open",
        "title": "ggml : bound-check row index in get_rows_back (CWE-125/787)",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-08-17T18:51:08Z",
        "updated_at": "2026-08-17T23:22:22Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27283",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 27285,
        "kind": "pr",
        "state": "open",
        "title": "mtmd : check for missing optional tensors before deref",
        "labels": [
          "mtmd"
        ],
        "created_at": "2026-08-17T19:42:52Z",
        "updated_at": "2026-08-19T08:22:16Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27285",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27286,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cpu : validate expert id in mul_mat_id",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-08-17T20:32:32Z",
        "updated_at": "2026-08-18T22:18:00Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27286",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27293,
        "kind": "pr",
        "state": "open",
        "title": "server : document --poll 0 recipe for always-on deployments",
        "labels": [
          "documentation",
          "server"
        ],
        "created_at": "2026-08-18T03:34:48Z",
        "updated_at": "2026-08-20T00:52:27Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27293",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27296,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: MTP breaks long/short inference?",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-18T05:04:20Z",
        "updated_at": "2026-08-25T09:33:05Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27296",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27301,
        "kind": "pr",
        "state": "open",
        "title": "ggml: allow passing alloc dependencies in graph_optimize",
        "labels": [
          "testing",
          "Vulkan",
          "ggml",
          "Apple Metal",
          "Hexagon",
          "CUDA"
        ],
        "created_at": "2026-08-18T06:29:26Z",
        "updated_at": "2026-08-24T13:36:14Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27301",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27303,
        "kind": "issue",
        "state": "open",
        "title": "HIP build fails on ROCm 6.2.x — `vendors/hip.h` typedefs `__hip_fp8_e4m3`, which does not exist before ROCm 6.3.0",
        "labels": [],
        "created_at": "2026-08-18T06:53:06Z",
        "updated_at": "2026-08-26T12:09:57Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27303",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27306,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: draft-mtp DeviceLost during *prompt* on AMD RADV — `common_speculative_process` runs `llama_decode(ctx_dft)` after every prefill ubatch",
        "labels": [],
        "created_at": "2026-08-18T07:58:22Z",
        "updated_at": "2026-08-22T05:37:38Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27306",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27309,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: server reports \"model loaded\" and binds the port after a fatal Metal OOM during init; every request then 500s",
        "labels": [],
        "created_at": "2026-08-18T08:39:37Z",
        "updated_at": "2026-08-21T18:23:30Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27309",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27310,
        "kind": "pr",
        "state": "open",
        "title": "spec : fuse the encoder into the KV injection decode",
        "labels": [
          "model"
        ],
        "created_at": "2026-08-18T09:03:40Z",
        "updated_at": "2026-08-26T13:20:38Z",
        "closed_at": null,
        "comments": 9,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27310",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "speculative_generation",
          "build_platform_ops"
        ]
      },
      {
        "number": 27311,
        "kind": "pr",
        "state": "open",
        "title": "Scheduler UMA ring buffer (+ sanitizer and fixes)",
        "labels": [
          "documentation",
          "testing",
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-08-18T09:24:33Z",
        "updated_at": "2026-08-26T17:24:46Z",
        "closed_at": null,
        "comments": 9,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27311",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27312,
        "kind": "pr",
        "state": "open",
        "title": "ui: Added agent skills support",
        "labels": [
          "documentation",
          "testing",
          "server",
          "server/ui"
        ],
        "created_at": "2026-08-18T09:25:55Z",
        "updated_at": "2026-08-24T19:27:10Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27312",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27316,
        "kind": "issue",
        "state": "open",
        "title": "CI: clean up release.yml webui build logic and remove HF bucket UI hosting",
        "labels": [],
        "created_at": "2026-08-18T10:40:59Z",
        "updated_at": "2026-08-21T18:20:18Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27316",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27318,
        "kind": "pr",
        "state": "open",
        "title": "server: fix completion_tokens/total_tokens undercount when n > 1",
        "labels": [
          "server"
        ],
        "created_at": "2026-08-18T11:42:20Z",
        "updated_at": "2026-08-18T11:54:15Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27318",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 27322,
        "kind": "pr",
        "state": "open",
        "title": "quantize: add IQ2_NL and IQ3_NL types (CPU)",
        "labels": [
          "testing",
          "examples",
          "ggml",
          "SYCL",
          "Apple Metal",
          "conversion"
        ],
        "created_at": "2026-08-18T13:31:34Z",
        "updated_at": "2026-08-23T13:59:32Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27322",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 27324,
        "kind": "pr",
        "state": "open",
        "title": "quantize: add IQ2_NL and IQ3_NL types (CPU + Metal)",
        "labels": [
          "testing",
          "examples",
          "ggml",
          "SYCL",
          "Apple Metal",
          "conversion"
        ],
        "created_at": "2026-08-18T13:35:42Z",
        "updated_at": "2026-08-22T00:28:12Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27324",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 27325,
        "kind": "pr",
        "state": "open",
        "title": "quantize: add IQ2_NL and IQ3_NL types (CPU + Metal + CUDA)",
        "labels": [
          "testing",
          "examples",
          "ggml",
          "SYCL",
          "Apple Metal",
          "CUDA",
          "conversion"
        ],
        "created_at": "2026-08-18T13:44:00Z",
        "updated_at": "2026-08-22T00:28:18Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27325",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 27326,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: WebUI Stop button cannot abort in-flight inference (incl. prefill) when API key auth is enabled",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-18T14:07:17Z",
        "updated_at": "2026-08-18T14:07:17Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27326",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27327,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Low speed of MoE models when unloading to RAM with VRAM. Degradation more than 3 times.",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-18T14:07:57Z",
        "updated_at": "2026-08-19T06:37:47Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27327",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 27328,
        "kind": "pr",
        "state": "open",
        "title": "[cuda]Make NCCL FP32/BF16 crossover threshold configurable",
        "labels": [
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-08-18T14:38:51Z",
        "updated_at": "2026-08-18T19:53:11Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27328",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27329,
        "kind": "issue",
        "state": "open",
        "title": "[CUDA][Blackwell] Qwen3.8-27B-NVFP4 decode hangs, CPU spin, no GPU work",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-18T14:51:07Z",
        "updated_at": "2026-08-23T22:49:59Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27329",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27330,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: CUDA graphs hang the GPU channel (RC watchdog + Xid 8) on RTX 5090 Laptop / sm_120; GGML_CUDA_DISABLE_GRAPHS=1 is a complete workaround",
        "labels": [],
        "created_at": "2026-08-18T15:00:05Z",
        "updated_at": "2026-08-19T12:43:11Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27330",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 27332,
        "kind": "pr",
        "state": "open",
        "title": "vulkan: use density gate for MUL_MAT_VEC_ID path",
        "labels": [
          "Vulkan",
          "ggml"
        ],
        "created_at": "2026-08-18T16:08:54Z",
        "updated_at": "2026-08-25T07:49:54Z",
        "closed_at": null,
        "comments": 9,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27332",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27333,
        "kind": "pr",
        "state": "open",
        "title": "llama-mmap: fix numa path hint inconsistency.",
        "labels": [],
        "created_at": "2026-08-18T16:12:12Z",
        "updated_at": "2026-08-21T16:45:00Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27333",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27334,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: errors and poor performance on vulkan tensor split on quad rx580",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-18T16:25:36Z",
        "updated_at": "2026-08-19T09:33:37Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27334",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27335,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: crash on M2 Ultra for Qwen3.8 27B defaults",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-18T18:03:22Z",
        "updated_at": "2026-08-18T18:06:38Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27335",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27341,
        "kind": "pr",
        "state": "open",
        "title": "CUDA: fuse ffn_gate + glu into the mul_mat_q epilogue",
        "labels": [
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-08-18T20:38:39Z",
        "updated_at": "2026-08-19T16:02:06Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27341",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 27342,
        "kind": "pr",
        "state": "open",
        "title": "spec : add DFlash2 support (local convolution + candidate selector)",
        "labels": [
          "model",
          "testing",
          "server",
          "ggml",
          "CUDA",
          "conversion"
        ],
        "created_at": "2026-08-18T20:53:14Z",
        "updated_at": "2026-08-26T21:02:16Z",
        "closed_at": null,
        "comments": 72,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27342",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing"
        ]
      },
      {
        "number": 27343,
        "kind": "pr",
        "state": "open",
        "title": "server, ui: allow inject message to generation",
        "labels": [
          "documentation",
          "server",
          "server/ui"
        ],
        "created_at": "2026-08-18T21:30:15Z",
        "updated_at": "2026-08-18T21:57:44Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27343",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 27350,
        "kind": "pr",
        "state": "open",
        "title": "metal : add i-quant support to mul_mv_ext (small-batch decode path)",
        "labels": [
          "ggml",
          "Apple Metal"
        ],
        "created_at": "2026-08-19T02:41:05Z",
        "updated_at": "2026-08-19T02:45:58Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27350",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 27351,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: llama-server in router mode can spawn child processes also in router mode",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-19T02:54:03Z",
        "updated_at": "2026-08-22T13:15:57Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27351",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27356,
        "kind": "pr",
        "state": "open",
        "title": "convert: fix Nemotron-H LoRA GGUF conversion",
        "labels": [
          "conversion"
        ],
        "created_at": "2026-08-19T05:41:51Z",
        "updated_at": "2026-08-24T06:39:26Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27356",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 27357,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: HTTP accept gzip/brotli compression",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-08-19T05:41:52Z",
        "updated_at": "2026-08-19T05:41:52Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27357",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27359,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Inkling support",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-08-19T06:49:28Z",
        "updated_at": "2026-08-19T06:49:28Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27359",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 27360,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Vulkan/RADV Radeon 680M regression: --load-mode auto causes device loss/OOM on large Gemma model; --load-mode mmap fixes it",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-19T06:55:12Z",
        "updated_at": "2026-08-19T06:55:12Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27360",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27361,
        "kind": "issue",
        "state": "open",
        "title": "Tool: KV-gate proxy to prevent 'KV pool full' deadlock with idle slots",
        "labels": [],
        "created_at": "2026-08-19T06:59:00Z",
        "updated_at": "2026-08-19T06:59:00Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27361",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27364,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: token per second metrics broken",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-19T07:23:37Z",
        "updated_at": "2026-08-19T12:05:09Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27364",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27366,
        "kind": "issue",
        "state": "open",
        "title": "Bug: -sm tensor hangs and leaks VRAM on multi-GPU Volta (V100); -sm row unavailable on CUDA",
        "labels": [],
        "created_at": "2026-08-19T07:57:33Z",
        "updated_at": "2026-08-19T07:57:33Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27366",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27367,
        "kind": "issue",
        "state": "open",
        "title": "Bug: HTTP 500 when a system message appears mid-conversation (strict chat templates, e.g. Qwen3.x)",
        "labels": [],
        "created_at": "2026-08-19T07:57:45Z",
        "updated_at": "2026-08-20T05:15:56Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27367",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27368,
        "kind": "pr",
        "state": "open",
        "title": "Fix #26558: Eval bug: llama-server hard crash (cublasSgemm INVALID_VALUE) with --spec-type draft-mtp under KV-cache saturation",
        "labels": [
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-08-19T08:22:06Z",
        "updated_at": "2026-08-21T15:11:54Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27368",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27371,
        "kind": "pr",
        "state": "open",
        "title": "perplexity : support --parse-special",
        "labels": [
          "documentation",
          "examples"
        ],
        "created_at": "2026-08-19T09:31:15Z",
        "updated_at": "2026-08-19T10:30:50Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27371",
        "categories": [
          "quantization",
          "models_architectures",
          "structured_generation_tools",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 27372,
        "kind": "pr",
        "state": "open",
        "title": "tests: add `-b/--backend` option to test-llama-archs for testing a specific backend",
        "labels": [
          "testing"
        ],
        "created_at": "2026-08-19T09:39:31Z",
        "updated_at": "2026-08-19T14:19:15Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27372",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27373,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: [SYCL] Degraded performance with MTP and Q3 quants",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-19T10:43:15Z",
        "updated_at": "2026-08-20T20:45:27Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27373",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27377,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cpu : add AVX2 vec_dot kernel for STQ1_0",
        "labels": [
          "testing",
          "examples",
          "ggml",
          "conversion"
        ],
        "created_at": "2026-08-19T11:19:01Z",
        "updated_at": "2026-08-19T12:10:20Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27377",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27379,
        "kind": "pr",
        "state": "open",
        "title": "CUDA/HIP: do not classify gfx909 and gfx90c APUs as CDNA",
        "labels": [
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-08-19T12:42:24Z",
        "updated_at": "2026-08-20T11:39:15Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27379",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27380,
        "kind": "pr",
        "state": "open",
        "title": "common/args: add env vars for temperature, top-p, min-p and penalties",
        "labels": [
          "documentation",
          "server"
        ],
        "created_at": "2026-08-19T12:51:34Z",
        "updated_at": "2026-08-20T07:30:21Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27380",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27387,
        "kind": "issue",
        "state": "open",
        "title": "Token corruption (wrong-alphabet chars) with two concurrent long generations; identical solo run is clean",
        "labels": [],
        "created_at": "2026-08-19T13:57:55Z",
        "updated_at": "2026-08-19T18:40:08Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27387",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27388,
        "kind": "issue",
        "state": "open",
        "title": "Server wedge: generation stalls mid-decode; /health OK but /slots hangs; needs SIGKILL",
        "labels": [],
        "created_at": "2026-08-19T13:58:04Z",
        "updated_at": "2026-08-19T13:58:04Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27388",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27396,
        "kind": "pr",
        "state": "open",
        "title": "Prefil optimizations for draft model",
        "labels": [
          "model",
          "server",
          "ggml"
        ],
        "created_at": "2026-08-19T15:25:25Z",
        "updated_at": "2026-08-19T22:06:02Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27396",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "build_platform_ops"
        ]
      },
      {
        "number": 27397,
        "kind": "issue",
        "state": "open",
        "title": "[server] json_schema_to_grammar: regex escape \"\\/\" passed through verbatim into GBNF → \"failed to parse grammar\" (HTTP 400)",
        "labels": [],
        "created_at": "2026-08-19T15:34:46Z",
        "updated_at": "2026-08-19T15:34:46Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27397",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27398,
        "kind": "pr",
        "state": "open",
        "title": "metal : remove the in-kernel dequantization from the FA kernels",
        "labels": [
          "ggml",
          "Apple Metal"
        ],
        "created_at": "2026-08-19T16:00:21Z",
        "updated_at": "2026-08-21T07:29:23Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27398",
        "categories": [
          "native_backends",
          "quantization",
          "memory_kv_context"
        ]
      },
      {
        "number": 27401,
        "kind": "pr",
        "state": "open",
        "title": "ui: add conversation compaction",
        "labels": [
          "server/ui"
        ],
        "created_at": "2026-08-19T17:57:57Z",
        "updated_at": "2026-08-23T09:06:47Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27401",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "api_ui_docs"
        ]
      },
      {
        "number": 27402,
        "kind": "pr",
        "state": "open",
        "title": "AVX2: Speed up large batch size prompt processing of IQ models",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-08-19T18:06:04Z",
        "updated_at": "2026-08-26T17:30:03Z",
        "closed_at": null,
        "comments": 10,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27402",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27406,
        "kind": "pr",
        "state": "open",
        "title": "cuda : prefer MMQ for IQ4_NL on CDNA2",
        "labels": [
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-08-19T18:24:26Z",
        "updated_at": "2026-08-19T20:53:00Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27406",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures"
        ]
      },
      {
        "number": 27407,
        "kind": "issue",
        "state": "open",
        "title": "spec: greedy output diverges from non-speculative baseline under batched verification on CUDA (deterministic numerics; draft-simple alone reproduces, DFlash2 amplifies)",
        "labels": [],
        "created_at": "2026-08-19T18:43:48Z",
        "updated_at": "2026-08-25T17:51:05Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27407",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27408,
        "kind": "issue",
        "state": "open",
        "title": "spec: mtmd image chunks leave positional holes in the draft KV cache -> llama_decode(ctx_dft) rc=-1 -> HTTP 500 with draft-dflash (CUDA); zero-fill patch crash-eliminating but no speculative benefit on images",
        "labels": [],
        "created_at": "2026-08-19T18:43:57Z",
        "updated_at": "2026-08-23T20:52:08Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27408",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27410,
        "kind": "pr",
        "state": "open",
        "title": "cuda : prefer MMQ for IQ4_XS on CDNA2",
        "labels": [
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-08-19T19:38:26Z",
        "updated_at": "2026-08-19T21:30:39Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27410",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing"
        ]
      },
      {
        "number": 27411,
        "kind": "pr",
        "state": "open",
        "title": "cuda : prefer MMQ for MXFP4 on CDNA2",
        "labels": [
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-08-19T19:40:12Z",
        "updated_at": "2026-08-19T21:44:24Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27411",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving"
        ]
      },
      {
        "number": 27412,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: SYCL run fails with The program was built for 1 devices (Iris(R) Xe Graphics)",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-19T19:48:16Z",
        "updated_at": "2026-08-22T17:53:43Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27412",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27415,
        "kind": "issue",
        "state": "open",
        "title": "Compile bug: version: 0.1.0-dev (build 10454, commit 4df29be4f) built with GNU 16.2.1 for Linux x86_64",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-19T21:18:30Z",
        "updated_at": "2026-08-19T21:18:30Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27415",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27418,
        "kind": "pr",
        "state": "open",
        "title": "server: add --sleep-mode rst, reset the proc while keeping metrics",
        "labels": [
          "documentation",
          "server"
        ],
        "created_at": "2026-08-19T22:34:42Z",
        "updated_at": "2026-08-20T14:10:03Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27418",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27419,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: upper image content is lost while Vulkan/CPU are correct",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-19T22:37:45Z",
        "updated_at": "2026-08-20T10:40:31Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27419",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27420,
        "kind": "issue",
        "state": "open",
        "title": "Vulkan/MTP performance bug triggered by ubatch=256 + f16 KV cache at large context",
        "labels": [],
        "created_at": "2026-08-19T22:39:39Z",
        "updated_at": "2026-08-19T22:39:39Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27420",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27422,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: System prompt leaking between different consumers / slots",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-19T23:14:23Z",
        "updated_at": "2026-08-21T18:24:33Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27422",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 27423,
        "kind": "pr",
        "state": "open",
        "title": "cuda : shrink the ubatch adaptively, gated on pipeline parallel",
        "labels": [],
        "created_at": "2026-08-19T23:15:20Z",
        "updated_at": "2026-08-19T23:42:03Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27423",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 27425,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Autotune tool to determine best configuration for op offload min batch size",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-08-20T01:28:01Z",
        "updated_at": "2026-08-20T01:28:01Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27425",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27427,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: A ~50 KB request causes a crash on llama-server, exit 139, OOMKilled=false, restart count 0 -> 1",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-20T04:50:29Z",
        "updated_at": "2026-08-20T04:50:29Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27427",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27428,
        "kind": "issue",
        "state": "open",
        "title": "eval bug: draft-mtp roughly halves prompt processing on multi-GPU layer split (single GPU is fine)",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-20T05:33:53Z",
        "updated_at": "2026-08-21T04:18:19Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27428",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27431,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: llama-cli and llama-server both crash when running unsloth/Qwen3.8-27B-UD-Q4_K_M.gguf on Vulkan (AMD R9700) on Windows",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-20T07:14:56Z",
        "updated_at": "2026-08-20T07:42:27Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27431",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27432,
        "kind": "pr",
        "state": "open",
        "title": "ui: export conversations from database instead of cached store",
        "labels": [
          "server/ui"
        ],
        "created_at": "2026-08-20T07:24:54Z",
        "updated_at": "2026-08-20T07:55:50Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27432",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 27435,
        "kind": "pr",
        "state": "open",
        "title": " add TeleChat model support",
        "labels": [
          "model",
          "testing",
          "ggml",
          "CUDA",
          "conversion"
        ],
        "created_at": "2026-08-20T09:35:13Z",
        "updated_at": "2026-08-21T01:49:05Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27435",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27436,
        "kind": "issue",
        "state": "open",
        "title": "Metrics gauges prompt_tokens_seconds / predicted_tokens_seconds are almost always 0, making live dashboards unusable",
        "labels": [],
        "created_at": "2026-08-20T09:50:09Z",
        "updated_at": "2026-08-20T09:50:09Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27436",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 27439,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: llama_state_seq_set_data_ext: invalid ON_DEVICE state can throw across the C API or abort",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-20T12:46:27Z",
        "updated_at": "2026-08-20T12:46:27Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27439",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27440,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: std::out_of_range in load_tensors when a device reports 0 free bytes (NaN tensor split)",
        "labels": [],
        "created_at": "2026-08-20T13:17:05Z",
        "updated_at": "2026-08-20T13:17:05Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27440",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27441,
        "kind": "pr",
        "state": "open",
        "title": "metal : skinny-batch mul_mm tiles + deterministic split-K for small-batch decode",
        "labels": [
          "ggml",
          "Apple Metal"
        ],
        "created_at": "2026-08-20T13:27:36Z",
        "updated_at": "2026-08-23T13:23:24Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27441",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27442,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: qwen35moe (hybrid recurrent/attention) models produce empty generation when prompt exceeds ~16K tokens via llama-server (Metal)",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-20T14:08:25Z",
        "updated_at": "2026-08-25T08:21:27Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27442",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27445,
        "kind": "issue",
        "state": "open",
        "title": "qwen35 embedding models: llama_get_embeddings_seq returns NULL → fallback to ith limited to 512 tokens",
        "labels": [],
        "created_at": "2026-08-20T14:46:18Z",
        "updated_at": "2026-08-20T14:46:18Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27445",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27448,
        "kind": "issue",
        "state": "open",
        "title": "Research: Full-bandwidth transformer (latent feedback)",
        "labels": [
          "research 🔬"
        ],
        "created_at": "2026-08-20T15:48:53Z",
        "updated_at": "2026-08-20T15:49:32Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27448",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "correctness_testing",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 27449,
        "kind": "pr",
        "state": "open",
        "title": "vulkan: handle larger batch sizes (>4) efficiently for IQ3_S mat-vec …",
        "labels": [
          "testing",
          "Vulkan",
          "ggml"
        ],
        "created_at": "2026-08-20T16:40:59Z",
        "updated_at": "2026-08-20T16:41:24Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27449",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27451,
        "kind": "pr",
        "state": "open",
        "title": "server : share checkpoint state and harden prompt cache OOM handling",
        "labels": [
          "server"
        ],
        "created_at": "2026-08-20T17:29:51Z",
        "updated_at": "2026-08-20T17:47:17Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27451",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27452,
        "kind": "pr",
        "state": "open",
        "title": "add hardware fp16_to_fp32 ",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-08-20T17:42:18Z",
        "updated_at": "2026-08-20T17:55:29Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27452",
        "categories": [
          "quantization",
          "models_architectures",
          "build_platform_ops"
        ]
      },
      {
        "number": 27453,
        "kind": "pr",
        "state": "open",
        "title": "Feature: Added LIGHTNING_INDEXER support for Deepseek V4 ops on Vulkan Backend",
        "labels": [
          "testing",
          "Vulkan",
          "ggml"
        ],
        "created_at": "2026-08-20T18:13:44Z",
        "updated_at": "2026-08-26T21:11:08Z",
        "closed_at": null,
        "comments": 12,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27453",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing"
        ]
      },
      {
        "number": 27454,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Loading a draft model crashes with \"invalid vector subscript\" when the target model has filled the GPU (NaN in `splits` → `devices.at()` out of range)",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-20T18:21:40Z",
        "updated_at": "2026-08-21T03:44:44Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27454",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27455,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: expose the reasoning effort levels server side and consume them in llama-ui",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-08-20T19:18:09Z",
        "updated_at": "2026-08-20T19:18:13Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27455",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27456,
        "kind": "issue",
        "state": "open",
        "title": "llama-server (router mode) crashes with 0xC0000409 when request hits busy slot during long xhigh generation",
        "labels": [],
        "created_at": "2026-08-20T19:24:01Z",
        "updated_at": "2026-08-21T08:52:15Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27456",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27458,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: E ggml_vulkan: device lost on Vulkan0 when attempting to use DFlash",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-20T20:09:06Z",
        "updated_at": "2026-08-23T13:39:59Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27458",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27459,
        "kind": "issue",
        "state": "open",
        "title": "Compile bug: Assembly error on build with Cuda 12.8",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-20T20:25:38Z",
        "updated_at": "2026-08-25T19:41:21Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27459",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27460,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: draft-mtp (self-speculative) models crash on Vulkan/RADV after Linux kernel bump 7.1.3 → 7.1.7 — same class as #24492, different GPU/model",
        "labels": [],
        "created_at": "2026-08-20T20:54:33Z",
        "updated_at": "2026-08-20T20:54:33Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27460",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27461,
        "kind": "pr",
        "state": "open",
        "title": "metal: enable Metal 4.0 tensor API on M5+/A19+",
        "labels": [
          "ggml",
          "Apple Metal"
        ],
        "created_at": "2026-08-20T21:23:34Z",
        "updated_at": "2026-08-24T17:16:53Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27461",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27462,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Bailing V3 (Ling 3.0) autoparser lacks a <tool_call> reasoning terminator but emits these often mid-reasoning",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-20T22:47:12Z",
        "updated_at": "2026-08-23T11:20:11Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27462",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27463,
        "kind": "issue",
        "state": "open",
        "title": "cant force stop localhost:8080 (llama-ui) no matter what i try",
        "labels": [],
        "created_at": "2026-08-20T23:29:50Z",
        "updated_at": "2026-08-20T23:29:50Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27463",
        "categories": [
          "models_architectures",
          "api_ui_docs"
        ]
      },
      {
        "number": 27465,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: error loading model: device CUDA0 does not support split buffers when using split-mode row or tensor",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-21T01:16:05Z",
        "updated_at": "2026-08-21T01:16:05Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27465",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27466,
        "kind": "pr",
        "state": "open",
        "title": "ROCm: add radix TOP_K for long rows",
        "labels": [
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-08-21T01:38:50Z",
        "updated_at": "2026-08-26T08:36:11Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27466",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing"
        ]
      },
      {
        "number": 27467,
        "kind": "issue",
        "state": "open",
        "title": "Bug: --split-mode tensor on CUDA dual-GPU does not fully offload model weights + sampling falls back to CPU (regression between 9ee9fc04 and a302733)",
        "labels": [],
        "created_at": "2026-08-21T02:39:31Z",
        "updated_at": "2026-08-21T02:39:31Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27467",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27468,
        "kind": "pr",
        "state": "open",
        "title": "sycl: bind the f16 KV cache in place for the oneDNN SDPA path",
        "labels": [
          "ggml",
          "SYCL"
        ],
        "created_at": "2026-08-21T04:33:59Z",
        "updated_at": "2026-08-26T15:31:28Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27468",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 27469,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Expose Speculative Decoding (MTP / Draft) in public C API (llama.h) for downstream bindings",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-08-21T05:26:29Z",
        "updated_at": "2026-08-21T19:58:43Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27469",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27470,
        "kind": "pr",
        "state": "open",
        "title": "SVE porting of q4_k gemv kernel",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-08-21T06:36:44Z",
        "updated_at": "2026-08-24T04:15:12Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27470",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing"
        ]
      },
      {
        "number": 27471,
        "kind": "pr",
        "state": "open",
        "title": "vulkan: add f16 B-type matmul pipelines and warp tile size tuning for Intel coopmat1",
        "labels": [
          "Vulkan",
          "ggml"
        ],
        "created_at": "2026-08-21T06:38:12Z",
        "updated_at": "2026-08-26T14:12:01Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27471",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27472,
        "kind": "pr",
        "state": "open",
        "title": "SVE 256 bit implementation of gemm_q6_K_8x8_q8_K kernel",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-08-21T07:20:44Z",
        "updated_at": "2026-08-21T12:09:17Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27472",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27473,
        "kind": "issue",
        "state": "open",
        "title": "Compile bug: Metal tensor API silently inert on M5+/A19+ — MTLCompileOptions.languageVersion never set",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-21T07:25:29Z",
        "updated_at": "2026-08-26T19:30:35Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27473",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27475,
        "kind": "pr",
        "state": "open",
        "title": "common: set Muse Glimmer thinking tags in common_chat_params",
        "labels": [],
        "created_at": "2026-08-21T08:11:49Z",
        "updated_at": "2026-08-25T13:26:09Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27475",
        "categories": [
          "models_architectures",
          "memory_kv_context"
        ]
      },
      {
        "number": 27478,
        "kind": "pr",
        "state": "open",
        "title": "ggml : speed up batch-1 CPU decode, align large allocations",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-08-21T08:37:58Z",
        "updated_at": "2026-08-21T14:07:59Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27478",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27479,
        "kind": "issue",
        "state": "open",
        "title": "Compile bug:  Unable to build SYCL build 10556, commit ff14356e0",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-21T08:45:56Z",
        "updated_at": "2026-08-23T02:26:26Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27479",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27481,
        "kind": "issue",
        "state": "open",
        "title": "server: DELETE resumable stream can abort child or fail to cancel before first token",
        "labels": [],
        "created_at": "2026-08-21T09:17:23Z",
        "updated_at": "2026-08-21T09:17:23Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27481",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27482,
        "kind": "pr",
        "state": "open",
        "title": "server: cancel resumable streams before first token",
        "labels": [
          "server"
        ],
        "created_at": "2026-08-21T09:18:05Z",
        "updated_at": "2026-08-21T09:38:23Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27482",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27483,
        "kind": "pr",
        "state": "open",
        "title": "misc : prevent RAM peaking at model loading stage",
        "labels": [],
        "created_at": "2026-08-21T09:30:38Z",
        "updated_at": "2026-08-21T09:30:39Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27483",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "multimodal_embedding",
          "loader_conversion"
        ]
      },
      {
        "number": 27487,
        "kind": "pr",
        "state": "open",
        "title": "llama : return 0 on invalid ON_DEVICE seq restore",
        "labels": [
          "testing"
        ],
        "created_at": "2026-08-21T11:03:27Z",
        "updated_at": "2026-08-24T12:54:08Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27487",
        "categories": [
          "models_architectures",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 27489,
        "kind": "pr",
        "state": "open",
        "title": "ggml : reuse compute buffers for MTP (#27282)",
        "labels": [
          "testing",
          "ggml"
        ],
        "created_at": "2026-08-21T11:10:29Z",
        "updated_at": "2026-08-21T11:45:21Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27489",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27491,
        "kind": "pr",
        "state": "open",
        "title": "SVE 128 bit Implementation of gemm_q4_k_8x8_q8_k kernel",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-08-21T11:30:49Z",
        "updated_at": "2026-08-21T11:43:41Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27491",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing"
        ]
      },
      {
        "number": 27493,
        "kind": "pr",
        "state": "open",
        "title": "Optimize Krea Vulkan - MUL_MAT",
        "labels": [
          "testing",
          "Vulkan",
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-08-21T11:35:23Z",
        "updated_at": "2026-08-21T14:13:42Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27493",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing"
        ]
      },
      {
        "number": 27494,
        "kind": "pr",
        "state": "open",
        "title": "Optimize Krea Vulkan - Flash Attention",
        "labels": [
          "testing",
          "Vulkan",
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-08-21T11:36:40Z",
        "updated_at": "2026-08-22T04:27:21Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27494",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing"
        ]
      },
      {
        "number": 27495,
        "kind": "pr",
        "state": "open",
        "title": "Optimize Krea Vulkan - fusion and misc. changes",
        "labels": [
          "testing",
          "Vulkan",
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-08-21T11:37:45Z",
        "updated_at": "2026-08-21T13:46:48Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27495",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing"
        ]
      },
      {
        "number": 27499,
        "kind": "issue",
        "state": "open",
        "title": "Research: Parent conditioned drafting trees (PCTree) for DSpark",
        "labels": [
          "research 🔬"
        ],
        "created_at": "2026-08-21T12:02:37Z",
        "updated_at": "2026-08-21T15:17:01Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27499",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 27501,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: support for n-cpu-moe in fit",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-08-21T12:21:24Z",
        "updated_at": "2026-08-21T12:21:24Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27501",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 27505,
        "kind": "issue",
        "state": "open",
        "title": "server: deadlock under sustained chat workload on Gemma 4 26B-A4B + Vulkan (regression between v194 and v0.2.0-dev)",
        "labels": [],
        "created_at": "2026-08-21T14:42:29Z",
        "updated_at": "2026-08-22T07:54:51Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27505",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27506,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: [ROCm] Severe PPL explosion starting from b10040",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-21T14:50:03Z",
        "updated_at": "2026-08-26T12:12:12Z",
        "closed_at": null,
        "comments": 8,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27506",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27512,
        "kind": "pr",
        "state": "open",
        "title": "Quant: OCP FP8 E4M3 support",
        "labels": [
          "testing",
          "ggml",
          "conversion"
        ],
        "created_at": "2026-08-21T16:52:21Z",
        "updated_at": "2026-08-22T08:02:58Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27512",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 27514,
        "kind": "pr",
        "state": "open",
        "title": "common: handle empty forced_tokens gracefully in reasoning-budget sampler",
        "labels": [
          "testing"
        ],
        "created_at": "2026-08-21T18:01:24Z",
        "updated_at": "2026-08-21T18:06:39Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27514",
        "categories": [
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27515,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: --context-shift --no-kv-offload crashes",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-21T18:23:53Z",
        "updated_at": "2026-08-22T08:48:48Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27515",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27517,
        "kind": "issue",
        "state": "open",
        "title": "SYCL: MoE expert tensors never get the Q8_0 reorder path (opt_for_reorder_id)",
        "labels": [],
        "created_at": "2026-08-21T19:11:47Z",
        "updated_at": "2026-08-24T14:17:42Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27517",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27522,
        "kind": "pr",
        "state": "open",
        "title": "server: observe resumable stream cancellation during first result wait (#27481)",
        "labels": [
          "server"
        ],
        "created_at": "2026-08-21T23:00:32Z",
        "updated_at": "2026-08-22T00:02:34Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27522",
        "categories": [
          "native_backends",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 27523,
        "kind": "pr",
        "state": "open",
        "title": "cuda : fuse RWKV7 recurrent input and state paths",
        "labels": [
          "testing",
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-08-21T23:38:58Z",
        "updated_at": "2026-08-25T06:43:02Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27523",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27526,
        "kind": "pr",
        "state": "open",
        "title": "vulkan: combine duplicated fastdiv functions, rename the one optimizing small divs",
        "labels": [
          "Vulkan",
          "ggml"
        ],
        "created_at": "2026-08-22T02:21:34Z",
        "updated_at": "2026-08-22T04:32:03Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27526",
        "categories": [
          "native_backends",
          "models_architectures"
        ]
      },
      {
        "number": 27527,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Checkpoints not happening consistently",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-22T03:32:39Z",
        "updated_at": "2026-08-22T03:32:39Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27527",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27530,
        "kind": "pr",
        "state": "open",
        "title": "llama : fix K/V and recurrent state cleanup after failed restores",
        "labels": [
          "testing"
        ],
        "created_at": "2026-08-22T04:54:52Z",
        "updated_at": "2026-08-24T15:10:19Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27530",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing"
        ]
      },
      {
        "number": 27532,
        "kind": "issue",
        "state": "open",
        "title": "WebUI Edit LLM responses (Payload / KV cache manipulation )",
        "labels": [
          "server/ui"
        ],
        "created_at": "2026-08-22T05:53:20Z",
        "updated_at": "2026-08-25T15:21:40Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27532",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 27536,
        "kind": "issue",
        "state": "open",
        "title": "Compile bug: loop not unrolled ROCm warnings",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-22T11:11:29Z",
        "updated_at": "2026-08-22T11:12:25Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27536",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27537,
        "kind": "pr",
        "state": "open",
        "title": "server: return prompt-token logprobs when echo=true (#27174)",
        "labels": [
          "server"
        ],
        "created_at": "2026-08-22T11:40:36Z",
        "updated_at": "2026-08-25T14:56:44Z",
        "closed_at": null,
        "comments": 12,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27537",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27543,
        "kind": "issue",
        "state": "open",
        "title": "server: PEG chat parser returns HTTP 500 on complete generations containing invalid UTF-8 (no content fallback, unlike legacy parser)",
        "labels": [],
        "created_at": "2026-08-22T14:02:18Z",
        "updated_at": "2026-08-22T14:02:18Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27543",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "build_platform_ops"
        ]
      },
      {
        "number": 27544,
        "kind": "issue",
        "state": "open",
        "title": "spec + vulkan: Performance Drop with MTP n-max > 1 and -np > 1",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-22T14:24:28Z",
        "updated_at": "2026-08-23T19:16:32Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27544",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27546,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: OpenVINO on i5 1345u GPU throws exception",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-22T14:33:08Z",
        "updated_at": "2026-08-25T04:53:21Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27546",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27547,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: SYCL dual-GPU stuck during model load",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-22T14:36:25Z",
        "updated_at": "2026-08-24T19:04:02Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27547",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 27549,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: llama.cpp\\llama.cpp\\ggml\\src\\ggml-cuda\\fattn.cu:574: fatal error",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-22T15:32:36Z",
        "updated_at": "2026-08-25T11:06:16Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27549",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27550,
        "kind": "pr",
        "state": "open",
        "title": "[SYCL] enhance the api to support peer-to-peer copy",
        "labels": [
          "ggml",
          "SYCL"
        ],
        "created_at": "2026-08-22T16:00:26Z",
        "updated_at": "2026-08-24T14:33:59Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27550",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "api_ui_docs"
        ]
      },
      {
        "number": 27553,
        "kind": "issue",
        "state": "open",
        "title": "Vulkan dense prefill on gfx1151 (RDNA3 iGPU): ~1.8x from a larger K-quant int-mmq tile",
        "labels": [],
        "created_at": "2026-08-22T17:05:06Z",
        "updated_at": "2026-08-23T07:31:07Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27553",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 27556,
        "kind": "issue",
        "state": "open",
        "title": "HIP backend silently corrupts Qwen3.5-27B (Gated DeltaNet) inference on gfx1151 - oldest context lost; Vulkan correct at identical commit",
        "labels": [],
        "created_at": "2026-08-22T17:35:12Z",
        "updated_at": "2026-08-26T04:36:47Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27556",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27557,
        "kind": "issue",
        "state": "open",
        "title": "HIP/gfx1151: SOLVE_TRI aborts with rocBLAS 'hipErrorInvalidImage' (device kernel image is invalid)",
        "labels": [],
        "created_at": "2026-08-22T17:35:19Z",
        "updated_at": "2026-08-22T17:35:19Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27557",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27558,
        "kind": "pr",
        "state": "open",
        "title": "HIP: Expand Q5_K and Q6_K tile widths for RDNA2",
        "labels": [
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-08-22T17:41:04Z",
        "updated_at": "2026-08-25T05:50:38Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27558",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 27559,
        "kind": "pr",
        "state": "open",
        "title": "[SYCL] Support limit max alloc memory within 2GB for host-pinned memory",
        "labels": [
          "documentation",
          "ggml",
          "merge ready",
          "SYCL"
        ],
        "created_at": "2026-08-22T17:49:44Z",
        "updated_at": "2026-08-24T14:34:33Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27559",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 27560,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: llama-server crashes (0xC0000005) on Windows/Vulkan with default ctx-checkpoints (Qwen3.8-27B)",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-22T18:32:30Z",
        "updated_at": "2026-08-22T18:32:30Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27560",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27562,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Just-in-time MoE expert streaming from storage",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-08-22T18:42:25Z",
        "updated_at": "2026-08-22T18:50:07Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27562",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 27563,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Show Model Info dialog for unloaded models in Router Mode",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-08-22T19:14:20Z",
        "updated_at": "2026-08-22T19:15:42Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27563",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27564,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Model navigation and load controls in Router Model Info dialog (Follow-up to #27563)",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-08-22T19:28:15Z",
        "updated_at": "2026-08-22T19:28:15Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27564",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27567,
        "kind": "pr",
        "state": "open",
        "title": "UI: allow mic input when an audio model is loaded",
        "labels": [
          "server/ui"
        ],
        "created_at": "2026-08-22T19:59:43Z",
        "updated_at": "2026-08-26T17:34:23Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27567",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27568,
        "kind": "pr",
        "state": "open",
        "title": "metal: fix flash-attention unroll heuristic at DK=512",
        "labels": [
          "ggml",
          "Apple Metal"
        ],
        "created_at": "2026-08-22T20:02:48Z",
        "updated_at": "2026-08-22T20:07:34Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27568",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27569,
        "kind": "pr",
        "state": "open",
        "title": "speculative: cap the draft context batch instead of inheriting the target's",
        "labels": [],
        "created_at": "2026-08-22T20:05:25Z",
        "updated_at": "2026-08-22T20:10:15Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27569",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "build_platform_ops"
        ]
      },
      {
        "number": 27571,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: --reasoning-budget like argument to control the budget based on the conversation length",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-08-22T20:26:45Z",
        "updated_at": "2026-08-23T09:23:40Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27571",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 27572,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: [draft-mtp] Draft acceptance collapses to 0.0 under `-np N` with multi-ubatch batches — async `t_h_nextn` device→host copy race",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-22T20:32:03Z",
        "updated_at": "2026-08-26T21:40:15Z",
        "closed_at": null,
        "comments": 11,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27572",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 27577,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: `-sm tensor` crashes with CUDA error on Pascal (sm_61) during graph compute — loads fine, all GPU subsets affected",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-22T22:28:25Z",
        "updated_at": "2026-08-22T22:28:25Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27577",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27578,
        "kind": "pr",
        "state": "open",
        "title": "reasoning-budget: opt-in soft wrap-up hint and bounded grace with hard termination cap",
        "labels": [
          "testing"
        ],
        "created_at": "2026-08-22T22:32:31Z",
        "updated_at": "2026-08-22T22:43:02Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27578",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27579,
        "kind": "issue",
        "state": "open",
        "title": "HIP/ROCm backend produces corrupted output on gfx1151: two dense architectures fail, Vulkan is correct with byte-identical flags",
        "labels": [],
        "created_at": "2026-08-22T22:40:59Z",
        "updated_at": "2026-08-26T08:58:02Z",
        "closed_at": null,
        "comments": 10,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27579",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27580,
        "kind": "issue",
        "state": "open",
        "title": "json_schema_to_grammar: regex shorthand classes (\\d, \\w, \\s) in \"pattern\" are copied verbatim into GBNF, producing an unparseable grammar with no error",
        "labels": [],
        "created_at": "2026-08-23T02:25:06Z",
        "updated_at": "2026-08-23T02:25:06Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27580",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "build_platform_ops"
        ]
      },
      {
        "number": 27581,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: llama-server container ignores SIGINT when downloading models from huggingface",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-23T03:05:59Z",
        "updated_at": "2026-08-23T03:05:59Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27581",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27583,
        "kind": "pr",
        "state": "open",
        "title": "llama : add StreamingLLM-style KV eviction for bounded long-context memory",
        "labels": [
          "documentation"
        ],
        "created_at": "2026-08-23T04:07:08Z",
        "updated_at": "2026-08-26T19:29:13Z",
        "closed_at": null,
        "comments": 9,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27583",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 27584,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: efficient MoE serving with bandwidth-adaptive CPU–GPU co-execution ( q ⋆ policy), full-layer double-buffered prefill streaming, global LRU expert caching, graph-compatible execution",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-08-23T04:30:37Z",
        "updated_at": "2026-08-23T15:38:22Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27584",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27585,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: rpc tensor caching is  working but not using",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-23T05:31:20Z",
        "updated_at": "2026-08-23T05:31:37Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27585",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 27587,
        "kind": "issue",
        "state": "open",
        "title": "Video input > ~10 s hangs llama-server forever, no response and no error (deadlock in mmproj video probe)",
        "labels": [],
        "created_at": "2026-08-23T06:11:15Z",
        "updated_at": "2026-08-23T19:00:34Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27587",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27589,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cpu: add TQ2_0 8x8 blocked GEMM with i8mm (SMMLA)",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-08-23T07:48:35Z",
        "updated_at": "2026-08-23T07:53:39Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27589",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27590,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cpu: add AVX-512 and VNNI paths for Q5_K/Q6_K dot products",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-08-23T07:49:49Z",
        "updated_at": "2026-08-23T19:32:51Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27590",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27592,
        "kind": "pr",
        "state": "open",
        "title": "reasoning: support fractional --reasoning-budget (fixes #27571)",
        "labels": [
          "server"
        ],
        "created_at": "2026-08-23T09:21:39Z",
        "updated_at": "2026-08-23T09:55:23Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27592",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 27595,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: loading SYCL with --fit doesn't fully account for memory usage, causing runtime overflow",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-23T10:14:14Z",
        "updated_at": "2026-08-24T14:20:31Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27595",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27597,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: GBNF parser rejects escaped hyphen (`\\-`) in character classes",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-23T11:45:10Z",
        "updated_at": "2026-08-23T11:52:01Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27597",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27604,
        "kind": "issue",
        "state": "open",
        "title": "llama-server (Vulkan): server-wide hang after client aborts a streaming request while another request is in flight",
        "labels": [],
        "created_at": "2026-08-23T12:55:34Z",
        "updated_at": "2026-08-23T12:55:34Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27604",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27607,
        "kind": "pr",
        "state": "open",
        "title": "examples/funasr: add Fun-ASR-Nano single-binary ASR CLI (CPU + CUDA)",
        "labels": [
          "examples"
        ],
        "created_at": "2026-08-23T13:35:46Z",
        "updated_at": "2026-08-23T23:37:52Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27607",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 27610,
        "kind": "pr",
        "state": "open",
        "title": "sycl: fuse rms_norm+mul+add and add+add residual chains",
        "labels": [
          "documentation",
          "testing",
          "ggml",
          "SYCL"
        ],
        "created_at": "2026-08-23T14:12:32Z",
        "updated_at": "2026-08-24T15:00:47Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27610",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27612,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: QWEN3.8:27b + lemonade server's rocm b10472  + cline (vscode) = tools partially working (MCP failures), while vulkan works flawlessly",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-23T15:02:14Z",
        "updated_at": "2026-08-26T13:02:32Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27612",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27613,
        "kind": "issue",
        "state": "open",
        "title": "server: generation terminates mid-tag (finish=stop) when reasoning stream contains DSML closing fragments — DeepSeek-V4-Flash, peg-native, long context",
        "labels": [],
        "created_at": "2026-08-23T16:13:05Z",
        "updated_at": "2026-08-23T16:13:05Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27613",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27616,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Copying prompt's KV-cache of the slots that are in use.",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-08-23T17:27:02Z",
        "updated_at": "2026-08-23T17:31:33Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27616",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27618,
        "kind": "pr",
        "state": "open",
        "title": "speculative: draft-mtp for dots3note",
        "labels": [
          "model"
        ],
        "created_at": "2026-08-23T17:48:37Z",
        "updated_at": "2026-08-25T16:42:36Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27618",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "correctness_testing",
          "training_adapters"
        ]
      },
      {
        "number": 27619,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: llama-server tool-calling grammar crashes with \"Unexpected empty grammar stack\" depending only on parameter description text",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-23T17:57:18Z",
        "updated_at": "2026-08-24T05:55:12Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27619",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 27621,
        "kind": "pr",
        "state": "open",
        "title": "CUDA: extend MOE fusion to specdec, earlier MOE glu fusion and topk-router fusion were restricted to 1 token",
        "labels": [
          "testing",
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-08-23T18:22:38Z",
        "updated_at": "2026-08-25T17:48:57Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27621",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27622,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Multiple Model Names, One LLM",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-08-23T19:17:21Z",
        "updated_at": "2026-08-25T03:20:10Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27622",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27623,
        "kind": "issue",
        "state": "open",
        "title": "Qwen3.8-27B (hybrid Gated DeltaNet): decode throughput collapses ~25x at context positions >~80K while prompt processing stays fast",
        "labels": [],
        "created_at": "2026-08-23T19:25:46Z",
        "updated_at": "2026-08-26T06:32:36Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27623",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27624,
        "kind": "pr",
        "state": "open",
        "title": "server: clear stale state when reusing LRU-selected slots",
        "labels": [
          "server"
        ],
        "created_at": "2026-08-23T19:34:42Z",
        "updated_at": "2026-08-23T19:53:04Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27624",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27625,
        "kind": "pr",
        "state": "open",
        "title": "model : add support for HrmTextForCausalLM (DFM Mimir 1B)",
        "labels": [
          "model",
          "testing",
          "conversion"
        ],
        "created_at": "2026-08-23T19:56:21Z",
        "updated_at": "2026-08-24T20:46:42Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27625",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27627,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: webui: attaching video file can wipe IndexedDB chat history",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-23T21:04:33Z",
        "updated_at": "2026-08-23T21:04:33Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27627",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27628,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Reasoning Canceled When Reading Image Tags From File",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-23T21:50:13Z",
        "updated_at": "2026-08-23T21:50:13Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27628",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27629,
        "kind": "pr",
        "state": "open",
        "title": "sycl: make --fit respect --fit-target better",
        "labels": [
          "documentation",
          "ggml",
          "SYCL"
        ],
        "created_at": "2026-08-23T22:42:26Z",
        "updated_at": "2026-08-25T08:25:25Z",
        "closed_at": null,
        "comments": 6,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27629",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 27630,
        "kind": "pr",
        "state": "open",
        "title": "opencl: fix several bugs where the backend aborts",
        "labels": [
          "ggml",
          "OpenCL"
        ],
        "created_at": "2026-08-23T22:46:09Z",
        "updated_at": "2026-08-23T22:46:25Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27630",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27631,
        "kind": "pr",
        "state": "open",
        "title": "sycl: attribute device allocations by site (GGML_SYCL_MEMTRACE)",
        "labels": [
          "documentation",
          "ggml",
          "SYCL"
        ],
        "created_at": "2026-08-23T22:53:11Z",
        "updated_at": "2026-08-24T11:16:42Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27631",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "api_ui_docs"
        ]
      },
      {
        "number": 27632,
        "kind": "pr",
        "state": "open",
        "title": "opencl: fix out‐of‐contract reads in the Adreno image kernels",
        "labels": [
          "ggml",
          "OpenCL"
        ],
        "created_at": "2026-08-23T23:06:26Z",
        "updated_at": "2026-08-23T23:06:44Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27632",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27633,
        "kind": "pr",
        "state": "open",
        "title": "opencl: extend the elementwise and data‐movement op coverage",
        "labels": [
          "ggml",
          "OpenCL"
        ],
        "created_at": "2026-08-23T23:15:11Z",
        "updated_at": "2026-08-23T23:15:32Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27633",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27634,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Vulkan on Intel iGPU / i915: kernel request watchdog silently cancels queued submissions, embeddings collapse with no error",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-23T23:29:26Z",
        "updated_at": "2026-08-26T18:26:27Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27634",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27635,
        "kind": "pr",
        "state": "open",
        "title": "opencl: use right-sized matmul kernels for speculative decoding's verification step",
        "labels": [
          "ggml",
          "OpenCL"
        ],
        "created_at": "2026-08-23T23:36:42Z",
        "updated_at": "2026-08-23T23:37:19Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27635",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27636,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: example: single-binary Fun-ASR-Nano ASR (CPU + CUDA)",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-08-23T23:36:52Z",
        "updated_at": "2026-08-23T23:36:52Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27636",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27637,
        "kind": "pr",
        "state": "open",
        "title": "opencl: choose the MoE expert matmul by batch size for speculative decoding/MTP",
        "labels": [
          "ggml",
          "OpenCL"
        ],
        "created_at": "2026-08-23T23:50:59Z",
        "updated_at": "2026-08-23T23:51:14Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27637",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "build_platform_ops"
        ]
      },
      {
        "number": 27638,
        "kind": "issue",
        "state": "open",
        "title": "Vulkan (Intel Mesa/ANV): GATED_DELTA_NET (KDA) prompt processing degrades to ~O(N²) and triggers device lost above -ub 2048 — bailingmoe3 (Ling-3.0-tiny)",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-24T00:28:36Z",
        "updated_at": "2026-08-24T14:44:52Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27638",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27640,
        "kind": "pr",
        "state": "open",
        "title": "opencl: use a better matmul path on two Adreno GPU generations",
        "labels": [
          "ggml",
          "OpenCL"
        ],
        "created_at": "2026-08-24T02:45:56Z",
        "updated_at": "2026-08-24T02:46:13Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27640",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27641,
        "kind": "pr",
        "state": "open",
        "title": "tests: add mul_mat, flash-attention and copy shapes that real models emit",
        "labels": [
          "testing"
        ],
        "created_at": "2026-08-24T03:10:25Z",
        "updated_at": "2026-08-24T03:10:42Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27641",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27643,
        "kind": "pr",
        "state": "open",
        "title": "json-schema : handle escaped slashes in regex patterns",
        "labels": [
          "testing"
        ],
        "created_at": "2026-08-24T05:41:37Z",
        "updated_at": "2026-08-24T05:46:12Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27643",
        "categories": [
          "models_architectures",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing"
        ]
      },
      {
        "number": 27646,
        "kind": "pr",
        "state": "open",
        "title": "ggml-hexagon: add HTP unary ops for ABS and LOG",
        "labels": [
          "testing",
          "ggml",
          "Hexagon"
        ],
        "created_at": "2026-08-24T07:14:49Z",
        "updated_at": "2026-08-26T17:15:32Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27646",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing"
        ]
      },
      {
        "number": 27647,
        "kind": "pr",
        "state": "open",
        "title": "llama : fix NaN default tensor split when devices report zero free memory",
        "labels": [],
        "created_at": "2026-08-24T07:18:41Z",
        "updated_at": "2026-08-24T07:24:39Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27647",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "speculative_generation",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27648,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: llama-finetune crashes with 'failed to allocate buffer of size 18446744073709547520' before training starts (all model precisions)",
        "labels": [],
        "created_at": "2026-08-24T08:18:22Z",
        "updated_at": "2026-08-25T07:30:10Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27648",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 27649,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Performance regression: mtmd vision encode ~2.7x slower on master (b10516 -> a130532) with CPU-only Android build",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-24T08:26:04Z",
        "updated_at": "2026-08-24T08:26:04Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27649",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27650,
        "kind": "pr",
        "state": "open",
        "title": "ui: edit assistant messages through their stored fields",
        "labels": [
          "server/ui"
        ],
        "created_at": "2026-08-24T08:33:50Z",
        "updated_at": "2026-08-25T15:41:36Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27650",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "api_ui_docs"
        ]
      },
      {
        "number": 27652,
        "kind": "pr",
        "state": "open",
        "title": "vulkan : add GGML_VK_MMV_MAX_COLS override for mul_mat_vec dispatch (#25356)",
        "labels": [
          "Vulkan",
          "ggml"
        ],
        "created_at": "2026-08-24T09:03:51Z",
        "updated_at": "2026-08-24T15:24:42Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27652",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 27653,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Transfer of model data to RPC node is very slow.",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-24T09:50:09Z",
        "updated_at": "2026-08-24T09:54:40Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27653",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27656,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: add SYCL / oneAPI detection and prebuilt install support",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-08-24T10:47:26Z",
        "updated_at": "2026-08-24T14:11:41Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27656",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27657,
        "kind": "pr",
        "state": "open",
        "title": "llama-model: fix NaN device splits when no memory info is available (Windows load crash)",
        "labels": [],
        "created_at": "2026-08-24T11:47:19Z",
        "updated_at": "2026-08-24T11:52:31Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27657",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27662,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Q7 quants",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-08-24T15:05:39Z",
        "updated_at": "2026-08-24T15:05:39Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27662",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 27663,
        "kind": "pr",
        "state": "open",
        "title": "ggml-cuda : fix cpy transposed path corrupting non-contiguous dst",
        "labels": [
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-08-24T15:12:21Z",
        "updated_at": "2026-08-24T15:12:54Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27663",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27664,
        "kind": "pr",
        "state": "open",
        "title": "ui: bug fix, chat bar overflows on mobile during reasoning",
        "labels": [
          "server/ui"
        ],
        "created_at": "2026-08-24T15:28:59Z",
        "updated_at": "2026-08-24T16:00:45Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27664",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 27665,
        "kind": "issue",
        "state": "open",
        "title": "WebUI: New 'Browser' tools are not opt-in, dispite the '--tools' CLI argument being omitted",
        "labels": [],
        "created_at": "2026-08-24T15:50:03Z",
        "updated_at": "2026-08-24T15:50:03Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27665",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27667,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: RPC + Metal: main node retains the full model with default mmap load mode, capping split models to a single node's memory",
        "labels": [],
        "created_at": "2026-08-24T16:17:43Z",
        "updated_at": "2026-08-26T18:09:36Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27667",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27669,
        "kind": "pr",
        "state": "open",
        "title": "rpc: add timestamps to rpc logs",
        "labels": [
          "examples",
          "ggml"
        ],
        "created_at": "2026-08-24T17:21:18Z",
        "updated_at": "2026-08-24T20:13:27Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27669",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27670,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Windows HIP gfx1201: Qwen3.6-35B-A3B NVFP4 loads, then first MUL_MAT fails with ROCm invalid argument",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-24T18:03:59Z",
        "updated_at": "2026-08-25T10:25:12Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27670",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27673,
        "kind": "pr",
        "state": "open",
        "title": "opencl: route the FA prefill split kernels through the per-kernel workgroup guard (a bug exposed in Adreno 660)",
        "labels": [
          "ggml",
          "OpenCL"
        ],
        "created_at": "2026-08-24T19:45:34Z",
        "updated_at": "2026-08-24T20:48:48Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27673",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27674,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: assertion failed in launch_fattn",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-24T20:00:35Z",
        "updated_at": "2026-08-24T20:00:35Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27674",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27677,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: Hexagon session fails when /vendor/lib64 is on LD_LIBRARY_PATH (Termux, SM8850)",
        "labels": [],
        "created_at": "2026-08-24T22:53:21Z",
        "updated_at": "2026-08-24T22:53:21Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27677",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27678,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: CUDA FlashAttention synccheck reports divergent __syncthreads",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-24T23:34:35Z",
        "updated_at": "2026-08-25T12:18:25Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27678",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27680,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Possible regression: DSV4 CUDA compute buffer size",
        "labels": [],
        "created_at": "2026-08-25T00:13:33Z",
        "updated_at": "2026-08-25T19:30:29Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27680",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 27682,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: OpenVINO: low tg",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-25T01:02:26Z",
        "updated_at": "2026-08-25T01:03:37Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27682",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27683,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: broken, Garbage output from Qwen3-2507 4B",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-25T01:16:08Z",
        "updated_at": "2026-08-25T01:16:08Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27683",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 27684,
        "kind": "pr",
        "state": "open",
        "title": "fix(gguf_filename_is_model): avoid excluding valid primary GGUFs with speculative-role keywords",
        "labels": [
          "testing"
        ],
        "created_at": "2026-08-25T01:28:28Z",
        "updated_at": "2026-08-25T01:38:48Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27684",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27685,
        "kind": "issue",
        "state": "open",
        "title": "Research: FreeToken is faster than llama.cpp",
        "labels": [
          "research 🔬"
        ],
        "created_at": "2026-08-25T03:22:35Z",
        "updated_at": "2026-08-26T05:16:05Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27685",
        "categories": [
          "models_architectures",
          "api_ui_docs"
        ]
      },
      {
        "number": 27687,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: llama-server (HIP/ROCm build) spins 2 threads at 100% CPU when idle in router mode (WSL2)",
        "labels": [],
        "created_at": "2026-08-25T05:28:04Z",
        "updated_at": "2026-08-25T05:28:04Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27687",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27688,
        "kind": "issue",
        "state": "open",
        "title": "Research: Adding model fingerprinting / watermarking for detecting use.",
        "labels": [
          "research 🔬"
        ],
        "created_at": "2026-08-25T06:18:13Z",
        "updated_at": "2026-08-26T16:53:16Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27688",
        "categories": [
          "quantization",
          "models_architectures",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 27689,
        "kind": "pr",
        "state": "open",
        "title": "sycl: fix zero size scratchpad crash and >4GB allocation limit on iGPU",
        "labels": [
          "ggml",
          "SYCL"
        ],
        "created_at": "2026-08-25T06:35:48Z",
        "updated_at": "2026-08-25T10:55:03Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27689",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27690,
        "kind": "pr",
        "state": "open",
        "title": "common : fix unsigned underflow in common_opt_dataset_init (#27648)",
        "labels": [
          "testing"
        ],
        "created_at": "2026-08-25T07:29:00Z",
        "updated_at": "2026-08-25T07:29:18Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27690",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 27692,
        "kind": "pr",
        "state": "open",
        "title": "Speculative prefill",
        "labels": [
          "documentation",
          "testing",
          "examples",
          "server"
        ],
        "created_at": "2026-08-25T07:46:23Z",
        "updated_at": "2026-08-26T13:35:51Z",
        "closed_at": null,
        "comments": 10,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27692",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 27693,
        "kind": "issue",
        "state": "open",
        "title": "mtmd: one-sample audio input causes process abort",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-25T08:00:46Z",
        "updated_at": "2026-08-26T16:08:35Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27693",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27694,
        "kind": "pr",
        "state": "open",
        "title": "Make the drafter probabilistic and the target verify by rejection sampling for simple draft and MTP",
        "labels": [
          "server"
        ],
        "created_at": "2026-08-25T08:15:11Z",
        "updated_at": "2026-08-25T08:39:19Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27694",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "build_platform_ops"
        ]
      },
      {
        "number": 27697,
        "kind": "issue",
        "state": "open",
        "title": "mtmd: audio with a duration that is a multiple of 30s produces one extra silent chunk",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-25T09:39:49Z",
        "updated_at": "2026-08-25T09:39:49Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27697",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27698,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: ROCm MoE models (qwen35moe, gemma4moe) fail to load / segfault on RX 6700 XT after b10108",
        "labels": [],
        "created_at": "2026-08-25T11:25:08Z",
        "updated_at": "2026-08-25T11:25:08Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27698",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27701,
        "kind": "pr",
        "state": "open",
        "title": "metal: Fix memory query under low-memory conditions",
        "labels": [
          "ggml",
          "Apple Metal"
        ],
        "created_at": "2026-08-25T12:37:29Z",
        "updated_at": "2026-08-25T13:52:45Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27701",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27703,
        "kind": "pr",
        "state": "open",
        "title": "vulkan : copy strided f16 KV once on AMD so reads spread across memory channels",
        "labels": [
          "testing",
          "Vulkan",
          "ggml"
        ],
        "created_at": "2026-08-25T13:37:39Z",
        "updated_at": "2026-08-26T14:43:40Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27703",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27704,
        "kind": "pr",
        "state": "open",
        "title": "rpc : do not advertise mmap support",
        "labels": [
          "ggml"
        ],
        "created_at": "2026-08-25T14:30:37Z",
        "updated_at": "2026-08-26T06:12:17Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27704",
        "categories": [
          "models_architectures",
          "scheduling_serving"
        ]
      },
      {
        "number": 27705,
        "kind": "pr",
        "state": "open",
        "title": "Fix/output reorder index space",
        "labels": [
          "testing",
          "server"
        ],
        "created_at": "2026-08-25T15:01:36Z",
        "updated_at": "2026-08-26T20:17:52Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27705",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27707,
        "kind": "pr",
        "state": "open",
        "title": "ci: Drop HF bucket hosting for prebuilt UI in favor of GH Release Artifacts",
        "labels": [
          "build",
          "devops",
          "server/ui"
        ],
        "created_at": "2026-08-25T16:27:53Z",
        "updated_at": "2026-08-26T13:15:38Z",
        "closed_at": null,
        "comments": 5,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27707",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27711,
        "kind": "pr",
        "state": "open",
        "title": "spec: Add benchmark-only synthetic speculative acceptance options",
        "labels": [
          "documentation",
          "testing",
          "server"
        ],
        "created_at": "2026-08-25T16:57:08Z",
        "updated_at": "2026-08-26T19:40:40Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27711",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27715,
        "kind": "pr",
        "state": "open",
        "title": "fix: Prevent Qwen3.6 reranker from allocating exponential memory as physical batch size increases",
        "labels": [
          "model"
        ],
        "created_at": "2026-08-25T18:21:40Z",
        "updated_at": "2026-08-25T18:22:11Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27715",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27717,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: --n-cpu-moe below a threshold crashes MTP draft model load with 'invalid vector subscript' (Nemotron-3.5-Lightning-30B-A3B)",
        "labels": [],
        "created_at": "2026-08-25T20:01:13Z",
        "updated_at": "2026-08-26T02:15:25Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27717",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27719,
        "kind": "pr",
        "state": "open",
        "title": "server : make /v1/models created timestamp stable",
        "labels": [
          "server"
        ],
        "created_at": "2026-08-25T20:16:26Z",
        "updated_at": "2026-08-25T20:21:23Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27719",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 27721,
        "kind": "pr",
        "state": "open",
        "title": "CUDA graphs on Pascal (sm_61): +40% on MoE, +7% on dense, no regression",
        "labels": [
          "ggml",
          "CUDA"
        ],
        "created_at": "2026-08-25T22:49:19Z",
        "updated_at": "2026-08-26T17:27:48Z",
        "closed_at": null,
        "comments": 9,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27721",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27723,
        "kind": "pr",
        "state": "open",
        "title": "ggml-vulkan: allow fp32-only devices (Haswell hasvk)",
        "labels": [
          "Vulkan",
          "ggml"
        ],
        "created_at": "2026-08-25T23:59:47Z",
        "updated_at": "2026-08-26T13:07:59Z",
        "closed_at": null,
        "comments": 4,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27723",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing"
        ]
      },
      {
        "number": 27724,
        "kind": "issue",
        "state": "open",
        "title": "input_video rejects data URLs: \"Failed to load image or audio file\" — data: URLs fall into the raw-base64 branch (accept_base64_uri=false)",
        "labels": [],
        "created_at": "2026-08-26T00:35:01Z",
        "updated_at": "2026-08-26T11:50:48Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27724",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing"
        ]
      },
      {
        "number": 27725,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Memory Leak Issue Report for llama.cpp",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-26T03:31:24Z",
        "updated_at": "2026-08-26T08:15:03Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27725",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27727,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: Garbled output from bartowski/huihui-ai_Qwen3-Coder-Next-abliterated-GGUF running in dockerised llama.cpp build from source for ROCM with gfx1150",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-26T07:47:28Z",
        "updated_at": "2026-08-26T07:51:01Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27727",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 27730,
        "kind": "pr",
        "state": "open",
        "title": "[feat] support dspark (Nanbeige4.2-3B)",
        "labels": [
          "model",
          "merge ready"
        ],
        "created_at": "2026-08-26T09:57:27Z",
        "updated_at": "2026-08-26T10:07:36Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27730",
        "categories": [
          "quantization",
          "models_architectures",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27732,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug:",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-26T10:44:44Z",
        "updated_at": "2026-08-26T10:44:44Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27732",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27733,
        "kind": "issue",
        "state": "open",
        "title": "[server] final peg-native parse failure on trailing \"<think>\" tail discards an entire completed generation (Qwen3.8)",
        "labels": [],
        "created_at": "2026-08-26T11:06:37Z",
        "updated_at": "2026-08-26T11:06:37Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27733",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "build_platform_ops"
        ]
      },
      {
        "number": 27734,
        "kind": "issue",
        "state": "open",
        "title": "Vulkan: ~78% decode throughput cliff at 131072 context — suballocation fragmentation; fixed via GGML_VK_SUBALLOCATION_BLOCK_SIZE=4 GiB",
        "labels": [],
        "created_at": "2026-08-26T11:35:40Z",
        "updated_at": "2026-08-26T11:35:40Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27734",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 27735,
        "kind": "pr",
        "state": "open",
        "title": "server : accept data: URLs for input_video and input_audio",
        "labels": [
          "server"
        ],
        "created_at": "2026-08-26T11:48:17Z",
        "updated_at": "2026-08-26T19:04:40Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27735",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27736,
        "kind": "pr",
        "state": "open",
        "title": "fix: model loader tensor count mismatch when MTP block contains NVFP4 weights with scales / input scales.",
        "labels": [],
        "created_at": "2026-08-26T12:18:48Z",
        "updated_at": "2026-08-26T12:18:49Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27736",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "speculative_generation",
          "loader_conversion",
          "correctness_testing"
        ]
      },
      {
        "number": 27740,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: pvs-studio -report",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-26T13:28:52Z",
        "updated_at": "2026-08-26T13:35:57Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27740",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 27741,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Possible to support Qwen3.8-Flash-Next?",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-08-26T13:29:00Z",
        "updated_at": "2026-08-26T17:37:19Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27741",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 27742,
        "kind": "pr",
        "state": "open",
        "title": "model: add Qwen3.8-Flash-Next (qwen4exp)",
        "labels": [
          "model",
          "testing",
          "conversion"
        ],
        "created_at": "2026-08-26T13:43:12Z",
        "updated_at": "2026-08-26T21:54:03Z",
        "closed_at": null,
        "comments": 62,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27742",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27745,
        "kind": "pr",
        "state": "open",
        "title": "ui: Replace per-conversation MCP overrides with per-conversation tool policy",
        "labels": [
          "server/ui"
        ],
        "created_at": "2026-08-26T13:48:57Z",
        "updated_at": "2026-08-26T19:10:45Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27745",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "api_ui_docs"
        ]
      },
      {
        "number": 27746,
        "kind": "pr",
        "state": "open",
        "title": "ui: Improve Chat Form Actions UI/UX (models selector, add panel)",
        "labels": [
          "server/ui"
        ],
        "created_at": "2026-08-26T13:51:38Z",
        "updated_at": "2026-08-26T19:11:00Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27746",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "api_ui_docs"
        ]
      },
      {
        "number": 27748,
        "kind": "issue",
        "state": "open",
        "title": "Official Windows prebuilt binaries crash with heap corruption (0xC0000374) on Windows 11 Insider build 26220 — LLVM OpenMP vs MSVC OpenMP",
        "labels": [],
        "created_at": "2026-08-26T13:57:11Z",
        "updated_at": "2026-08-26T13:57:11Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27748",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27749,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug:",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-26T14:52:38Z",
        "updated_at": "2026-08-26T14:52:38Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27749",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27750,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: 2-GPU CUDA tensor split is unstable with Qwen3.8-27B on Windows, including with MTP disabled",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-26T14:54:00Z",
        "updated_at": "2026-08-26T14:54:00Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27750",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27751,
        "kind": "pr",
        "state": "open",
        "title": "server : fix Responses API item parsing for reasoning and Codex compatibility",
        "labels": [
          "draft"
        ],
        "created_at": "2026-08-26T15:04:32Z",
        "updated_at": "2026-08-26T15:35:22Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27751",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 27752,
        "kind": "pr",
        "state": "open",
        "title": "model : add GLM-5.3-Flash (glm5next)",
        "labels": [
          "model",
          "testing",
          "conversion"
        ],
        "created_at": "2026-08-26T16:07:23Z",
        "updated_at": "2026-08-26T17:35:37Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27752",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27753,
        "kind": "pr",
        "state": "open",
        "title": "ci : build only the ggml-hip backend for windows-rocm release",
        "labels": [
          "devops"
        ],
        "created_at": "2026-08-26T16:12:22Z",
        "updated_at": "2026-08-26T16:39:26Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27753",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27754,
        "kind": "pr",
        "state": "open",
        "title": "model: add GLM-5-Next (GLM-5.3-Flash)",
        "labels": [
          "model",
          "testing",
          "mtmd",
          "conversion"
        ],
        "created_at": "2026-08-26T16:53:24Z",
        "updated_at": "2026-08-26T17:48:12Z",
        "closed_at": null,
        "comments": 2,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27754",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "speculative_generation",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 27755,
        "kind": "pr",
        "state": "open",
        "title": "tests : run test-save-load-state across all architectures",
        "labels": [
          "model",
          "testing"
        ],
        "created_at": "2026-08-26T17:37:10Z",
        "updated_at": "2026-08-26T19:27:47Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27755",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 27756,
        "kind": "issue",
        "state": "open",
        "title": "Qwen3.5-hybrid 64-layer (Qwen3.8-27B): silent instant-EOS beyond ~130k context on both CUDA and CPU — consistent with DeltaNet recurrent-state depth × layer-count degradation",
        "labels": [],
        "created_at": "2026-08-26T17:41:57Z",
        "updated_at": "2026-08-26T18:51:42Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27756",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27757,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: tensor parallelism and kv offload crashes",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-26T17:53:42Z",
        "updated_at": "2026-08-26T18:17:23Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27757",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27758,
        "kind": "pr",
        "state": "open",
        "title": "metal : fix memory leaks due to missing autoreleasepools",
        "labels": [
          "ggml",
          "Apple Metal"
        ],
        "created_at": "2026-08-26T18:32:11Z",
        "updated_at": "2026-08-26T18:54:43Z",
        "closed_at": null,
        "comments": 3,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27758",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing"
        ]
      },
      {
        "number": 27760,
        "kind": "pr",
        "state": "open",
        "title": "server : report live generation throughput",
        "labels": [
          "server"
        ],
        "created_at": "2026-08-26T19:08:35Z",
        "updated_at": "2026-08-26T20:21:05Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27760",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27761,
        "kind": "issue",
        "state": "open",
        "title": "Misc. bug: ROCm flash attention falls back to 34 graph splits with q5_0 KV cache (pp512 drops ~68%)",
        "labels": [
          "bug-unconfirmed"
        ],
        "created_at": "2026-08-26T19:09:17Z",
        "updated_at": "2026-08-26T19:09:17Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27761",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27763,
        "kind": "issue",
        "state": "open",
        "title": "qwen4exp (Qwen3.8-Flash-Next) garbage output on SM110 when >8 layers on GPU",
        "labels": [],
        "created_at": "2026-08-26T20:13:39Z",
        "updated_at": "2026-08-26T20:13:39Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27763",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27764,
        "kind": "pr",
        "state": "open",
        "title": "chat : split specialized parsers into common/parsers",
        "labels": [],
        "created_at": "2026-08-26T20:17:28Z",
        "updated_at": "2026-08-26T21:43:06Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27764",
        "categories": [
          "models_architectures",
          "structured_generation_tools",
          "build_platform_ops"
        ]
      },
      {
        "number": 27765,
        "kind": "pr",
        "state": "open",
        "title": "vulkan: add TQ1_0 support (mm, mat-vec, mat-vec-id, dequant, get_rows)",
        "labels": [
          "testing",
          "Vulkan",
          "ggml"
        ],
        "created_at": "2026-08-26T20:56:18Z",
        "updated_at": "2026-08-26T21:47:28Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/pull/27765",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 27766,
        "kind": "issue",
        "state": "open",
        "title": "Feature Request: Make mmap vs read decision per tensor instead of per file.",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-08-26T20:58:55Z",
        "updated_at": "2026-08-26T20:59:54Z",
        "closed_at": null,
        "comments": 0,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27766",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 27767,
        "kind": "issue",
        "state": "open",
        "title": "Eval bug: named/required tool_choice not enforced for Qwen3.6 templates when enable_thinking=false",
        "labels": [],
        "created_at": "2026-08-26T21:10:02Z",
        "updated_at": "2026-08-26T21:16:03Z",
        "closed_at": null,
        "comments": 1,
        "url": "https://github.com/ggml-org/llama.cpp/issues/27767",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      }
    ]
  },
  "fak": {
    "repository": "anthony-chaudhary/fak",
    "queries": [
      "\"llama.cpp\"",
      "\"llama cpp\"",
      "llamacpp"
    ],
    "matching_item_count": 357,
    "open_count": 87,
    "closed_count": 270,
    "items": [
      {
        "number": 1,
        "state": "closed",
        "title": "perf(model): AMX/Accelerate prefill path (opt-in cgo) — the only route to llama.cpp CPU prefill parity on Apple Silicon",
        "labels": [
          "compute",
          "model",
          "performance",
          "priority/P2"
        ],
        "created_at": "2026-06-21T20:13:36Z",
        "updated_at": "2026-06-30T07:01:45Z",
        "closed_at": "2026-06-30T07:01:45Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/1",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 2,
        "state": "closed",
        "title": "perf(model): oracle-validate the shipped deferred-reduction NEON Q8 kernels on a real Qwen2.5-1.5B export (argmax-exact)",
        "labels": [
          "compute",
          "model",
          "performance"
        ],
        "created_at": "2026-06-21T20:13:40Z",
        "updated_at": "2026-06-24T19:50:20Z",
        "closed_at": "2026-06-24T19:50:20Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/2",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 6,
        "state": "closed",
        "title": "perf(model): arm64 decode forward-path de-serialization (group q/k/v + gate/up) + oracle-verify the deferred Q8 kernel",
        "labels": [
          "compute",
          "model",
          "performance",
          "priority/P2"
        ],
        "created_at": "2026-06-21T20:13:57Z",
        "updated_at": "2026-07-03T23:26:31Z",
        "closed_at": "2026-07-03T23:26:31Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/6",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 7,
        "state": "closed",
        "title": "fak(ggufload): add gemma3 (post_ffw_norm) + MoE (ffn_*_exps) tensor-name mappings",
        "labels": [
          "enhancement",
          "loader",
          "priority/P2",
          "track/C-gpu-backend"
        ],
        "created_at": "2026-06-21T20:14:01Z",
        "updated_at": "2026-07-03T23:26:28Z",
        "closed_at": "2026-07-03T23:26:28Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/7",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "loader_conversion"
        ]
      },
      {
        "number": 18,
        "state": "closed",
        "title": "bench(gcp): re-run fak-cuda on a real GPU — build bug fixed (01ca1e2) but head-to-head still missing fak's device number",
        "labels": [
          "help wanted",
          "benchmark",
          "cuda",
          "gpu",
          "performance",
          "priority/P1"
        ],
        "created_at": "2026-06-21T20:14:45Z",
        "updated_at": "2026-06-29T16:26:11Z",
        "closed_at": "2026-06-29T16:26:11Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/18",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 28,
        "state": "closed",
        "title": "feat(serving): native prefill/decode role split over the continuous-batching scheduler",
        "labels": [
          "agentic-serving",
          "batching",
          "multi-gpu",
          "performance",
          "priority/P2",
          "track/B-performance"
        ],
        "created_at": "2026-06-21T20:15:25Z",
        "updated_at": "2026-07-03T23:26:24Z",
        "closed_at": "2026-07-03T23:26:24Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/28",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 44,
        "state": "closed",
        "title": "bench(serving): EXTEND the bench harness/webbench into a vLLM/SGLang/native parity harness",
        "labels": [
          "agentic-serving",
          "benchmark",
          "performance",
          "priority/P0",
          "testing",
          "track/E-testing-quality"
        ],
        "created_at": "2026-06-21T20:16:30Z",
        "updated_at": "2026-06-25T05:01:30Z",
        "closed_at": "2026-06-25T05:01:30Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/44",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 49,
        "state": "closed",
        "title": "docs(serving): dual-track RIDE+NATIVE serving decision doc with honest scope + de-dup map",
        "labels": [
          "documentation",
          "agentic-serving",
          "priority/P0",
          "substrate",
          "track/G-foundation"
        ],
        "created_at": "2026-06-21T20:16:50Z",
        "updated_at": "2026-06-22T13:10:49Z",
        "closed_at": "2026-06-22T13:10:49Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/49",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 59,
        "state": "open",
        "title": "track(metal): Apple-Silicon Metal GPU steps — close the M3 Pro SOTA gap [C-001 breakdown]",
        "labels": [
          "help wanted",
          "epic",
          "gpu",
          "metal",
          "performance",
          "priority/P1",
          "track/C-gpu-backend"
        ],
        "created_at": "2026-06-21T20:17:31Z",
        "updated_at": "2026-07-09T00:39:28Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/59",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "build_platform_ops"
        ]
      },
      {
        "number": 60,
        "state": "closed",
        "title": "perf(cpu): close the 4.5x fak-CPU vs llama.cpp-CPU prefill gap (extract-once int8 q4kGemm)",
        "labels": [
          "help wanted",
          "gpu",
          "metal",
          "performance",
          "priority/P1",
          "track/C-gpu-backend"
        ],
        "created_at": "2026-06-21T20:17:35Z",
        "updated_at": "2026-07-03T23:28:37Z",
        "closed_at": "2026-07-03T23:28:37Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/60",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 61,
        "state": "closed",
        "title": "refactor(metal): reconcile the two Metal lanes (compute/metal.go vs metalgemm)",
        "labels": [
          "gpu",
          "metal",
          "performance",
          "priority/P2",
          "track/C-gpu-backend"
        ],
        "created_at": "2026-06-21T20:17:39Z",
        "updated_at": "2026-06-29T10:24:16Z",
        "closed_at": "2026-06-29T10:24:16Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/61",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 62,
        "state": "closed",
        "title": "feat(metal): auto-enable Metal on darwin/arm64 (retire the -tags fakmetal opt-in)",
        "labels": [
          "gpu",
          "metal",
          "performance",
          "priority/P2",
          "track/C-gpu-backend"
        ],
        "created_at": "2026-06-21T20:17:43Z",
        "updated_at": "2026-06-29T17:14:53Z",
        "closed_at": "2026-06-29T17:14:53Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/62",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 63,
        "state": "closed",
        "title": "bench(metal): witness fak-Metal M3 Pro numbers in BENCHMARK-AUTHORITY + experiments",
        "labels": [
          "gpu",
          "metal",
          "performance",
          "priority/P2",
          "track/C-gpu-backend"
        ],
        "created_at": "2026-06-21T20:17:47Z",
        "updated_at": "2026-06-29T00:18:03Z",
        "closed_at": "2026-06-29T00:18:03Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/63",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 64,
        "state": "closed",
        "title": "test(metal): fak-Metal vs llama.cpp-Metal parity + perf gate (toward 51.55/7.29)",
        "labels": [
          "help wanted",
          "gpu",
          "metal",
          "performance",
          "priority/P1",
          "track/C-gpu-backend"
        ],
        "created_at": "2026-06-21T20:17:51Z",
        "updated_at": "2026-07-03T23:28:34Z",
        "closed_at": "2026-07-03T23:28:34Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/64",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 65,
        "state": "closed",
        "title": "feat(metal): Gated-DeltaNet recurrence — GPU kernel vs CPU-hybrid decision",
        "labels": [
          "gpu",
          "metal",
          "performance",
          "priority/P2",
          "track/C-gpu-backend"
        ],
        "created_at": "2026-06-21T20:17:55Z",
        "updated_at": "2026-06-29T15:21:16Z",
        "closed_at": "2026-06-29T15:21:16Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/65",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 67,
        "state": "open",
        "title": "perf(metal): one MTLCommandBuffer per forward pass (async/stream pipelining)",
        "labels": [
          "help wanted",
          "gpu",
          "metal",
          "performance",
          "priority/P1",
          "track/C-gpu-backend"
        ],
        "created_at": "2026-06-21T20:18:03Z",
        "updated_at": "2026-08-26T20:50:46Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/67",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 68,
        "state": "closed",
        "title": "feat(metal): q4_k decode GEMV on Metal",
        "labels": [
          "gpu",
          "metal",
          "performance",
          "priority/P2",
          "track/C-gpu-backend"
        ],
        "created_at": "2026-06-21T20:18:07Z",
        "updated_at": "2026-06-28T11:15:44Z",
        "closed_at": "2026-06-28T11:15:44Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/68",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving"
        ]
      },
      {
        "number": 69,
        "state": "closed",
        "title": "feat(metal): resident weights in unified memory (no per-call upload)",
        "labels": [
          "help wanted",
          "gpu",
          "metal",
          "performance",
          "priority/P1",
          "track/C-gpu-backend"
        ],
        "created_at": "2026-06-21T20:18:11Z",
        "updated_at": "2026-06-29T11:58:04Z",
        "closed_at": "2026-06-29T11:58:04Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/69",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion"
        ]
      },
      {
        "number": 70,
        "state": "closed",
        "title": "feat(metal): q4_k device GEMM — upload q4_k (f16/int8), not Q8-dequant->f32",
        "labels": [
          "help wanted",
          "gpu",
          "metal",
          "performance",
          "priority/P1",
          "track/C-gpu-backend"
        ],
        "created_at": "2026-06-21T20:18:15Z",
        "updated_at": "2026-06-28T22:17:30Z",
        "closed_at": "2026-06-28T22:17:30Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/70",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving"
        ]
      },
      {
        "number": 71,
        "state": "closed",
        "title": "feat(metal): hybrid (Qwen3.6 Gated-DeltaNet) Metal prefill — lift requirePreNorm",
        "labels": [
          "help wanted",
          "gpu",
          "metal",
          "performance",
          "priority/P1",
          "track/C-gpu-backend"
        ],
        "created_at": "2026-06-21T20:18:19Z",
        "updated_at": "2026-06-28T22:28:14Z",
        "closed_at": "2026-06-28T22:28:14Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/71",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 92,
        "state": "closed",
        "title": "examples: consolidated cmd/ diagnostic-cheatsheet for the 13 benchmark/diagnostic tools",
        "labels": [
          "documentation",
          "enhancement",
          "performance"
        ],
        "created_at": "2026-06-21T20:19:45Z",
        "updated_at": "2026-06-27T10:56:17Z",
        "closed_at": "2026-06-27T10:56:17Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/92",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 93,
        "state": "closed",
        "title": "feat(model): AWQ Quantization Support [A-001]",
        "labels": [
          "enhancement",
          "compute",
          "model-support",
          "priority/P0",
          "quantization",
          "track/A-model-support"
        ],
        "created_at": "2026-06-21T20:19:49Z",
        "updated_at": "2026-06-23T01:40:17Z",
        "closed_at": "2026-06-23T01:40:17Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/93",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 96,
        "state": "closed",
        "title": "docs(benchmark-author): several headline numbers are absent from BENCHMARK-AUTHORITY despite its 'single source of truth' mandate",
        "labels": [
          "documentation"
        ],
        "created_at": "2026-06-21T20:20:01Z",
        "updated_at": "2026-06-27T10:56:13Z",
        "closed_at": "2026-06-27T10:56:13Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/96",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 112,
        "state": "closed",
        "title": "docs(qwen36-amd-vulkan): \"7.29 tok/s bar\" silently borrowed from a different machine",
        "labels": [
          "documentation"
        ],
        "created_at": "2026-06-21T20:21:04Z",
        "updated_at": "2026-06-27T10:37:07Z",
        "closed_at": "2026-06-27T10:37:07Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/112",
        "categories": [
          "native_backends",
          "models_architectures",
          "api_ui_docs"
        ]
      },
      {
        "number": 117,
        "state": "closed",
        "title": "docs(model-batching vs llamacpp-headtohead): batched decode peak disagrees (862 vs 2916 tok/s)",
        "labels": [
          "documentation"
        ],
        "created_at": "2026-06-21T20:21:24Z",
        "updated_at": "2026-06-27T10:34:21Z",
        "closed_at": "2026-06-27T10:34:21Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/117",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "api_ui_docs"
        ]
      },
      {
        "number": 118,
        "state": "closed",
        "title": "docs(qwen36-amd-vulkan): \"ratio to M3 Pro bar\" uses two different decode numbers (2.697 vs 2.80)",
        "labels": [
          "documentation"
        ],
        "created_at": "2026-06-21T20:21:28Z",
        "updated_at": "2026-06-27T10:33:00Z",
        "closed_at": "2026-06-27T10:33:00Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/118",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 119,
        "state": "closed",
        "title": "docs(qwen36-amd-vulkan vs qwen36-parity): fak BEATS llama.cpp 1.88–3.15× on AMD but is 0.12× of llama.cpp on M3",
        "labels": [
          "documentation"
        ],
        "created_at": "2026-06-21T20:21:32Z",
        "updated_at": "2026-06-27T10:39:50Z",
        "closed_at": "2026-06-27T10:39:50Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/119",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "api_ui_docs"
        ]
      },
      {
        "number": 122,
        "state": "closed",
        "title": "docs(qwen25-7b vs fak-native-chat): llama.cpp Qwen2.5-7B Metal decode disagrees (17.27 vs 17.6 tok/s)",
        "labels": [
          "documentation"
        ],
        "created_at": "2026-06-21T20:21:44Z",
        "updated_at": "2026-06-27T10:27:03Z",
        "closed_at": "2026-06-27T10:27:02Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/122",
        "categories": [
          "native_backends",
          "models_architectures",
          "api_ui_docs"
        ]
      },
      {
        "number": 124,
        "state": "closed",
        "title": "docs(m3-llamacpp vs fak-native-chat): Qwen2.5-1.5B llama.cpp references disagree ~10%",
        "labels": [
          "documentation"
        ],
        "created_at": "2026-06-21T20:21:52Z",
        "updated_at": "2026-06-27T10:26:14Z",
        "closed_at": "2026-06-27T10:26:14Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/124",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 133,
        "state": "closed",
        "title": "docs(claims): Vulkan-on-RX-7600 throughput row is ~8× stale vs the current committed number",
        "labels": [
          "documentation"
        ],
        "created_at": "2026-06-21T20:22:28Z",
        "updated_at": "2026-06-27T10:35:35Z",
        "closed_at": "2026-06-27T10:35:35Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/133",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 140,
        "state": "closed",
        "title": "docs(hardware-catalog): CUDA Graph described as \"marginal\" but GPU.md credits it with the 16× parity jump",
        "labels": [
          "documentation"
        ],
        "created_at": "2026-06-21T20:22:56Z",
        "updated_at": "2026-06-27T10:11:19Z",
        "closed_at": "2026-06-27T10:11:19Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/140",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 223,
        "state": "closed",
        "title": "docs(documentation): Migration Guides [E-006]",
        "labels": [
          "documentation",
          "migration",
          "priority/P2",
          "track/E-testing-quality"
        ],
        "created_at": "2026-06-21T20:28:24Z",
        "updated_at": "2026-06-24T19:32:57Z",
        "closed_at": "2026-06-24T19:32:57Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/223",
        "categories": [
          "models_architectures",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 269,
        "state": "closed",
        "title": "feat(metal): Metal Backend (Apple Silicon) [C-001]",
        "labels": [
          "enhancement",
          "gpu",
          "metal",
          "priority/P1",
          "track/C-gpu-backend"
        ],
        "created_at": "2026-06-21T20:31:28Z",
        "updated_at": "2026-06-22T18:55:44Z",
        "closed_at": "2026-06-22T18:55:44Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/269",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 275,
        "state": "closed",
        "title": "feat(quant): INT4/INT2 Quantization [B-007]",
        "labels": [
          "enhancement",
          "model",
          "performance",
          "priority/P2",
          "quantization",
          "track/B-performance"
        ],
        "created_at": "2026-06-21T20:31:53Z",
        "updated_at": "2026-06-28T07:51:11Z",
        "closed_at": "2026-06-28T07:51:11Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/275",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 287,
        "state": "closed",
        "title": "perf(vulkan): Vulkan Backend Optimization [B-002]",
        "labels": [
          "help wanted",
          "gpu",
          "performance",
          "priority/P0",
          "track/B-performance",
          "vulkan"
        ],
        "created_at": "2026-06-21T20:32:40Z",
        "updated_at": "2026-06-29T00:27:07Z",
        "closed_at": "2026-06-27T15:11:13Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/287",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 289,
        "state": "closed",
        "title": "perf(prefill): Close Prefill Throughput Gap [B-001]",
        "labels": [
          "help wanted",
          "cuda",
          "gpu",
          "performance",
          "priority/P0",
          "track/B-performance"
        ],
        "created_at": "2026-06-21T20:32:48Z",
        "updated_at": "2026-06-27T15:11:11Z",
        "closed_at": "2026-06-27T15:11:11Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/289",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 292,
        "state": "closed",
        "title": "feat(model): GGUF Format Completion [A-010]",
        "labels": [
          "enhancement",
          "gguf",
          "loader",
          "model-support",
          "priority/P2",
          "track/A-model-support"
        ],
        "created_at": "2026-06-21T20:33:00Z",
        "updated_at": "2026-06-27T14:45:43Z",
        "closed_at": "2026-06-27T14:45:43Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/292",
        "categories": [
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "loader_conversion",
          "build_platform_ops"
        ]
      },
      {
        "number": 298,
        "state": "closed",
        "title": "feat(model): Production Llama 3.x Checkpoints [A-004]",
        "labels": [
          "enhancement",
          "help wanted",
          "llama",
          "model",
          "model-support",
          "priority/P0",
          "track/A-model-support"
        ],
        "created_at": "2026-06-21T20:33:24Z",
        "updated_at": "2026-06-26T08:45:40Z",
        "closed_at": "2026-06-26T08:45:40Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/298",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 299,
        "state": "closed",
        "title": "feat(model): EXL2 Format Support [A-003]",
        "labels": [
          "enhancement",
          "help wanted",
          "model",
          "model-support",
          "priority/P1",
          "quantization",
          "track/A-model-support"
        ],
        "created_at": "2026-06-21T20:33:27Z",
        "updated_at": "2026-06-26T09:00:38Z",
        "closed_at": "2026-06-26T09:00:38Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/299",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "build_platform_ops"
        ]
      },
      {
        "number": 300,
        "state": "closed",
        "title": "feat(model): GPTQ Quantization Support [A-002]",
        "labels": [
          "enhancement",
          "help wanted",
          "model",
          "model-support",
          "priority/P1",
          "quantization",
          "track/A-model-support"
        ],
        "created_at": "2026-06-21T20:33:31Z",
        "updated_at": "2026-06-26T10:15:47Z",
        "closed_at": "2026-06-26T10:15:47Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/300",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 305,
        "state": "open",
        "title": "[epic] Track C - GPU/Backend Parity",
        "labels": [
          "epic",
          "gpu",
          "priority/P2",
          "research",
          "track/C-gpu-backend"
        ],
        "created_at": "2026-06-21T20:33:52Z",
        "updated_at": "2026-06-29T22:34:25Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/305",
        "categories": [
          "native_backends"
        ]
      },
      {
        "number": 329,
        "state": "closed",
        "title": "examples: GPU smoke walkthrough (Vulkan AMD + CUDA NVIDIA — single adoption-shaped entry)",
        "labels": [
          "documentation",
          "enhancement",
          "compute",
          "gpu"
        ],
        "created_at": "2026-06-21T20:35:26Z",
        "updated_at": "2026-06-29T01:37:06Z",
        "closed_at": "2026-06-29T01:37:06Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/329",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 355,
        "state": "closed",
        "title": "docs(readme): \"parity\" used in two senses (bit-exact vs throughput), and the throughput claim tensions with \"doesn't compete on throughput\"",
        "labels": [
          "documentation"
        ],
        "created_at": "2026-06-21T20:37:11Z",
        "updated_at": "2026-06-22T14:15:53Z",
        "closed_at": "2026-06-22T14:15:53Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/355",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 374,
        "state": "closed",
        "title": "docs(readme): headline GPU parity claim (~120 tok/s RTX 4070) absent from the cited 'single source of truth' BENCHMARK-AUTHORITY.md",
        "labels": [
          "documentation"
        ],
        "created_at": "2026-06-21T20:38:28Z",
        "updated_at": "2026-06-22T13:15:49Z",
        "closed_at": "2026-06-22T13:15:49Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/374",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 378,
        "state": "closed",
        "title": "gpu(cuda): WSL2 missing CUDA runtime libraries for Q8 quantization",
        "labels": [
          "bug",
          "gpu"
        ],
        "created_at": "2026-06-21T20:38:43Z",
        "updated_at": "2026-06-25T08:32:38Z",
        "closed_at": "2026-06-25T08:32:38Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/378",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 379,
        "state": "closed",
        "title": "gpu(perf): Q8 benchmarks WSL-launch-bound at 9.2 tok/s vs 20-30 target",
        "labels": [
          "gpu",
          "performance"
        ],
        "created_at": "2026-06-21T20:38:47Z",
        "updated_at": "2026-06-24T18:02:03Z",
        "closed_at": "2026-06-24T18:02:03Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/379",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 389,
        "state": "closed",
        "title": "docs(fak): Add FAQ and common issues page",
        "labels": [
          "documentation",
          "enhancement"
        ],
        "created_at": "2026-06-21T20:39:30Z",
        "updated_at": "2026-06-22T08:47:17Z",
        "closed_at": "2026-06-22T08:47:17Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/389",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 400,
        "state": "closed",
        "title": "feat(compute): achieve SIMD throughput parity with llama.cpp",
        "labels": [
          "enhancement",
          "performance"
        ],
        "created_at": "2026-06-21T20:40:13Z",
        "updated_at": "2026-06-25T07:47:20Z",
        "closed_at": "2026-06-25T07:47:20Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/400",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 415,
        "state": "closed",
        "title": "Hardware catalog: specs, onboarding, baseline runs",
        "labels": [
          "documentation",
          "substrate"
        ],
        "created_at": "2026-06-21T20:41:11Z",
        "updated_at": "2026-06-27T04:41:45Z",
        "closed_at": "2026-06-27T04:41:45Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/415",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 430,
        "state": "closed",
        "title": "bench(fanbench): add tuned shared-prefix serving baselines for vLLM/SGLang/llama.cpp",
        "labels": [
          "help wanted",
          "agentic-serving",
          "performance",
          "priority/P1",
          "research"
        ],
        "created_at": "2026-06-21T20:47:29Z",
        "updated_at": "2026-06-27T15:30:54Z",
        "closed_at": "2026-06-27T15:30:54Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/430",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 431,
        "state": "closed",
        "title": "bench(fanbench): validate 256K-1M prefix-scale points with real long-context model wall-clock",
        "labels": [
          "model",
          "performance",
          "priority/P1",
          "research"
        ],
        "created_at": "2026-06-21T20:47:33Z",
        "updated_at": "2026-06-25T07:47:18Z",
        "closed_at": "2026-06-25T07:47:18Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/431",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving"
        ]
      },
      {
        "number": 438,
        "state": "closed",
        "title": "perf(qwen35): profile and accelerate Qwen3.6 first-token prefill/decode",
        "labels": [
          "help wanted",
          "compute",
          "model",
          "model-arch",
          "performance",
          "priority/P1"
        ],
        "created_at": "2026-06-21T20:48:02Z",
        "updated_at": "2026-06-26T07:15:41Z",
        "closed_at": "2026-06-26T07:15:41Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/438",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding"
        ]
      },
      {
        "number": 439,
        "state": "closed",
        "title": "perf(qwen35): keep Qwen3.6 GGUF weights native q4/q6 instead of Q8 expansion",
        "labels": [
          "compute",
          "loader",
          "model",
          "model-arch",
          "performance",
          "priority/P2"
        ],
        "created_at": "2026-06-21T20:48:06Z",
        "updated_at": "2026-06-25T12:47:41Z",
        "closed_at": "2026-06-25T12:47:41Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/439",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion"
        ]
      },
      {
        "number": 442,
        "state": "closed",
        "title": "test(qwen35): add real-artifact oracle and llama.cpp parity witness",
        "labels": [
          "enhancement",
          "model",
          "model-arch",
          "priority/P1"
        ],
        "created_at": "2026-06-21T20:48:18Z",
        "updated_at": "2026-06-24T02:16:22Z",
        "closed_at": "2026-06-24T02:16:22Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/442",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing"
        ]
      },
      {
        "number": 443,
        "state": "closed",
        "title": "perf(qwen35): add chunked/Q8/device Gated-DeltaNet path",
        "labels": [
          "enhancement",
          "compute",
          "gpu",
          "model",
          "model-arch",
          "performance",
          "priority/P2"
        ],
        "created_at": "2026-06-21T20:48:22Z",
        "updated_at": "2026-06-24T02:01:27Z",
        "closed_at": "2026-06-24T02:01:27Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/443",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 445,
        "state": "closed",
        "title": "tokenizer: roll Go BPE + byte-exact validation vs reference tokenizers; wire chat detok",
        "labels": [
          "enhancement",
          "model",
          "model-arch"
        ],
        "created_at": "2026-06-21T20:48:30Z",
        "updated_at": "2026-06-22T07:30:04Z",
        "closed_at": "2026-06-22T07:30:04Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/445",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "structured_generation_tools",
          "loader_conversion"
        ]
      },
      {
        "number": 447,
        "state": "closed",
        "title": "[epic] qwen35 (Qwen3.6-27B) hybrid Gated-DeltaNet arch in fak's own engine",
        "labels": [
          "enhancement",
          "model",
          "model-arch"
        ],
        "created_at": "2026-06-21T20:48:38Z",
        "updated_at": "2026-06-23T22:31:25Z",
        "closed_at": "2026-06-23T22:31:25Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/447",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion"
        ]
      },
      {
        "number": 462,
        "state": "closed",
        "title": "Track: AMD RX 7600 Vulkan decode parity with llama.cpp (~17 tok/s f32, ~8× off)",
        "labels": [],
        "created_at": "2026-06-21T20:49:39Z",
        "updated_at": "2026-06-24T00:31:31Z",
        "closed_at": "2026-06-24T00:31:31Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/462",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 469,
        "state": "closed",
        "title": "research: grammar-constrained tool-call decoding so weak models emit valid structured calls (upstream fix vs downstream parse)",
        "labels": [
          "enhancement",
          "research"
        ],
        "created_at": "2026-06-21T20:50:07Z",
        "updated_at": "2026-06-22T20:16:08Z",
        "closed_at": "2026-06-22T20:16:08Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/469",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 472,
        "state": "closed",
        "title": "Land model-level Q8 routing through Vulkan HAL",
        "labels": [
          "enhancement",
          "gpu",
          "priority/P2"
        ],
        "created_at": "2026-06-21T20:50:18Z",
        "updated_at": "2026-06-28T06:30:38Z",
        "closed_at": "2026-06-28T06:30:37Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/472",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 475,
        "state": "closed",
        "title": "loader: mmap/streaming single-file safetensors (LoadSafetensors / LoadSafetensorsQuant os.ReadFile the whole file)",
        "labels": [
          "enhancement",
          "loader",
          "model"
        ],
        "created_at": "2026-06-21T20:50:29Z",
        "updated_at": "2026-06-22T07:30:08Z",
        "closed_at": "2026-06-22T07:30:08Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/475",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing"
        ]
      },
      {
        "number": 476,
        "state": "closed",
        "title": "perf(model): NEON activation quantization (quantizeVecQ8 is scalar on arm64)",
        "labels": [
          "compute",
          "model",
          "performance"
        ],
        "created_at": "2026-06-21T20:50:33Z",
        "updated_at": "2026-06-22T20:11:06Z",
        "closed_at": "2026-06-22T20:11:06Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/476",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 477,
        "state": "closed",
        "title": "perf(model): arm64 decode kernel — amortized-FP-reduction / i8mm SMMLA path (own argmax+cosine gate)",
        "labels": [
          "compute",
          "model",
          "performance",
          "research"
        ],
        "created_at": "2026-06-21T20:50:37Z",
        "updated_at": "2026-06-22T20:11:03Z",
        "closed_at": "2026-06-22T20:11:03Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/477",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 478,
        "state": "closed",
        "title": "perf(model): arm64 NEON register-blocked Q8_0 GEMM tile (prefill parity lever)",
        "labels": [
          "compute",
          "model",
          "model-arch",
          "performance"
        ],
        "created_at": "2026-06-21T20:50:41Z",
        "updated_at": "2026-06-22T07:54:18Z",
        "closed_at": "2026-06-22T07:54:18Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/478",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 480,
        "state": "closed",
        "title": "perf(gpu): throughput-parity tracking — Go-CUDA tok/s vs llama.cpp on Qwen2.5-7B-Q4_K_M",
        "labels": [
          "gpu",
          "performance",
          "research"
        ],
        "created_at": "2026-06-21T20:50:49Z",
        "updated_at": "2026-06-22T07:54:16Z",
        "closed_at": "2026-06-22T07:54:16Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/480",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "build_platform_ops"
        ]
      },
      {
        "number": 484,
        "state": "closed",
        "title": "gpu(compute): fp16 compute path (cuBLAS HGEMM / tensor cores) with a per-backend cosine threshold",
        "labels": [
          "compute",
          "gpu",
          "performance"
        ],
        "created_at": "2026-06-21T20:51:05Z",
        "updated_at": "2026-06-22T17:30:59Z",
        "closed_at": "2026-06-22T17:30:59Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/484",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 487,
        "state": "closed",
        "title": "[epic] Model-arch seam: adapt the in-kernel fusion past Llama-only to the top-10 model families",
        "labels": [
          "enhancement",
          "model-arch"
        ],
        "created_at": "2026-06-21T20:51:17Z",
        "updated_at": "2026-06-22T07:54:07Z",
        "closed_at": "2026-06-22T07:54:07Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/487",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 489,
        "state": "closed",
        "title": "Ecosystem loaders: GGUF (new internal/ggufload leaf) + int4/MXFP4 dequant-on-load",
        "labels": [
          "enhancement",
          "model-arch"
        ],
        "created_at": "2026-06-21T20:51:24Z",
        "updated_at": "2026-06-22T07:42:30Z",
        "closed_at": "2026-06-22T07:42:30Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/489",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 568,
        "state": "closed",
        "title": "epic(wirescreen): local model on the wire — the witnessed-lossy-proposer spine",
        "labels": [],
        "created_at": "2026-06-23T23:18:15Z",
        "updated_at": "2026-06-24T05:02:13Z",
        "closed_at": "2026-06-24T05:02:13Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/568",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 569,
        "state": "closed",
        "title": "feat(wirescreen): model-backed semantic poison screener (FAK_WIRE_SCREEN=model)",
        "labels": [],
        "created_at": "2026-06-23T23:18:16Z",
        "updated_at": "2026-06-24T03:01:23Z",
        "closed_at": "2026-06-24T03:01:23Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/569",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 570,
        "state": "closed",
        "title": "feat(ctxmmu): useful page-out — model-authored digest via the reserved ScreenDigest disposition",
        "labels": [],
        "created_at": "2026-06-23T23:18:17Z",
        "updated_at": "2026-06-24T00:31:29Z",
        "closed_at": "2026-06-24T00:31:29Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/570",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 571,
        "state": "closed",
        "title": "feat(ctxmmu): multi-modal screenshot triage — perceptual-hash frame dedup as a reversible Transform",
        "labels": [],
        "created_at": "2026-06-23T23:18:18Z",
        "updated_at": "2026-06-24T03:16:20Z",
        "closed_at": "2026-06-24T03:16:20Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/571",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 572,
        "state": "closed",
        "title": "feat(wirescreen): pre-send PII/secret redaction (the privacy compressor)",
        "labels": [],
        "created_at": "2026-06-23T23:18:19Z",
        "updated_at": "2026-06-24T03:01:21Z",
        "closed_at": "2026-06-24T03:01:21Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/572",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 581,
        "state": "closed",
        "title": "feat(serve): wire RadixAttention prefix-tree KV reuse into the fak serve mainline",
        "labels": [
          "enhancement",
          "agentic-serving",
          "performance",
          "prompt-caching"
        ],
        "created_at": "2026-06-24T14:04:23Z",
        "updated_at": "2026-06-24T14:24:58Z",
        "closed_at": "2026-06-24T14:24:58Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/581",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 633,
        "state": "open",
        "title": "epic(edge): make fak compelling to mobile / edge / IoT platform owners",
        "labels": [
          "enhancement",
          "epic",
          "priority/P2",
          "security"
        ],
        "created_at": "2026-06-25T01:14:09Z",
        "updated_at": "2026-06-29T22:34:08Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/633",
        "categories": [
          "native_backends",
          "models_architectures",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 905,
        "state": "closed",
        "title": "research(sota): learn production kernel stacks before more native GLM/Qwen optimization",
        "labels": [
          "compute",
          "cuda",
          "gpu",
          "model",
          "performance",
          "priority/P2",
          "quantization",
          "research"
        ],
        "created_at": "2026-06-26T15:49:47Z",
        "updated_at": "2026-06-26T17:56:09Z",
        "closed_at": "2026-06-26T17:56:09Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/905",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 931,
        "state": "open",
        "title": "epic(serving): run exact Qwen3.6-27B-FP8 on GPU server in the pure fak kernel with native L2/L3",
        "labels": [
          "agentic-serving",
          "cuda",
          "epic",
          "gpu",
          "multi-gpu",
          "performance",
          "priority/P0",
          "track/B-performance",
          "current",
          "gen/now"
        ],
        "created_at": "2026-06-27T03:42:31Z",
        "updated_at": "2026-08-15T16:09:23Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/931",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 934,
        "state": "closed",
        "title": "model(inkernel): forward panics on the real Qwen3.6-27B GGUF — it's a GDN/SSM hybrid (fused attn_qkv + ssm_*), no self_attn.q_proj",
        "labels": [
          "cuda",
          "gpu",
          "model",
          "model-arch",
          "model-support",
          "priority/P1",
          "track/A-model-support"
        ],
        "created_at": "2026-06-27T04:19:26Z",
        "updated_at": "2026-06-29T00:15:43Z",
        "closed_at": "2026-06-29T00:15:43Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/934",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "loader_conversion",
          "build_platform_ops"
        ]
      },
      {
        "number": 949,
        "state": "closed",
        "title": "perf(cuda): hold raw Q4_K resident for standard-arch on --backend cuda (FAK_Q4K ignored on device → loads Q8, ~2x VRAM)",
        "labels": [
          "cuda",
          "gpu",
          "model-support",
          "performance",
          "priority/P2",
          "track/B-performance"
        ],
        "created_at": "2026-06-27T04:53:49Z",
        "updated_at": "2026-06-27T15:04:49Z",
        "closed_at": "2026-06-27T10:11:12Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/949",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 968,
        "state": "open",
        "title": "A100 bench-node coverage frontier is blocked on local weights (gpucheck/modelbench need a model)",
        "labels": [
          "enhancement",
          "benchmark",
          "class:dev"
        ],
        "created_at": "2026-06-27T06:12:37Z",
        "updated_at": "2026-07-10T00:02:04Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/968",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "multimodal_embedding"
        ]
      },
      {
        "number": 974,
        "state": "closed",
        "title": "fak serve (CPU reference path): add a pre-load memory-fit guard so an all-resident GLM-5.2 load can't OOM-wedge the host",
        "labels": [
          "benchmark",
          "model"
        ],
        "created_at": "2026-06-27T07:49:31Z",
        "updated_at": "2026-06-27T14:51:11Z",
        "closed_at": "2026-06-27T14:51:11Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/974",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion"
        ]
      },
      {
        "number": 976,
        "state": "closed",
        "title": "Register a CPU-server bench node + collect the GLM-5.2 fak-vs-llama.cpp CPU apples-to-apples throughput row (B3 rung)",
        "labels": [
          "benchmark"
        ],
        "created_at": "2026-06-27T07:59:47Z",
        "updated_at": "2026-06-27T12:38:31Z",
        "closed_at": "2026-06-27T09:55:31Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/976",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving"
        ]
      },
      {
        "number": 977,
        "state": "open",
        "title": "epic(perf): the GPU decode-parity playbook — apply the #67 isolation-split method to every off-parity backend",
        "labels": [
          "epic",
          "gpu",
          "performance",
          "priority/P1",
          "track/B-performance"
        ],
        "created_at": "2026-06-27T10:04:38Z",
        "updated_at": "2026-06-29T22:34:26Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/977",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 996,
        "state": "closed",
        "title": "GLM-5.2 fak-kernel serve: 0.03 tok/s (cpu-offload wall) + verify coherence on current code",
        "labels": [],
        "created_at": "2026-06-27T14:14:55Z",
        "updated_at": "2026-06-28T02:22:54Z",
        "closed_at": "2026-06-28T02:22:54Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/996",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 1012,
        "state": "closed",
        "title": "test(swebench): replay a known-solved ticket through the GLM-5.2 fak-kernel gateway on dgx; capture cache-value + grade (epic #1010 CHILD-B)",
        "labels": [
          "enhancement",
          "class:dev"
        ],
        "created_at": "2026-06-27T16:01:46Z",
        "updated_at": "2026-07-15T00:40:14Z",
        "closed_at": "2026-07-15T00:40:14Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/1012",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 1030,
        "state": "closed",
        "title": "tokenizer+template: Ornith Qwen3.5 tokenizer (vocab 248320, 2-id eos) + model-driven chat_template.jinja",
        "labels": [
          "model-support",
          "qwen"
        ],
        "created_at": "2026-06-27T16:50:02Z",
        "updated_at": "2026-06-27T17:41:23Z",
        "closed_at": "2026-06-27T17:41:23Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/1030",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 1041,
        "state": "closed",
        "title": "edge(engine): ship one reference EngineDriver wrapping a phone-class on-device runtime (Ollama/llama.cpp -> MLC)",
        "labels": [
          "deployment",
          "integration"
        ],
        "created_at": "2026-06-27T16:59:38Z",
        "updated_at": "2026-06-28T08:30:38Z",
        "closed_at": "2026-06-28T08:30:38Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/1041",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 1045,
        "state": "open",
        "title": "epic(serve): automatic KV-cache memory sizing — auto-fit the context to the host, don't refuse or push a manual knob",
        "labels": [
          "compute",
          "epic",
          "performance",
          "priority/P1"
        ],
        "created_at": "2026-06-27T23:00:56Z",
        "updated_at": "2026-06-29T22:34:16Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/1045",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "build_platform_ops"
        ]
      },
      {
        "number": 1047,
        "state": "closed",
        "title": "feat(compute): KV-cache precision tiers (q8/f32) so denser KV multiplies the fittable context",
        "labels": [
          "compute",
          "performance",
          "priority/P2"
        ],
        "created_at": "2026-06-27T23:01:42Z",
        "updated_at": "2026-06-28T06:15:40Z",
        "closed_at": "2026-06-28T06:15:40Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/1047",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "build_platform_ops"
        ]
      },
      {
        "number": 1048,
        "state": "closed",
        "title": "feat(compute): tiered KV residency — spill cold KV host<->device instead of shrinking context",
        "labels": [
          "compute",
          "gpu",
          "performance",
          "priority/P2"
        ],
        "created_at": "2026-06-27T23:01:44Z",
        "updated_at": "2026-06-27T23:30:46Z",
        "closed_at": "2026-06-27T23:30:46Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/1048",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "training_adapters"
        ]
      },
      {
        "number": 1051,
        "state": "closed",
        "title": "scorecard(latency): measure the gateway request-path latency percentiles (TTFT/TPOT/E2EL/tail) — the tax fak can honestly report",
        "labels": [
          "agentic-serving",
          "benchmark",
          "performance"
        ],
        "created_at": "2026-06-27T23:09:58Z",
        "updated_at": "2026-06-28T06:15:38Z",
        "closed_at": "2026-06-28T06:15:38Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/1051",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 1055,
        "state": "closed",
        "title": "scorecard(model): follow through on MLPerf v6.0 / DeepSeek-R1 671B — make fak's MoE position one coherent story across the rows",
        "labels": [
          "benchmark",
          "model-support",
          "moe"
        ],
        "created_at": "2026-06-27T23:10:23Z",
        "updated_at": "2026-06-28T22:53:31Z",
        "closed_at": "2026-06-28T22:53:31Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/1055",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 1056,
        "state": "closed",
        "title": "epic(local-coding): one command for a small local model + your coding harness + kernel governance",
        "labels": [
          "enhancement",
          "epic",
          "priority/P1"
        ],
        "created_at": "2026-06-27T23:16:53Z",
        "updated_at": "2026-06-28T04:19:31Z",
        "closed_at": "2026-06-28T04:19:31Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/1056",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 1057,
        "state": "closed",
        "title": "feat(guard): fak guard --local — auto-detect a running local model server (Ollama/LM Studio/llama.cpp)",
        "labels": [
          "enhancement",
          "priority/P1"
        ],
        "created_at": "2026-06-27T23:19:05Z",
        "updated_at": "2026-06-28T03:54:07Z",
        "closed_at": "2026-06-28T03:54:07Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/1057",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 1062,
        "state": "closed",
        "title": "perf(cpu): witnessed GLM-5.2 q4 CPU serving — 0.06 tok/s decode + NFS-vs-NVMe load-path tax (CPU server)",
        "labels": [
          "benchmark",
          "gguf",
          "moe",
          "performance",
          "simd",
          "track/B-performance"
        ],
        "created_at": "2026-06-27T23:20:33Z",
        "updated_at": "2026-08-15T16:09:21Z",
        "closed_at": "2026-06-28T03:15:41Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/1062",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 1063,
        "state": "open",
        "title": "epic(benchmarks): fak benchmark-entry portfolio - which competitive boards fak enters (ranked, with the per-lane honesty fence) + the SKIP list",
        "labels": [
          "enhancement",
          "agentic-serving",
          "benchmark",
          "epic",
          "priority/P1",
          "prompt-caching",
          "track/E-testing-quality",
          "trust-floor"
        ],
        "created_at": "2026-06-27T23:56:06Z",
        "updated_at": "2026-08-20T20:07:59Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/1063",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 1067,
        "state": "closed",
        "title": "fix(model): --metal q4k frees qt.raw but CPU prefill GEMM still reads it -> slice-bounds panic on multi-K prompts",
        "labels": [
          "compute",
          "gpu",
          "metal",
          "priority/P1"
        ],
        "created_at": "2026-06-27T23:58:35Z",
        "updated_at": "2026-06-28T04:11:10Z",
        "closed_at": "2026-06-28T04:11:10Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/1067",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing"
        ]
      },
      {
        "number": 1071,
        "state": "closed",
        "title": "perf(model): in-kernel q4k prefill runs the slow Q8/CPU GEMM, not the Metal Q4_K batched GEMM — ~0.4 tok/s, times out on real prompts",
        "labels": [
          "compute",
          "gpu",
          "metal",
          "performance",
          "priority/P1"
        ],
        "created_at": "2026-06-28T00:10:37Z",
        "updated_at": "2026-06-28T03:14:20Z",
        "closed_at": "2026-06-28T03:00:38Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/1071",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "training_adapters"
        ]
      },
      {
        "number": 1085,
        "state": "closed",
        "title": "perf(metalgemm): q4k_gemm prefill GEMM is occupancy-bound at small P — SIMD-group dot reduction",
        "labels": [
          "performance"
        ],
        "created_at": "2026-06-28T04:12:41Z",
        "updated_at": "2026-06-28T04:21:11Z",
        "closed_at": "2026-06-28T04:21:11Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/1085",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 1113,
        "state": "closed",
        "title": "q4k Metal prefill: per-call GPU round-trip + lazy weight upload cap warm prefill at ~7x under llama.cpp",
        "labels": [
          "gpu",
          "metal",
          "performance",
          "priority/P1"
        ],
        "created_at": "2026-06-28T15:16:40Z",
        "updated_at": "2026-06-28T15:36:13Z",
        "closed_at": "2026-06-28T15:36:13Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/1113",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 1124,
        "state": "closed",
        "title": "epic(perf): the native-code doctrine — where raw C / Go-assembly earns its place, and where it must never go",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-06-28T16:17:55Z",
        "updated_at": "2026-08-15T16:09:18Z",
        "closed_at": "2026-06-29T17:36:14Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/1124",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 1127,
        "state": "closed",
        "title": "perf(simd): AVX2 Q8 prefill GEMM falls to scalar — add the register-blocked AVX2 tile (the CPU server prefill floor) (#1124 C3)",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-06-28T16:25:59Z",
        "updated_at": "2026-08-15T16:09:16Z",
        "closed_at": "2026-06-29T19:21:12Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/1127",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 1138,
        "state": "closed",
        "title": "nightrun: add a first-class SATURATED selector verdict + enqueue the next datum as a pickable task",
        "labels": [
          "enhancement"
        ],
        "created_at": "2026-06-29T00:11:56Z",
        "updated_at": "2026-06-29T04:15:50Z",
        "closed_at": "2026-06-29T04:15:50Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/1138",
        "categories": [
          "native_backends",
          "models_architectures",
          "build_platform_ops"
        ]
      },
      {
        "number": 1146,
        "state": "closed",
        "title": "Epic: durable built-in visibility for fleet session crash + resume (one source, prove-the-resume-took)",
        "labels": [
          "enhancement",
          "epic"
        ],
        "created_at": "2026-06-29T02:27:14Z",
        "updated_at": "2026-07-03T23:22:36Z",
        "closed_at": "2026-07-03T23:22:36Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/1146",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 1147,
        "state": "closed",
        "title": "epic(self-tax): the self-tax plane — first-class evidence that fak's own methods don't degrade performance",
        "labels": [
          "epic",
          "observability",
          "performance",
          "priority/P1",
          "track/G-foundation"
        ],
        "created_at": "2026-06-29T02:29:22Z",
        "updated_at": "2026-06-29T03:11:13Z",
        "closed_at": "2026-06-29T03:11:13Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/1147",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 1242,
        "state": "closed",
        "title": "test(metal): execute the Qwen3.6-27B Mac M3 Pro parity gate + multi-agent verification run",
        "labels": [],
        "created_at": "2026-06-29T04:56:10Z",
        "updated_at": "2026-06-29T07:30:40Z",
        "closed_at": "2026-06-29T07:30:40Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/1242",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 1382,
        "state": "closed",
        "title": "perf(metal): resident one-command-buffer decode forward for Qwen3.6-27B hybrid-MoE (lift metal_decode decline)",
        "labels": [
          "gpu",
          "metal",
          "performance",
          "priority/P1",
          "track/C-gpu-backend"
        ],
        "created_at": "2026-06-29T22:09:20Z",
        "updated_at": "2026-07-01T05:41:13Z",
        "closed_at": "2026-07-01T05:41:13Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/1382",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 1457,
        "state": "open",
        "title": "epic(roadmap): the 6 live-but-not-quite-there workstreams — milestone-mapped + witnessed status",
        "labels": [
          "epic"
        ],
        "created_at": "2026-06-30T04:26:27Z",
        "updated_at": "2026-06-30T10:48:42Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/1457",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 1476,
        "state": "open",
        "title": "epic(serving): the GCP-H100 GLM-5.2 kernel demo — one command, witnessed end to end",
        "labels": [
          "agentic-serving",
          "cuda",
          "epic",
          "model-support",
          "priority/P1"
        ],
        "created_at": "2026-06-30T10:22:20Z",
        "updated_at": "2026-06-30T22:58:35Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/1476",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 1478,
        "state": "open",
        "title": "bench(cuda): collect the three H100 kernel-lever witnesses — Q8 decode parity, TF32 prefill, prefill phase-profile (epic #1476 C2)",
        "labels": [
          "enhancement",
          "cuda",
          "class:dev"
        ],
        "created_at": "2026-06-30T10:27:05Z",
        "updated_at": "2026-07-09T23:59:58Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/1478",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "api_ui_docs"
        ]
      },
      {
        "number": 1479,
        "state": "open",
        "title": "perf(cuda): device-resident pos/nPos so the decode CUDA graph is capture-once-replay-many (Lever 2, epic #1476 C3)",
        "labels": [
          "enhancement",
          "cuda",
          "class:dev"
        ],
        "created_at": "2026-06-30T10:27:08Z",
        "updated_at": "2026-07-10T00:01:04Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/1479",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "build_platform_ops"
        ]
      },
      {
        "number": 1480,
        "state": "closed",
        "title": "bench(serving): measure TTFT/TPOT/ITL for fak gateway vs raw SGLang on a GPU server",
        "labels": [
          "enhancement",
          "agentic-serving",
          "benchmark",
          "model-support"
        ],
        "created_at": "2026-06-30T10:27:18Z",
        "updated_at": "2026-07-04T14:32:44Z",
        "closed_at": "2026-07-04T14:32:44Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/1480",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 1490,
        "state": "open",
        "title": "epic(cache): turn the vCache gates ON by default + honest per-mechanism attribution",
        "labels": [
          "enhancement",
          "agentic-serving",
          "epic",
          "prompt-caching"
        ],
        "created_at": "2026-06-30T10:40:03Z",
        "updated_at": "2026-08-19T08:56:16Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/1490",
        "categories": [
          "native_backends",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 1498,
        "state": "open",
        "title": "feat(cache): push the vBlock/anchor abstraction across pure-fak, sglang, vllm, llama.cpp + the API surface",
        "labels": [
          "enhancement",
          "agentic-serving",
          "prompt-caching",
          "class:dev"
        ],
        "created_at": "2026-06-30T10:41:42Z",
        "updated_at": "2026-08-21T15:27:34Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/1498",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 1549,
        "state": "closed",
        "title": "cache-default[31]: external engines: Build a cache capability inventory for SGLang, vLLM, llama.cpp, Ollama, and LM Studio",
        "labels": [
          "enhancement",
          "agentic-serving",
          "generation",
          "gen/second-next"
        ],
        "created_at": "2026-06-30T16:01:43Z",
        "updated_at": "2026-07-07T20:50:11Z",
        "closed_at": "2026-07-07T20:49:52Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/1549",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 1550,
        "state": "closed",
        "title": "cache-default[32]: external engines: Add a wire-neutral `engine.CacheCapability` contract",
        "labels": [
          "enhancement",
          "generation",
          "gen/second-next"
        ],
        "created_at": "2026-06-30T16:01:45Z",
        "updated_at": "2026-07-07T20:59:52Z",
        "closed_at": "2026-07-07T20:59:52Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/1550",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 1551,
        "state": "closed",
        "title": "cache-default[33]: external engines: Add a vLLM prefix-cache observation adapter",
        "labels": [
          "enhancement",
          "agentic-serving",
          "generation",
          "gen/second-next"
        ],
        "created_at": "2026-06-30T16:01:46Z",
        "updated_at": "2026-07-08T00:00:06Z",
        "closed_at": "2026-07-08T00:00:06Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/1551",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 1552,
        "state": "closed",
        "title": "cache-default[34]: external engines: Add an SGLang radix/prefix-cache observation adapter",
        "labels": [
          "enhancement",
          "agentic-serving"
        ],
        "created_at": "2026-06-30T16:01:48Z",
        "updated_at": "2026-07-08T02:23:41Z",
        "closed_at": "2026-07-08T02:23:41Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/1552",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 1553,
        "state": "closed",
        "title": "cache-default[35]: external engines: Add a llama.cpp/llama-server session-cache observation adapter",
        "labels": [
          "enhancement",
          "generation",
          "gen/second-next"
        ],
        "created_at": "2026-06-30T16:01:50Z",
        "updated_at": "2026-07-07T22:43:12Z",
        "closed_at": "2026-07-07T22:43:12Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/1553",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 1554,
        "state": "closed",
        "title": "cache-default[36]: external engines: Add adapter conformance tests for \"fronted\" vs \"cache-integrated.\"",
        "labels": [
          "enhancement",
          "class:dev"
        ],
        "created_at": "2026-06-30T16:01:51Z",
        "updated_at": "2026-07-16T11:08:12Z",
        "closed_at": "2026-07-16T11:08:12Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/1554",
        "categories": [
          "native_backends",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 1556,
        "state": "closed",
        "title": "cache-default[38]: external engines: Add per-engine cold-path correctness witnesses",
        "labels": [
          "enhancement",
          "agentic-serving",
          "class:dev"
        ],
        "created_at": "2026-06-30T16:01:54Z",
        "updated_at": "2026-07-15T02:46:15Z",
        "closed_at": "2026-07-15T02:46:15Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/1556",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 1678,
        "state": "open",
        "title": "Epic: fak as the vendor-neutral binding layer for neo-silicon and neo-clouds",
        "labels": [
          "enhancement",
          "epic",
          "research",
          "substrate",
          "gen/next"
        ],
        "created_at": "2026-06-30T16:47:11Z",
        "updated_at": "2026-06-30T17:17:10Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/1678",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 1688,
        "state": "closed",
        "title": "GPU-bias correction: add a hardware-shape-neutrality dimension to the industry scorecard",
        "labels": [
          "enhancement",
          "research",
          "substrate",
          "gen/next"
        ],
        "created_at": "2026-06-30T16:49:52Z",
        "updated_at": "2026-06-30T17:33:23Z",
        "closed_at": "2026-06-30T17:33:23Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/1688",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 1704,
        "state": "closed",
        "title": "fak serve: KVCache.Evict panics on Gated-DeltaNet hybrid (Qwen3.6-27B) — decodes then RemoteDisconnects, blocks local agentic serving",
        "labels": [],
        "created_at": "2026-06-30T18:02:41Z",
        "updated_at": "2026-06-30T18:21:35Z",
        "closed_at": "2026-06-30T18:21:35Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/1704",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 2120,
        "state": "closed",
        "title": "feat(guard): `fak guard --local` — print a positive backend-detection banner (which server + model was chosen)",
        "labels": [
          "priority/P2",
          "ux",
          "dev-ex"
        ],
        "created_at": "2026-07-01T16:11:05Z",
        "updated_at": "2026-07-01T19:52:46Z",
        "closed_at": "2026-07-01T19:52:46Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/2120",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 2236,
        "state": "open",
        "title": "epic(superset): fak > best of vLLM + SGLang + Dynamo + TRT-LLM/LMDeploy/llama.cpp - ranked concept-by-concept, conceptual to benchmark, memory-first",
        "labels": [
          "epic",
          "pagedattention",
          "performance",
          "prompt-caching",
          "research"
        ],
        "created_at": "2026-07-02T03:35:12Z",
        "updated_at": "2026-07-10T15:05:20Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/2236",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 2237,
        "state": "open",
        "title": "research(superset): memory-concept ranking dossier - 10 concepts ranked fak-vs-engines with evidence + adopt-or-SKIP verdicts (epic #2236)",
        "labels": [
          "documentation",
          "pagedattention",
          "prompt-caching",
          "research",
          "class:dev"
        ],
        "created_at": "2026-07-02T03:35:43Z",
        "updated_at": "2026-07-09T23:14:18Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/2237",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 2283,
        "state": "closed",
        "title": "docs(explainer): \"fak is not a serving engine\" — the honest boundary",
        "labels": [
          "popularization",
          "adoption"
        ],
        "created_at": "2026-07-02T10:48:18Z",
        "updated_at": "2026-07-03T19:51:01Z",
        "closed_at": "2026-07-03T19:51:01Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/2283",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 2722,
        "state": "open",
        "title": "epic(mac): Mac CPU+RAM+SSD offload serving — close the #2691 slow-prefill gap, beat llama.cpp/MLX provably, land fak's cache-value on top",
        "labels": [
          "agentic-serving",
          "epic",
          "metal",
          "performance",
          "priority/P1",
          "track/C-gpu-backend",
          "generation",
          "gen/next"
        ],
        "created_at": "2026-07-04T22:27:58Z",
        "updated_at": "2026-07-18T05:57:47Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/2722",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "loader_conversion",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 2723,
        "state": "open",
        "title": "bench(macbench): head-to-head fak vs llama.cpp vs MLX on real Mac hardware — TTFT/ITL/throughput, prefill and decode split",
        "labels": [
          "benchmark",
          "metal",
          "performance",
          "generation",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-07-04T22:28:12Z",
        "updated_at": "2026-07-09T23:10:07Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/2723",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 2724,
        "state": "closed",
        "title": "feat(engine): MLX ride-adapter — front mlx-lm/vllm-mlx the way fak already fronts vLLM/SGLang",
        "labels": [
          "integration",
          "performance",
          "generation",
          "gen/next",
          "class:dev"
        ],
        "created_at": "2026-07-04T22:28:15Z",
        "updated_at": "2026-07-15T10:40:34Z",
        "closed_at": "2026-07-15T10:40:34Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/2724",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 2725,
        "state": "closed",
        "title": "perf(metal): root-cause and close the claude-mac-fak 10-15 min prefill gap named in #2691",
        "labels": [
          "metal",
          "performance",
          "priority/P1",
          "generation",
          "gen/next"
        ],
        "created_at": "2026-07-04T22:28:19Z",
        "updated_at": "2026-07-07T18:39:50Z",
        "closed_at": "2026-07-07T18:39:50Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/2725",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 3030,
        "state": "closed",
        "title": "kernel(gptq): native packed GPTQ GPU parity and throughput witness",
        "labels": [
          "cuda",
          "gpu",
          "model-support",
          "performance",
          "priority/P1",
          "quantization",
          "testing",
          "track/A-model-support"
        ],
        "created_at": "2026-07-06T05:00:45Z",
        "updated_at": "2026-07-06T13:00:38Z",
        "closed_at": "2026-07-06T13:00:38Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/3030",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 3060,
        "state": "open",
        "title": "run(livecodebench): GLM-5.2 on DGX — first graded pass@1 (raw + fak arms)",
        "labels": [
          "benchmark",
          "epic",
          "livecodebench"
        ],
        "created_at": "2026-07-06T20:52:12Z",
        "updated_at": "2026-08-18T02:35:29Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/3060",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 3073,
        "state": "closed",
        "title": "epic(perf): GLM-5.2 on DGX 2/3 — drive single-stream 23.2→~150 tok/s (80% of roofline) + stand up batched aggregate, in a day",
        "labels": [
          "agentic-serving",
          "epic",
          "gpu",
          "multi-gpu",
          "performance",
          "track/B-performance",
          "fleet-400iph"
        ],
        "created_at": "2026-07-06T21:16:54Z",
        "updated_at": "2026-08-17T05:05:42Z",
        "closed_at": "2026-08-17T05:05:42Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/3073",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "speculative_generation"
        ]
      },
      {
        "number": 3074,
        "state": "closed",
        "title": "measure(model): pin GLM-5.2 active-params + active-bytes/token from the GGUF header + a decode-only per-op trace [DGX 2 · Lane F]",
        "labels": [
          "benchmark",
          "moe",
          "performance",
          "track/B-performance",
          "class:dev"
        ],
        "created_at": "2026-07-06T21:17:37Z",
        "updated_at": "2026-07-16T13:43:29Z",
        "closed_at": "2026-07-16T13:43:29Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/3074",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 3075,
        "state": "closed",
        "title": "perf(serve): 8-GPU tensor/row split for GLM-5.2 — kill llama.cpp `-sm layer`, put all 8 GPUs on every token [DGX 3 · Lane A · L1]",
        "labels": [
          "agentic-serving",
          "gpu",
          "multi-gpu",
          "performance",
          "track/B-performance"
        ],
        "created_at": "2026-07-06T21:17:39Z",
        "updated_at": "2026-07-10T00:29:33Z",
        "closed_at": "2026-07-07T03:00:47Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/3075",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving"
        ]
      },
      {
        "number": 3076,
        "state": "closed",
        "title": "perf(serve): enable flash-attention + CUDA-graph decode for GLM-5.2 8-GPU serve [DGX 3 · Lane A · L4]",
        "labels": [
          "cuda",
          "gpu",
          "performance",
          "track/B-performance",
          "generation",
          "gen/now"
        ],
        "created_at": "2026-07-06T21:17:41Z",
        "updated_at": "2026-07-08T12:45:57Z",
        "closed_at": "2026-07-08T12:45:57Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/3076",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 3077,
        "state": "closed",
        "title": "perf(quant): GLM-5.2 decode quant sweep — UD-Q4_K_M vs UD-Q4_K_S (CUDA resident-Q4_K fast path) vs Q4_K-pure vs Q3 [DGX 2 · Lane A · L5]",
        "labels": [
          "moe",
          "performance",
          "quantization",
          "track/B-performance",
          "generation",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-07-06T21:17:43Z",
        "updated_at": "2026-08-17T05:05:40Z",
        "closed_at": "2026-08-17T05:05:39Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/3077",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 3078,
        "state": "closed",
        "title": "perf(decode): speculative decoding for GLM-5.2 — draft-model / EAGLE / prompt-lookup, measure acceptance + tok/s [DGX 2 · Lane A-spec · L3]",
        "labels": [
          "agentic-serving",
          "gpu",
          "performance",
          "track/B-performance",
          "class:dev"
        ],
        "created_at": "2026-07-06T21:17:45Z",
        "updated_at": "2026-07-16T13:43:24Z",
        "closed_at": "2026-07-16T13:43:24Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/3078",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 3079,
        "state": "closed",
        "title": "perf(serve): stand up continuous batching for GLM-5.2 — concurrency sweep 1→128, aggregate tok/s + p50/p95 TTFT [DGX 2 · Lane B · L2]",
        "labels": [
          "agentic-serving",
          "gpu",
          "performance",
          "track/B-performance",
          "generation",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-07-06T21:18:23Z",
        "updated_at": "2026-08-17T05:05:37Z",
        "closed_at": "2026-08-17T05:05:37Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/3079",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 3080,
        "state": "closed",
        "title": "perf(serve): KV paging + context-budget tuning for many concurrent GLM-5.2 streams within 206 GiB free VRAM [DGX 2 · Lane B]",
        "labels": [
          "agentic-serving",
          "gpu",
          "performance",
          "track/B-performance",
          "generation",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-07-06T21:18:25Z",
        "updated_at": "2026-07-16T13:33:00Z",
        "closed_at": "2026-07-16T13:33:00Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/3080",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 3081,
        "state": "open",
        "title": "bench(cache): RadixAttention/prefix-cache warm-fleet A/B under concurrency — fak cache-value on shared agent prefixes [DGX 3 · Lane B-cache]",
        "labels": [
          "agentic-serving",
          "benchmark",
          "performance",
          "track/B-performance",
          "ablation",
          "class:dev"
        ],
        "created_at": "2026-07-06T21:18:27Z",
        "updated_at": "2026-07-09T23:07:39Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/3081",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 3082,
        "state": "closed",
        "title": "tooling(bench): one-command DGX GLM-5.2 serve+bench harness — any lever measurable in <10 min per node [DGX 2/3 · Lane C]",
        "labels": [
          "benchmark",
          "observability",
          "performance",
          "track/B-performance"
        ],
        "created_at": "2026-07-06T21:18:29Z",
        "updated_at": "2026-07-07T03:00:42Z",
        "closed_at": "2026-07-07T03:00:42Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/3082",
        "categories": [
          "models_architectures"
        ]
      },
      {
        "number": 3083,
        "state": "closed",
        "title": "serve: warm-start GLM-5.2 readiness on DGX (bind #3051) — first real turn is warm, not paying the ~500s cold tax [DGX 2/3 · Lane C · L8]",
        "labels": [
          "enhancement",
          "agentic-serving",
          "performance",
          "track/B-performance",
          "class:dev"
        ],
        "created_at": "2026-07-06T21:18:31Z",
        "updated_at": "2026-07-16T13:32:23Z",
        "closed_at": "2026-07-16T13:32:23Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/3083",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 3084,
        "state": "open",
        "title": "serve/infra: persist DeepGEMM/CUDA-graph/compile caches on DGX (bind #3052) — pay the GLM-5.2 cold tax once, not per boot [DGX 2/3 · Lane C · L8]",
        "labels": [
          "enhancement",
          "cuda",
          "performance",
          "track/B-performance",
          "class:dev"
        ],
        "created_at": "2026-07-06T21:18:33Z",
        "updated_at": "2026-07-09T23:07:35Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/3084",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 3085,
        "state": "closed",
        "title": "bench(prefill): first real GLM-5.2 prefill sweep on DGX — 128/512/2048/4096/8192, establish the baseline (currently unmeasured) [DGX 3 · Lane D]",
        "labels": [
          "benchmark",
          "gpu",
          "performance",
          "track/B-performance"
        ],
        "created_at": "2026-07-06T21:19:11Z",
        "updated_at": "2026-07-09T03:30:51Z",
        "closed_at": "2026-07-09T03:30:51Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/3085",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 3086,
        "state": "open",
        "title": "perf(prefill): chunked prefill + flash-attention for long-context GLM-5.2 [DGX 2 · Lane D · L9]",
        "labels": [
          "cuda",
          "gpu",
          "performance",
          "track/B-performance",
          "class:dev"
        ],
        "created_at": "2026-07-06T21:19:13Z",
        "updated_at": "2026-07-09T23:07:33Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/3086",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 3087,
        "state": "closed",
        "title": "perf(kernel): INT8 tensor-core expert GEMM for GLM-5.2 on sm_80 — the 2x aggregate compute lever [DGX 2 · Lane E · L6]",
        "labels": [
          "cuda",
          "moe",
          "performance",
          "quantization",
          "track/B-performance",
          "class:dev"
        ],
        "created_at": "2026-07-06T21:19:16Z",
        "updated_at": "2026-07-16T13:31:54Z",
        "closed_at": "2026-07-16T13:31:54Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/3087",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 3088,
        "state": "closed",
        "title": "perf(attn): serve GLM-5.2 DSA as TRUE sparse attention, not full MLA — long-context decode/prefill delta [DGX 2 · Lane E · L7]",
        "labels": [
          "model-support",
          "moe",
          "performance",
          "track/B-performance",
          "gen/second-next",
          "class:dev"
        ],
        "created_at": "2026-07-06T21:19:18Z",
        "updated_at": "2026-08-17T05:05:34Z",
        "closed_at": "2026-08-17T05:05:34Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/3088",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 3089,
        "state": "open",
        "title": "feat(multi-gpu): live native fak resident-EP device-NCCL witness on DGX (bind #1482) — the pure-fak 8-GPU tok/s [DGX 2 · Lane E · L10]",
        "labels": [
          "cuda",
          "moe",
          "multi-gpu",
          "performance",
          "track/B-performance",
          "class:dev"
        ],
        "created_at": "2026-07-06T21:19:20Z",
        "updated_at": "2026-07-09T23:07:27Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/3089",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 3090,
        "state": "closed",
        "title": "bench(roofline): live current-vs-ceiling dashboard for the GLM-5.2 DGX drive — updated from the run artifacts [DGX 2 · Lane F]",
        "labels": [
          "benchmark",
          "dashboard",
          "observability",
          "performance",
          "track/B-performance",
          "class:infra"
        ],
        "created_at": "2026-07-06T21:19:22Z",
        "updated_at": "2026-07-16T13:31:46Z",
        "closed_at": "2026-07-16T13:31:46Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/3090",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 3142,
        "state": "closed",
        "title": "research(superset): add LMCache row to the external-engine cache capability inventory (#1549 pattern) — the epic names it, the inventory can't point at it",
        "labels": [
          "prompt-caching",
          "research",
          "context-engineering",
          "class:dev"
        ],
        "created_at": "2026-07-07T21:16:36Z",
        "updated_at": "2026-08-09T01:00:25Z",
        "closed_at": "2026-08-09T01:00:25Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/3142",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 3205,
        "state": "open",
        "title": "feat(model): batch-size-invariance rung + determinism posture on fronted engines (borrow Thinking Machines batch_invariant_ops)",
        "labels": [
          "model-arch",
          "research",
          "class:dev"
        ],
        "created_at": "2026-07-08T11:03:08Z",
        "updated_at": "2026-07-09T23:17:50Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/3205",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 3237,
        "state": "closed",
        "title": "feat(fleetmetrics): cross-vendor device-telemetry spine — normalized-optional struct + {field,query} table + lock-free double-buffer + Prom/CSV/JSON + fleet-federation TUI (epic #3236)",
        "labels": [
          "enhancement",
          "dashboard",
          "observability",
          "generation",
          "gen/next",
          "class:infra"
        ],
        "created_at": "2026-07-08T15:59:37Z",
        "updated_at": "2026-07-11T21:43:11Z",
        "closed_at": "2026-07-11T21:43:11Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/3237",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "api_ui_docs"
        ]
      },
      {
        "number": 3279,
        "state": "open",
        "title": "docs(deployment): air-gapped single-binary deployment kit + SBOM — local-first but governed (epic #3256)",
        "labels": [
          "documentation",
          "deployment",
          "generation",
          "gen/next",
          "enterprise",
          "class:infra"
        ],
        "created_at": "2026-07-08T17:17:06Z",
        "updated_at": "2026-08-07T00:56:11Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/3279",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 3407,
        "state": "closed",
        "title": "docs(enginecache): re-witness external-engine cache capability inventory against the landed observation adapters (epic #1490)",
        "labels": [
          "enhancement",
          "gen/future"
        ],
        "created_at": "2026-07-09T00:38:23Z",
        "updated_at": "2026-07-09T01:00:38Z",
        "closed_at": "2026-07-09T01:00:38Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/3407",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 3418,
        "state": "open",
        "title": "research(metal): chunked WY-GEMM parallel GDN scan for prefill - close the 51.55 tok/s prefill gap (epic #59)",
        "labels": [
          "research",
          "track/B-performance",
          "class:dev"
        ],
        "created_at": "2026-07-09T00:38:42Z",
        "updated_at": "2026-08-23T09:38:14Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/3418",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving"
        ]
      },
      {
        "number": 3809,
        "state": "open",
        "title": "epic(mac): pick the model that proves fak's kernel+cache value for agentic long-horizon MANY-AGENT use on a MacBook — model-selection driven by cache economics, not single-stream speed",
        "labels": [
          "agentic-serving",
          "epic",
          "metal",
          "performance",
          "priority/P1",
          "track/C-gpu-backend",
          "generation",
          "gen/next"
        ],
        "created_at": "2026-07-10T07:27:30Z",
        "updated_at": "2026-07-10T07:28:29Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/3809",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 3813,
        "state": "open",
        "title": "bench(macbench,cachevalue): Mac many-agent shared-prefix cache-value A/B — the Mac analogue of #3081 (#3809)",
        "labels": [
          "help wanted",
          "agentic-serving",
          "metal",
          "performance",
          "track/C-gpu-backend",
          "gen/next",
          "class:dev"
        ],
        "created_at": "2026-07-10T07:30:17Z",
        "updated_at": "2026-07-10T09:56:40Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/3813",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 3814,
        "state": "closed",
        "title": "feat(model): verify/land the selected small GQA model on the in-kernel Metal serve path (#3809, builds on #67)",
        "labels": [
          "metal",
          "performance",
          "track/A-model-support",
          "track/C-gpu-backend",
          "gen/next",
          "class:dev"
        ],
        "created_at": "2026-07-10T07:30:20Z",
        "updated_at": "2026-07-16T13:31:46Z",
        "closed_at": "2026-07-16T13:31:46Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/3814",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 4254,
        "state": "open",
        "title": "epic(observability): hot-path wall-clock plane — first-launch + dispatch spawn, and the 5× on the most-called stage",
        "labels": [
          "epic",
          "observability"
        ],
        "created_at": "2026-07-11T01:47:50Z",
        "updated_at": "2026-07-11T01:47:50Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/4254",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 4273,
        "state": "closed",
        "title": "fix(inkernel): Qwen3.6-27B GGUF degenerates into repetition on ~1.3k-token prompts — short prompts coherent, non-thinking + recommended sampling already applied",
        "labels": [
          "bug",
          "gguf",
          "priority/P1",
          "qwen",
          "generation",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-07-11T05:36:13Z",
        "updated_at": "2026-08-15T16:10:44Z",
        "closed_at": "2026-07-14T23:46:05Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/4273",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 4381,
        "state": "open",
        "title": "bench(mac): collect the Qwen3.6-27B Metal serving-curve sweep (decode/longgen/prefill A-B/2-stream)",
        "labels": [
          "benchmark",
          "metal",
          "qwen",
          "class:dev"
        ],
        "created_at": "2026-07-12T01:17:20Z",
        "updated_at": "2026-07-12T04:03:05Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/4381",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 4509,
        "state": "closed",
        "title": "epic(quality): build the missing middle validation ladder from decode parity to executive-report dogfood",
        "labels": [
          "epic",
          "research",
          "testing",
          "track/E-testing-quality",
          "gen/now"
        ],
        "created_at": "2026-07-13T16:00:33Z",
        "updated_at": "2026-07-16T13:40:23Z",
        "closed_at": "2026-07-16T13:40:23Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/4509",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 4534,
        "state": "closed",
        "title": "test(quality): compare prefix-cache on/off output parity",
        "labels": [
          "testing",
          "track/E-testing-quality",
          "gen/now"
        ],
        "created_at": "2026-07-13T16:06:01Z",
        "updated_at": "2026-07-13T19:22:11Z",
        "closed_at": "2026-07-13T19:17:27Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/4534",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 4535,
        "state": "closed",
        "title": "test(quality): validate KV eviction and recomputation quality parity",
        "labels": [
          "testing",
          "track/E-testing-quality",
          "gen/now",
          "class:infra"
        ],
        "created_at": "2026-07-13T16:06:03Z",
        "updated_at": "2026-07-14T18:31:10Z",
        "closed_at": "2026-07-14T18:31:10Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/4535",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 4536,
        "state": "closed",
        "title": "test(quality): gate chunked versus monolithic prefill parity",
        "labels": [
          "testing",
          "track/E-testing-quality",
          "gen/now"
        ],
        "created_at": "2026-07-13T16:06:06Z",
        "updated_at": "2026-07-13T18:15:52Z",
        "closed_at": "2026-07-13T18:15:52Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/4536",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 4537,
        "state": "closed",
        "title": "test(quality): qualify every scheduler policy for output parity",
        "labels": [
          "testing",
          "track/E-testing-quality",
          "gen/now"
        ],
        "created_at": "2026-07-13T16:06:12Z",
        "updated_at": "2026-07-13T19:44:49Z",
        "closed_at": "2026-07-13T19:32:39Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/4537",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 4538,
        "state": "closed",
        "title": "test(quality): prove retry and mid-generation resume preserve semantics",
        "labels": [
          "testing",
          "track/E-testing-quality",
          "gen/now"
        ],
        "created_at": "2026-07-13T16:06:18Z",
        "updated_at": "2026-07-13T18:15:47Z",
        "closed_at": "2026-07-13T18:15:47Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/4538",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 4539,
        "state": "closed",
        "title": "test(quality): validate speculative decoding against target-only decode",
        "labels": [
          "testing",
          "track/E-testing-quality",
          "gen/now"
        ],
        "created_at": "2026-07-13T16:06:22Z",
        "updated_at": "2026-07-13T19:32:15Z",
        "closed_at": "2026-07-13T19:32:15Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/4539",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "speculative_generation",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 4540,
        "state": "closed",
        "title": "test(quality): set quantization-specific quality budgets",
        "labels": [
          "testing",
          "track/E-testing-quality",
          "gen/now",
          "class:infra"
        ],
        "created_at": "2026-07-13T16:06:24Z",
        "updated_at": "2026-07-13T20:30:45Z",
        "closed_at": "2026-07-13T20:30:45Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/4540",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 4541,
        "state": "closed",
        "title": "test(quality): qualify FP32, BF16, and FP16 execution quality deltas",
        "labels": [
          "testing",
          "track/E-testing-quality",
          "gen/now"
        ],
        "created_at": "2026-07-13T16:06:26Z",
        "updated_at": "2026-07-13T19:18:40Z",
        "closed_at": "2026-07-13T19:16:47Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/4541",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 4542,
        "state": "closed",
        "title": "test(quality): gate tensor-parallel degree output parity",
        "labels": [
          "testing",
          "track/E-testing-quality",
          "gen/now"
        ],
        "created_at": "2026-07-13T16:06:29Z",
        "updated_at": "2026-07-13T19:47:03Z",
        "closed_at": "2026-07-13T19:47:03Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/4542",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 4543,
        "state": "closed",
        "title": "test(quality): gate pipeline-parallel output parity and resumes",
        "labels": [
          "testing",
          "track/E-testing-quality",
          "gen/now"
        ],
        "created_at": "2026-07-13T16:06:31Z",
        "updated_at": "2026-07-13T19:33:33Z",
        "closed_at": "2026-07-13T19:31:44Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/4543",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 4544,
        "state": "closed",
        "title": "test(quality): maintain CPU, CUDA, ROCm, Metal, and Vulkan quality matrices",
        "labels": [
          "testing",
          "track/E-testing-quality",
          "gen/now"
        ],
        "created_at": "2026-07-13T16:06:33Z",
        "updated_at": "2026-07-13T19:36:28Z",
        "closed_at": "2026-07-13T19:31:27Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/4544",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 4545,
        "state": "closed",
        "title": "test(quality): verify model load and conversion preserve generation quality",
        "labels": [
          "testing",
          "track/E-testing-quality",
          "gen/now"
        ],
        "created_at": "2026-07-13T16:06:35Z",
        "updated_at": "2026-07-13T18:45:54Z",
        "closed_at": "2026-07-13T18:45:54Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/4545",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 4546,
        "state": "closed",
        "title": "test(quality): detect quality cliffs across context length and position scaling",
        "labels": [
          "testing",
          "track/E-testing-quality",
          "gen/now"
        ],
        "created_at": "2026-07-13T16:06:38Z",
        "updated_at": "2026-07-13T19:46:57Z",
        "closed_at": "2026-07-13T19:46:57Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/4546",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 4547,
        "state": "closed",
        "title": "test(quality): qualify OpenAI-compatible request and response semantics",
        "labels": [
          "openai-api",
          "testing",
          "track/E-testing-quality",
          "gen/now"
        ],
        "created_at": "2026-07-13T16:06:40Z",
        "updated_at": "2026-07-13T19:46:50Z",
        "closed_at": "2026-07-13T19:46:50Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/4547",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 4548,
        "state": "closed",
        "title": "test(quality): validate constrained and structured-output semantic quality",
        "labels": [
          "substrate",
          "testing",
          "track/E-testing-quality",
          "gen/now"
        ],
        "created_at": "2026-07-13T16:06:43Z",
        "updated_at": "2026-07-13T19:48:01Z",
        "closed_at": "2026-07-13T19:46:44Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/4548",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 4549,
        "state": "closed",
        "title": "test(quality): validate multi-turn state and report consistency",
        "labels": [
          "testing",
          "track/E-testing-quality",
          "gen/now"
        ],
        "created_at": "2026-07-13T16:06:45Z",
        "updated_at": "2026-07-13T19:46:37Z",
        "closed_at": "2026-07-13T19:46:37Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/4549",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 4550,
        "state": "closed",
        "title": "test(report-quality): create a versioned executive-report dogfood corpus",
        "labels": [
          "testing",
          "track/E-testing-quality",
          "gen/now",
          "class:infra"
        ],
        "created_at": "2026-07-13T16:06:47Z",
        "updated_at": "2026-07-14T18:31:05Z",
        "closed_at": "2026-07-14T18:31:05Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/4550",
        "categories": [
          "native_backends",
          "models_architectures",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 4559,
        "state": "closed",
        "title": "test(report-quality): enforce executive-report structure without freezing style",
        "labels": [
          "testing",
          "track/E-testing-quality",
          "gen/now",
          "class:infra"
        ],
        "created_at": "2026-07-13T16:07:10Z",
        "updated_at": "2026-07-14T18:35:32Z",
        "closed_at": "2026-07-14T18:35:32Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/4559",
        "categories": [
          "native_backends",
          "models_architectures",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 4627,
        "state": "closed",
        "title": "C4: coherence gate — prove long-prompt coherence holds on the int8 Q4_K decode path",
        "labels": [
          "gguf",
          "qwen",
          "testing",
          "generation",
          "gen/now",
          "class:infra"
        ],
        "created_at": "2026-07-14T05:55:20Z",
        "updated_at": "2026-07-15T00:59:56Z",
        "closed_at": "2026-07-15T00:59:56Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/4627",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "structured_generation_tools",
          "correctness_testing"
        ]
      },
      {
        "number": 4867,
        "state": "open",
        "title": "epic: Support prism-ml Ternary-Bonsai-27B (Qwen3.6 ternary GGUF) — VLM, hybrid attention, DSpark spec-decode",
        "labels": [
          "enhancement",
          "epic",
          "model-support",
          "track/A-model-support"
        ],
        "created_at": "2026-07-15T14:38:33Z",
        "updated_at": "2026-07-15T14:40:47Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/4867",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "build_platform_ops"
        ]
      },
      {
        "number": 4868,
        "state": "closed",
        "title": "feat(ggufload): ternary Q2_0 GGUF quant type — load + dequant (Bonsai spine)",
        "labels": [
          "model-support",
          "priority/P0",
          "quantization",
          "track/A-model-support",
          "generation",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-07-15T14:39:19Z",
        "updated_at": "2026-07-15T15:33:24Z",
        "closed_at": "2026-07-15T15:27:06Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/4868",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 4871,
        "state": "closed",
        "title": "feat(ggufload): 1-bit Q1_0 (g128) GGUF quant type for the Bonsai-27B binary build",
        "labels": [
          "model-support",
          "priority/P1",
          "quantization",
          "track/A-model-support",
          "generation",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-07-15T14:39:26Z",
        "updated_at": "2026-07-28T13:01:29Z",
        "closed_at": "2026-07-28T13:01:29Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/4871",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 4872,
        "state": "closed",
        "title": "feat(compute): CUDA ternary GEMM kernel (k_q2_0_gemm) + HAL Q2_0 Dtype",
        "labels": [
          "compute",
          "cuda",
          "gpu",
          "priority/P2",
          "track/C-gpu-backend",
          "generation",
          "gen/next",
          "class:dev"
        ],
        "created_at": "2026-07-15T14:39:29Z",
        "updated_at": "2026-07-16T20:00:46Z",
        "closed_at": "2026-07-16T20:00:46Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/4872",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 4875,
        "state": "closed",
        "title": "feat(model): wire Bonsai vision tower (VLM) end-to-end",
        "labels": [
          "model",
          "model-support",
          "priority/P2",
          "track/A-model-support",
          "generation",
          "gen/next",
          "class:dev"
        ],
        "created_at": "2026-07-15T14:39:52Z",
        "updated_at": "2026-08-07T06:46:14Z",
        "closed_at": "2026-08-07T06:46:14Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/4875",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 4880,
        "state": "open",
        "title": "bench: Ternary-Bonsai-27B head-to-head vs llama.cpp (numeric parity + tok/s)",
        "labels": [
          "benchmark",
          "performance",
          "priority/P2",
          "track/B-performance",
          "generation",
          "gen/next",
          "class:dev"
        ],
        "created_at": "2026-07-15T14:40:12Z",
        "updated_at": "2026-07-15T19:42:10Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/4880",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 4952,
        "state": "closed",
        "title": "fix(glm52): apply rank-local EP capacity plan before device-load refusal",
        "labels": [
          "priority/P0",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-07-15T15:09:53Z",
        "updated_at": "2026-08-17T05:01:36Z",
        "closed_at": "2026-08-17T05:01:36Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/4952",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 5261,
        "state": "closed",
        "title": "feat(model/spec): model-free prompt-lookup (n-gram) drafter feeding the shipped verify-accept substrate — zero draft head",
        "labels": [
          "agentic-serving",
          "model-arch",
          "performance",
          "speculative-decoding",
          "class:dev"
        ],
        "created_at": "2026-07-19T06:11:20Z",
        "updated_at": "2026-08-21T05:56:26Z",
        "closed_at": "2026-08-21T05:56:25Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/5261",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 5280,
        "state": "closed",
        "title": "feat(model): user-supplied per-tensor placement override (name-regex, first-match-wins) for the CPU-offload split",
        "labels": [
          "priority/P2",
          "class:dev"
        ],
        "created_at": "2026-07-19T06:15:46Z",
        "updated_at": "2026-07-21T06:12:13Z",
        "closed_at": "2026-07-21T06:12:13Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/5280",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 5281,
        "state": "closed",
        "title": "feat(model): graded expert-spill knob (--n-cpu-moe N) + auto-fit N to a measured device budget",
        "labels": [
          "priority/P2",
          "class:dev"
        ],
        "created_at": "2026-07-19T06:15:48Z",
        "updated_at": "2026-08-05T22:15:52Z",
        "closed_at": "2026-07-21T06:26:38Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/5281",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 5282,
        "state": "closed",
        "title": "feat(model/spec): model-free n-gram (prompt-lookup) drafter feeding the shipped verify-accept substrate",
        "labels": [
          "priority/P2",
          "class:dev"
        ],
        "created_at": "2026-07-19T06:16:01Z",
        "updated_at": "2026-07-19T06:39:03Z",
        "closed_at": "2026-07-19T06:39:03Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/5282",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 5283,
        "state": "closed",
        "title": "feat(engine): Hadamard-rotation outlier smoothing for quantized K/V cache (KV-quant ladder enhancement)",
        "labels": [
          "priority/P2",
          "class:dev"
        ],
        "created_at": "2026-07-19T06:16:03Z",
        "updated_at": "2026-07-21T08:41:47Z",
        "closed_at": "2026-07-21T08:41:47Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/5283",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 5284,
        "state": "closed",
        "title": "perf(model): NUMA-aware weight-mmap policy — MADV_RANDOM under NUMA, sequential hints otherwise",
        "labels": [
          "priority/P2",
          "class:dev"
        ],
        "created_at": "2026-07-19T06:16:05Z",
        "updated_at": "2026-07-21T07:27:59Z",
        "closed_at": "2026-07-21T07:27:59Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/5284",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 5285,
        "state": "closed",
        "title": "perf(cpu): online weight repack into SIMD-interleaved blocks selected by runtime CPU-feature detection",
        "labels": [
          "priority/P2",
          "class:dev"
        ],
        "created_at": "2026-07-19T06:16:08Z",
        "updated_at": "2026-08-07T21:01:04Z",
        "closed_at": "2026-08-07T21:01:04Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/5285",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 5306,
        "state": "open",
        "title": "docs(mac): cross-link the epic #2722 Mac results artifacts into the newcomer-facing Mac docs once #2723 lands numbers",
        "labels": [
          "documentation",
          "class:dev"
        ],
        "created_at": "2026-07-20T10:40:01Z",
        "updated_at": "2026-07-20T14:36:21Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/5306",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 5461,
        "state": "open",
        "title": "docs(repo): 31 markdown files in the repository root bury the front door",
        "labels": [
          "documentation",
          "class:frontdoor"
        ],
        "created_at": "2026-07-29T11:04:49Z",
        "updated_at": "2026-08-03T13:43:29Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/5461",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 5612,
        "state": "closed",
        "title": "feat(moe): ship the graded --n-cpu-moe N knob and admit on the activated working set (epic #5606 R1)",
        "labels": [
          "enhancement",
          "model-arch",
          "moe",
          "priority/P0",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-05T22:14:27Z",
        "updated_at": "2026-08-05T23:29:14Z",
        "closed_at": "2026-08-05T23:00:41Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/5612",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 6094,
        "state": "closed",
        "title": "bench(tokenizer): compare native BPE with llama.cpp and Hugging Face",
        "labels": [],
        "created_at": "2026-08-10T01:54:51Z",
        "updated_at": "2026-08-10T02:31:09Z",
        "closed_at": "2026-08-10T02:31:09Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/6094",
        "categories": [
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion"
        ]
      },
      {
        "number": 6196,
        "state": "open",
        "title": "Benchmark model/backend/precision coverage against runtime and external matrices",
        "labels": [
          "class:dev"
        ],
        "created_at": "2026-08-10T17:42:52Z",
        "updated_at": "2026-08-11T02:33:26Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/6196",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 6221,
        "state": "open",
        "title": "epic(model): broad quantization interoperability without a format winner",
        "labels": [
          "epic",
          "model-support",
          "priority/P1",
          "research",
          "gen/next"
        ],
        "created_at": "2026-08-10T18:53:06Z",
        "updated_at": "2026-08-10T21:47:33Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/6221",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 6231,
        "state": "closed",
        "title": "feat(model): map GGUF quantization types and metadata",
        "labels": [
          "model-support",
          "priority/P1",
          "gen/next",
          "class:dev"
        ],
        "created_at": "2026-08-10T19:03:48Z",
        "updated_at": "2026-08-19T01:08:06Z",
        "closed_at": "2026-08-19T01:08:06Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/6231",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 6232,
        "state": "closed",
        "title": "feat(engine): delegate quantized models to llama.cpp with capability checks",
        "labels": [
          "model-support",
          "priority/P1",
          "gen/next",
          "class:dev"
        ],
        "created_at": "2026-08-10T19:03:55Z",
        "updated_at": "2026-08-19T01:13:33Z",
        "closed_at": "2026-08-19T01:13:33Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/6232",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 6370,
        "state": "closed",
        "title": "Benchmark answer degeneration against provider and generation controls",
        "labels": [
          "class:dev"
        ],
        "created_at": "2026-08-10T22:56:10Z",
        "updated_at": "2026-08-17T04:58:10Z",
        "closed_at": "2026-08-17T04:58:10Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/6370",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 6851,
        "state": "open",
        "title": "bench(cache): witness pure-fak host L2 restore on exact Qwen3.6-27B-FP8 on GPU server",
        "labels": [
          "cuda",
          "gpu",
          "priority/P0",
          "current",
          "gen/now",
          "class:dev",
          "gated/hardware"
        ],
        "created_at": "2026-08-14T23:04:28Z",
        "updated_at": "2026-08-18T03:49:41Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/6851",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 6870,
        "state": "closed",
        "title": "feat(model): load the exact Qwen3.6-27B-FP8 checkpoint in the pure fak engine",
        "labels": [
          "cuda",
          "gpu",
          "model-arch",
          "priority/P0",
          "current",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-15T15:20:21Z",
        "updated_at": "2026-08-18T03:46:59Z",
        "closed_at": "2026-08-18T03:31:06Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/6870",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 8011,
        "state": "open",
        "title": "epic(model): first-class Qwen3.8-27B support across MacBook Metal and A100 CUDA",
        "labels": [
          "epic",
          "model-support",
          "priority/P0",
          "gen/now"
        ],
        "created_at": "2026-08-18T20:41:32Z",
        "updated_at": "2026-08-25T03:36:47Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/8011",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "build_platform_ops"
        ]
      },
      {
        "number": 8061,
        "state": "closed",
        "title": "bench(qwen38): capture exact MacBook Metal and A100 acceptance reports",
        "labels": [],
        "created_at": "2026-08-19T01:18:32Z",
        "updated_at": "2026-08-20T23:15:10Z",
        "closed_at": "2026-08-20T23:15:10Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8061",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "build_platform_ops"
        ]
      },
      {
        "number": 8064,
        "state": "closed",
        "title": "fix(guard): resolve friendly GGUF model references before local launch",
        "labels": [
          "gen/now"
        ],
        "created_at": "2026-08-19T01:32:37Z",
        "updated_at": "2026-08-22T13:43:30Z",
        "closed_at": "2026-08-22T13:43:30Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8064",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 8067,
        "state": "closed",
        "title": "fix(metal): prevent Qwen3.8 27B post-load memory overcommit",
        "labels": [],
        "created_at": "2026-08-19T01:53:35Z",
        "updated_at": "2026-08-19T08:57:40Z",
        "closed_at": "2026-08-19T06:09:24Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8067",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 8101,
        "state": "closed",
        "title": "perf(model): bound native Metal host materialization for Qwen3.8-27B",
        "labels": [
          "model-support",
          "performance",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-19T08:17:58Z",
        "updated_at": "2026-08-26T15:46:47Z",
        "closed_at": "2026-08-26T15:46:47Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8101",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 8129,
        "state": "open",
        "title": "experiment(qwen38): run controlled DGX multi-arm runtime and cache campaigns",
        "labels": [
          "benchmark",
          "model-support",
          "priority/P1",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-19T14:30:32Z",
        "updated_at": "2026-08-25T07:17:08Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/8129",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 8130,
        "state": "closed",
        "title": "research(qwen38): map upstream runtime, tool, context, and cache support",
        "labels": [
          "model-support",
          "priority/P1",
          "research",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-19T14:30:34Z",
        "updated_at": "2026-08-26T18:14:47Z",
        "closed_at": "2026-08-26T18:14:46Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8130",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 8298,
        "state": "open",
        "title": "ops(qwen38): run an always-on GCP coding endpoint through native fak",
        "labels": [
          "gen/now",
          "class:infra"
        ],
        "created_at": "2026-08-20T21:42:40Z",
        "updated_at": "2026-08-25T07:17:37Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/8298",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 8320,
        "state": "open",
        "title": "bench(qwen38): publish the quality-constrained top-10 quantization frontier",
        "labels": [
          "benchmark",
          "model-support",
          "priority/P0",
          "research",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-20T22:55:54Z",
        "updated_at": "2026-08-25T07:17:48Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/8320",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 8324,
        "state": "open",
        "title": "perf(model): route dense Qwen3.8 through resident Metal decode",
        "labels": [
          "model-support",
          "performance",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-20T23:16:55Z",
        "updated_at": "2026-08-26T16:55:15Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/8324",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 8325,
        "state": "open",
        "title": "perf(model): map streamed Q4_K checkpoints into Metal views",
        "labels": [
          "model-support",
          "performance",
          "gen/next",
          "class:dev"
        ],
        "created_at": "2026-08-20T23:18:23Z",
        "updated_at": "2026-08-26T07:36:53Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/8325",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 8326,
        "state": "closed",
        "title": "perf(metalgemm): specialize Q4_K for 4-8 Metal vectors",
        "labels": [
          "model-support",
          "performance",
          "gen/next"
        ],
        "created_at": "2026-08-20T23:19:05Z",
        "updated_at": "2026-08-21T05:42:26Z",
        "closed_at": "2026-08-21T05:42:26Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8326",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 8344,
        "state": "closed",
        "title": "perf(qwen38): add device-resident CUDA sequence prefill for Codex",
        "labels": [
          "gen/now"
        ],
        "created_at": "2026-08-21T00:54:54Z",
        "updated_at": "2026-08-22T21:20:50Z",
        "closed_at": "2026-08-22T21:20:50Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8344",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 8345,
        "state": "closed",
        "title": "feat(qwen38): add resident CUDA GDN sequence parity spine",
        "labels": [
          "gen/now"
        ],
        "created_at": "2026-08-21T01:22:47Z",
        "updated_at": "2026-08-21T04:17:30Z",
        "closed_at": "2026-08-21T03:58:02Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8345",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "build_platform_ops"
        ]
      },
      {
        "number": 8356,
        "state": "open",
        "title": "Bound Qwen3.8 16K long-context retrieval on native CUDA",
        "labels": [
          "class:dev"
        ],
        "created_at": "2026-08-21T03:33:28Z",
        "updated_at": "2026-08-25T07:17:57Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/8356",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 8360,
        "state": "closed",
        "title": "bench(qwen38): run the comparable MacBook Metal campaign",
        "labels": [
          "gen/now"
        ],
        "created_at": "2026-08-21T04:10:28Z",
        "updated_at": "2026-08-22T17:26:53Z",
        "closed_at": "2026-08-22T17:26:53Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8360",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing"
        ]
      },
      {
        "number": 8370,
        "state": "closed",
        "title": "feat(qwen38): execute the complete resident CUDA sequence spine",
        "labels": [
          "gen/now"
        ],
        "created_at": "2026-08-21T06:50:33Z",
        "updated_at": "2026-08-22T21:04:21Z",
        "closed_at": "2026-08-22T21:04:21Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8370",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "build_platform_ops"
        ]
      },
      {
        "number": 8374,
        "state": "open",
        "title": "feat(qwen38quantrun): ship maintained llama.cpp and vLLM adapter examples",
        "labels": [
          "class:dev"
        ],
        "created_at": "2026-08-21T07:02:17Z",
        "updated_at": "2026-08-25T07:18:02Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/8374",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "multimodal_embedding",
          "loader_conversion",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 8384,
        "state": "closed",
        "title": "bench(qwen38): measure a quality-constrained speculative decoding frontier",
        "labels": [
          "benchmark",
          "model-support",
          "performance",
          "priority/P1",
          "gen/now"
        ],
        "created_at": "2026-08-21T09:02:19Z",
        "updated_at": "2026-08-23T08:01:36Z",
        "closed_at": "2026-08-23T08:01:36Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8384",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 8385,
        "state": "closed",
        "title": "bench(qwen38): gate Mac fak-native intra-generation decode decay",
        "labels": [
          "benchmark",
          "model-support",
          "performance",
          "priority/P1",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-21T09:02:48Z",
        "updated_at": "2026-08-26T17:03:41Z",
        "closed_at": "2026-08-26T16:56:37Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8385",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 8389,
        "state": "closed",
        "title": "feat(modelperfobs): observe inference bottlenecks end to end",
        "labels": [
          "gen/now"
        ],
        "created_at": "2026-08-21T09:23:11Z",
        "updated_at": "2026-08-21T10:24:01Z",
        "closed_at": "2026-08-21T10:24:01Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8389",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 8390,
        "state": "closed",
        "title": "bench(qwen38): make pinned llama.cpp the numeric and performance oracle",
        "labels": [
          "model-support",
          "priority/P1",
          "gen/now"
        ],
        "created_at": "2026-08-21T09:25:52Z",
        "updated_at": "2026-08-21T09:46:08Z",
        "closed_at": "2026-08-21T09:46:08Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8390",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 8391,
        "state": "closed",
        "title": "feat(qwen38): enforce provenance for directly copied inference kernels",
        "labels": [
          "model-support",
          "priority/P1",
          "gen/now"
        ],
        "created_at": "2026-08-21T09:44:18Z",
        "updated_at": "2026-08-22T21:04:21Z",
        "closed_at": "2026-08-22T21:04:21Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8391",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 8392,
        "state": "closed",
        "title": "bench(qwen38): make pinned llama.cpp the numeric and performance oracle",
        "labels": [
          "model-support",
          "priority/P1",
          "gen/now"
        ],
        "created_at": "2026-08-21T09:44:37Z",
        "updated_at": "2026-08-22T20:26:08Z",
        "closed_at": "2026-08-22T20:26:08Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8392",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 8393,
        "state": "open",
        "title": "perf(qwen38): copy or adapt one profiled CUDA hot path from a leading runtime",
        "labels": [
          "model-support",
          "priority/P1",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-21T09:44:43Z",
        "updated_at": "2026-08-25T07:18:13Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/8393",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 8394,
        "state": "open",
        "title": "perf(qwen38): copy or adapt one profiled Metal hot path from llama.cpp or MLX-LM",
        "labels": [
          "metal",
          "model-support",
          "performance",
          "priority/P1",
          "track/B-performance",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-21T09:44:53Z",
        "updated_at": "2026-08-25T07:18:15Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/8394",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 8395,
        "state": "open",
        "title": "perf(qwen38): close the serving baseline with paged KV, prefix reuse, chunked prefill, and batching",
        "labels": [
          "model-support",
          "priority/P1",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-21T09:45:22Z",
        "updated_at": "2026-08-26T07:36:55Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/8395",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 8397,
        "state": "open",
        "title": "ops(dgxbridge): recover stale GPU server and GPU server control daemons",
        "labels": [
          "class:infra"
        ],
        "created_at": "2026-08-21T10:05:48Z",
        "updated_at": "2026-08-25T07:18:17Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/8397",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding"
        ]
      },
      {
        "number": 8401,
        "state": "closed",
        "title": "fix(qwen38): pin no-thinking campaign profile for exact effects",
        "labels": [],
        "created_at": "2026-08-21T10:19:11Z",
        "updated_at": "2026-08-22T05:15:52Z",
        "closed_at": "2026-08-22T04:33:41Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8401",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 8403,
        "state": "open",
        "title": "feat(modelperfobs): join backend engine metrics to requests",
        "labels": [
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-21T10:21:30Z",
        "updated_at": "2026-08-25T07:18:21Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/8403",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 8593,
        "state": "open",
        "title": "epic(plugin): \"everything is a plugin\" — plugin-shape the four remaining seams: model adapter, agent loop, session store, runtime tool admit",
        "labels": [
          "epic",
          "substrate",
          "gen/next",
          "class:dev"
        ],
        "created_at": "2026-08-22T15:56:07Z",
        "updated_at": "2026-08-22T15:59:49Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/8593",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 8598,
        "state": "open",
        "title": "epic(edge): qualify useful small models on low-resource phones and laptops",
        "labels": [
          "epic",
          "integration",
          "performance",
          "priority/P1",
          "research",
          "managed-context",
          "gen/second-next",
          "harness-native",
          "class:dev",
          "product"
        ],
        "created_at": "2026-08-22T16:59:20Z",
        "updated_at": "2026-08-22T17:01:21Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/8598",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 8600,
        "state": "closed",
        "title": "feat(edgequal): prove the low-resource phone-and-laptop offline spine",
        "labels": [
          "integration",
          "performance",
          "priority/P1",
          "research",
          "model-routing",
          "managed-context",
          "gen/second-next",
          "harness-native",
          "class:dev",
          "product"
        ],
        "created_at": "2026-08-22T17:00:57Z",
        "updated_at": "2026-08-26T15:34:30Z",
        "closed_at": "2026-08-26T15:34:30Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8600",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 8623,
        "state": "closed",
        "title": "bench(model): close hill climb on exact Qwen3.8-27B",
        "labels": [
          "gen/now"
        ],
        "created_at": "2026-08-22T18:22:45Z",
        "updated_at": "2026-08-22T22:28:42Z",
        "closed_at": "2026-08-22T21:18:56Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8623",
        "categories": [
          "models_architectures",
          "memory_kv_context"
        ]
      },
      {
        "number": 8627,
        "state": "closed",
        "title": "perf(qwen38): make resident CUDA Q4_K the proven default",
        "labels": [
          "cuda",
          "performance",
          "gen/now"
        ],
        "created_at": "2026-08-22T19:30:21Z",
        "updated_at": "2026-08-22T19:49:14Z",
        "closed_at": "2026-08-22T19:49:14Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8627",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 8628,
        "state": "closed",
        "title": "perf(qwen38): replace scalar CUDA Q4_K decode with a llama.cpp-class kernel",
        "labels": [
          "cuda",
          "performance",
          "gen/now"
        ],
        "created_at": "2026-08-22T19:30:59Z",
        "updated_at": "2026-08-22T20:29:49Z",
        "closed_at": "2026-08-22T20:29:49Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8628",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 8631,
        "state": "closed",
        "title": "perf(qwen38): accelerate resident GDN Q8 projections with DP4A",
        "labels": [
          "compute",
          "performance",
          "gen/now"
        ],
        "created_at": "2026-08-22T21:55:14Z",
        "updated_at": "2026-08-22T21:55:46Z",
        "closed_at": "2026-08-22T21:55:46Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8631",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "api_ui_docs"
        ]
      },
      {
        "number": 8632,
        "state": "closed",
        "title": "perf(qwen38): implement chunked resident GDN sequence prefill",
        "labels": [
          "compute",
          "performance",
          "gen/now"
        ],
        "created_at": "2026-08-22T22:40:55Z",
        "updated_at": "2026-08-22T23:43:35Z",
        "closed_at": "2026-08-22T23:43:35Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8632",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 8634,
        "state": "closed",
        "title": "perf(qwen38): reuse Q4_K weights across prompt-token panels",
        "labels": [
          "gen/now"
        ],
        "created_at": "2026-08-23T00:36:47Z",
        "updated_at": "2026-08-23T00:38:39Z",
        "closed_at": "2026-08-23T00:38:39Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8634",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 8635,
        "state": "open",
        "title": "perf(qwen38): implement Q8_1 DP4A Q4_K decode MMVQ",
        "labels": [
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-23T00:49:00Z",
        "updated_at": "2026-08-25T13:18:43Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/8635",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 8636,
        "state": "closed",
        "title": "perf(qwen38): bind large Q4_K panels to transient SGEMM",
        "labels": [
          "gen/now"
        ],
        "created_at": "2026-08-23T01:51:20Z",
        "updated_at": "2026-08-23T01:52:26Z",
        "closed_at": "2026-08-23T01:52:26Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8636",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 8637,
        "state": "closed",
        "title": "Accelerate large GDN Q8 input panels with SGEMM",
        "labels": [
          "priority/P0",
          "gen/now"
        ],
        "created_at": "2026-08-23T02:32:07Z",
        "updated_at": "2026-08-23T03:14:21Z",
        "closed_at": "2026-08-23T03:14:21Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8637",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 8638,
        "state": "closed",
        "title": "Route large generic Q8 panels through SGEMM",
        "labels": [
          "priority/P0",
          "gen/now"
        ],
        "created_at": "2026-08-23T02:54:39Z",
        "updated_at": "2026-08-23T03:14:22Z",
        "closed_at": "2026-08-23T03:14:22Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8638",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 8639,
        "state": "closed",
        "title": "Route large GDN output panels through SGEMM",
        "labels": [
          "priority/P0",
          "gen/now"
        ],
        "created_at": "2026-08-23T03:00:21Z",
        "updated_at": "2026-08-23T03:14:24Z",
        "closed_at": "2026-08-23T03:14:24Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8639",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 8640,
        "state": "closed",
        "title": "Default large prefill SGEMMs to compensated TF32",
        "labels": [
          "priority/P0",
          "gen/now"
        ],
        "created_at": "2026-08-23T03:14:42Z",
        "updated_at": "2026-08-23T03:22:11Z",
        "closed_at": "2026-08-23T03:22:11Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8640",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 8641,
        "state": "closed",
        "title": "Use compensated FP16 operands for large panels",
        "labels": [
          "priority/P0",
          "gen/now"
        ],
        "created_at": "2026-08-23T03:32:13Z",
        "updated_at": "2026-08-23T03:35:50Z",
        "closed_at": "2026-08-23T03:35:44Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8641",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing"
        ]
      },
      {
        "number": 8642,
        "state": "closed",
        "title": "Reduce prompt attention dots with warp shuffles",
        "labels": [
          "priority/P0",
          "gen/now"
        ],
        "created_at": "2026-08-23T03:43:50Z",
        "updated_at": "2026-08-23T03:45:16Z",
        "closed_at": "2026-08-23T03:45:16Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8642",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 8643,
        "state": "closed",
        "title": "Map each Qwen prompt-attention head to one warp",
        "labels": [
          "priority/P0",
          "gen/now"
        ],
        "created_at": "2026-08-23T03:50:08Z",
        "updated_at": "2026-08-23T03:51:09Z",
        "closed_at": "2026-08-23T03:51:09Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8643",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "build_platform_ops"
        ]
      },
      {
        "number": 8648,
        "state": "closed",
        "title": "bench(ultracode): factorially separate micro-context scoping from ordinary prefix reuse",
        "labels": [
          "benchmark",
          "performance",
          "priority/P1",
          "managed-context",
          "gen/now"
        ],
        "created_at": "2026-08-23T05:14:54Z",
        "updated_at": "2026-08-24T01:40:09Z",
        "closed_at": "2026-08-24T01:40:09Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8648",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 8650,
        "state": "open",
        "title": "bench(ultracode): test scoped-prefix savings across radix-cache runtimes",
        "labels": [
          "gen/next",
          "class:infra"
        ],
        "created_at": "2026-08-23T05:14:57Z",
        "updated_at": "2026-08-25T13:18:52Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/8650",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 8656,
        "state": "closed",
        "title": "Precompute Q4_K decode coefficients",
        "labels": [
          "gen/now"
        ],
        "created_at": "2026-08-23T07:22:15Z",
        "updated_at": "2026-08-23T07:46:52Z",
        "closed_at": "2026-08-23T07:46:52Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8656",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 8659,
        "state": "closed",
        "title": "Keep GDN recurrent state thread-local during prefill",
        "labels": [
          "gen/now"
        ],
        "created_at": "2026-08-23T08:48:18Z",
        "updated_at": "2026-08-23T08:50:38Z",
        "closed_at": "2026-08-23T08:50:38Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8659",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 8662,
        "state": "closed",
        "title": "perf(qwen38): default to pinned llama.cpp MTP delegation on supported CUDA",
        "labels": [
          "model-support",
          "performance",
          "priority/P1",
          "gen/now"
        ],
        "created_at": "2026-08-23T11:19:08Z",
        "updated_at": "2026-08-23T12:34:28Z",
        "closed_at": "2026-08-23T11:34:17Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8662",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 8664,
        "state": "closed",
        "title": "docs(native): make beating llama.cpp the canonical fak-native invariant",
        "labels": [
          "priority/P0",
          "gen/now"
        ],
        "created_at": "2026-08-23T18:35:01Z",
        "updated_at": "2026-08-23T19:12:09Z",
        "closed_at": "2026-08-23T19:12:09Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8664",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 8665,
        "state": "closed",
        "title": "fix(serve): default Qwen3.8 execution to fak-native, never implicit llama.cpp",
        "labels": [
          "priority/P0",
          "gen/now"
        ],
        "created_at": "2026-08-23T18:35:03Z",
        "updated_at": "2026-08-23T19:19:12Z",
        "closed_at": "2026-08-23T19:19:12Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8665",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 8666,
        "state": "closed",
        "title": "guard(native): reject silent llama.cpp substitution while allowing explicit comparison",
        "labels": [
          "priority/P0",
          "gen/now"
        ],
        "created_at": "2026-08-23T18:35:04Z",
        "updated_at": "2026-08-23T19:19:14Z",
        "closed_at": "2026-08-23T19:19:14Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8666",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 8672,
        "state": "closed",
        "title": "bench(ultracode): put confidence intervals around the scoped-prefix attribution",
        "labels": [
          "benchmark",
          "performance",
          "priority/P1",
          "research",
          "managed-context",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-23T18:47:21Z",
        "updated_at": "2026-08-26T17:28:57Z",
        "closed_at": "2026-08-26T17:28:57Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8672",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 8673,
        "state": "closed",
        "title": "research(ultracode): run negative controls that can refute the micro-context win",
        "labels": [
          "benchmark",
          "priority/P1",
          "research",
          "testing",
          "managed-context",
          "gen/now",
          "class:infra"
        ],
        "created_at": "2026-08-23T18:47:30Z",
        "updated_at": "2026-08-26T13:22:30Z",
        "closed_at": "2026-08-26T13:22:30Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8673",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing"
        ]
      },
      {
        "number": 8674,
        "state": "closed",
        "title": "bench(ultracode): map the task-complexity crossover for micro-context scoping",
        "labels": [
          "benchmark",
          "performance",
          "priority/P1",
          "managed-context",
          "gen/next",
          "class:dev"
        ],
        "created_at": "2026-08-23T18:47:35Z",
        "updated_at": "2026-08-26T13:22:23Z",
        "closed_at": "2026-08-26T13:22:23Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8674",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing"
        ]
      },
      {
        "number": 8675,
        "state": "closed",
        "title": "bench(ultracode): separate tokenizer effects from semantic context omission",
        "labels": [
          "benchmark",
          "model-support",
          "priority/P1",
          "research",
          "managed-context",
          "gen/next",
          "class:dev"
        ],
        "created_at": "2026-08-23T18:47:48Z",
        "updated_at": "2026-08-26T13:01:58Z",
        "closed_at": "2026-08-26T13:01:58Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8675",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing"
        ]
      },
      {
        "number": 8676,
        "state": "closed",
        "title": "obs(ultracode): reconcile prompt, prefix-cache, and wall-work receipts without double counting",
        "labels": [
          "benchmark",
          "observability",
          "performance",
          "priority/P1",
          "gen/now",
          "class:infra"
        ],
        "created_at": "2026-08-23T18:47:53Z",
        "updated_at": "2026-08-26T13:01:57Z",
        "closed_at": "2026-08-26T13:01:57Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8676",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 8677,
        "state": "closed",
        "title": "bench(ultracode): measure tool-result shape as the micro-context savings driver",
        "labels": [
          "benchmark",
          "performance",
          "priority/P1",
          "managed-context",
          "gen/next",
          "class:dev"
        ],
        "created_at": "2026-08-23T18:47:57Z",
        "updated_at": "2026-08-26T15:28:19Z",
        "closed_at": "2026-08-26T15:28:19Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8677",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 8678,
        "state": "closed",
        "title": "dogfood(ultracode): survive cache TTL expiry, compaction, and provider-clear boundaries",
        "labels": [
          "observability",
          "performance",
          "priority/P1",
          "managed-context",
          "gen/next",
          "class:infra",
          "dogfood"
        ],
        "created_at": "2026-08-23T18:48:02Z",
        "updated_at": "2026-08-26T15:28:13Z",
        "closed_at": "2026-08-26T15:28:13Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8678",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 8679,
        "state": "closed",
        "title": "bench(ultracode): net micro-context savings against routing and orchestration overhead",
        "labels": [
          "benchmark",
          "performance",
          "priority/P1",
          "managed-context",
          "gen/next",
          "class:dev"
        ],
        "created_at": "2026-08-23T18:48:06Z",
        "updated_at": "2026-08-26T00:11:11Z",
        "closed_at": "2026-08-26T00:11:11Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8679",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 8680,
        "state": "closed",
        "title": "test(ultracode): lock the scoped-prefix claim to a versioned regression corpus",
        "labels": [
          "benchmark",
          "priority/P1",
          "testing",
          "managed-context",
          "gen/next",
          "class:infra"
        ],
        "created_at": "2026-08-23T18:48:16Z",
        "updated_at": "2026-08-26T12:13:41Z",
        "closed_at": "2026-08-26T12:13:41Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8680",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing"
        ]
      },
      {
        "number": 8681,
        "state": "closed",
        "title": "product(ultracode): turn scoped-prefix evidence into a conservative default decision",
        "labels": [
          "performance",
          "priority/P1",
          "managed-context",
          "gen/next",
          "class:dev",
          "product"
        ],
        "created_at": "2026-08-23T18:48:23Z",
        "updated_at": "2026-08-26T12:13:41Z",
        "closed_at": "2026-08-26T12:13:41Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8681",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 8692,
        "state": "open",
        "title": "feat(model): add bounded native dense GPU-layer placement for 8 GiB Qwen3.8 parity",
        "labels": [
          "model",
          "performance",
          "priority/P1",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-23T19:27:03Z",
        "updated_at": "2026-08-26T05:24:47Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/8692",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 8693,
        "state": "closed",
        "title": "guard(native): enforce native-first semantics at commit time",
        "labels": [
          "priority/P0",
          "gen/now"
        ],
        "created_at": "2026-08-23T19:29:14Z",
        "updated_at": "2026-08-24T01:51:27Z",
        "closed_at": "2026-08-24T01:51:27Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8693",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 8694,
        "state": "closed",
        "title": "fix(qwen38quant): bind performance claims to execution engine",
        "labels": [
          "priority/P0",
          "gen/now"
        ],
        "created_at": "2026-08-23T19:29:17Z",
        "updated_at": "2026-08-24T01:58:30Z",
        "closed_at": "2026-08-24T01:58:30Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8694",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 8697,
        "state": "closed",
        "title": "perf(metal): close Qwen3.8-27B native decode parity gap on M3 Pro",
        "labels": [
          "gpu",
          "metal",
          "performance",
          "priority/P1",
          "class:dev"
        ],
        "created_at": "2026-08-24T01:18:37Z",
        "updated_at": "2026-08-26T15:46:50Z",
        "closed_at": "2026-08-26T15:46:50Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8697",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 8703,
        "state": "closed",
        "title": "feat(qwen-workers): enforce input-amplification budgets",
        "labels": [
          "priority/P1",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-24T01:30:52Z",
        "updated_at": "2026-08-26T11:53:24Z",
        "closed_at": "2026-08-26T11:53:24Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8703",
        "categories": [
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 8710,
        "state": "closed",
        "title": "Add IQ3_S support for Qwen3.8 native loading",
        "labels": [],
        "created_at": "2026-08-24T03:09:28Z",
        "updated_at": "2026-08-24T03:57:53Z",
        "closed_at": "2026-08-24T03:57:53Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8710",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 8711,
        "state": "closed",
        "title": "Add native partial GPU-layer offload for oversized dense GGUF models",
        "labels": [
          "class:dev"
        ],
        "created_at": "2026-08-24T03:50:01Z",
        "updated_at": "2026-08-26T05:24:51Z",
        "closed_at": "2026-08-26T05:24:51Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8711",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 8738,
        "state": "closed",
        "title": "feat(nativeperf): expose the raw-model hill-climb rung graph",
        "labels": [
          "gen/now"
        ],
        "created_at": "2026-08-24T16:11:14Z",
        "updated_at": "2026-08-24T18:28:45Z",
        "closed_at": "2026-08-24T18:28:45Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8738",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 8752,
        "state": "closed",
        "title": "feat(nativeperf): expose per-lever ablations and deterministic next experiment",
        "labels": [
          "gpu",
          "performance",
          "priority/P1",
          "gen/now"
        ],
        "created_at": "2026-08-24T17:28:46Z",
        "updated_at": "2026-08-24T19:03:56Z",
        "closed_at": "2026-08-24T19:03:56Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8752",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 8757,
        "state": "closed",
        "title": "feat(nativeperf): shift left with baseline/candidate experiment receipts",
        "labels": [
          "gpu",
          "performance",
          "priority/P1",
          "gen/now"
        ],
        "created_at": "2026-08-24T18:02:23Z",
        "updated_at": "2026-08-24T19:19:22Z",
        "closed_at": "2026-08-24T19:19:22Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8757",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 8782,
        "state": "closed",
        "title": "docs(video): show microcontext end to end with a fak-native Qwen 27B Mac run",
        "labels": [
          "documentation",
          "priority/P2",
          "brand",
          "gen/now",
          "class:frontdoor"
        ],
        "created_at": "2026-08-24T19:39:15Z",
        "updated_at": "2026-08-26T15:46:43Z",
        "closed_at": "2026-08-26T15:46:43Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8782",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 8795,
        "state": "open",
        "title": "capture(video): record Qwen3.8-27B fak-native on Apple Metal",
        "labels": [
          "documentation",
          "priority/P2",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-24T20:01:51Z",
        "updated_at": "2026-08-25T13:19:58Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/8795",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "loader_conversion",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 8820,
        "state": "open",
        "title": "perf(qwen38): replace serial CUDA prompt loop with true panel prefill",
        "labels": [
          "gpu",
          "model-support",
          "performance",
          "priority/P1",
          "gen/now",
          "class:dev",
          "gated/hardware"
        ],
        "created_at": "2026-08-25T03:34:21Z",
        "updated_at": "2026-08-25T03:34:21Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/8820",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 8821,
        "state": "open",
        "title": "perf(qwen38): implement one profiled native CUDA decode lever",
        "labels": [
          "gpu",
          "model-support",
          "performance",
          "priority/P1",
          "gen/now",
          "class:dev",
          "gated/hardware"
        ],
        "created_at": "2026-08-25T03:34:43Z",
        "updated_at": "2026-08-25T03:34:43Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/8821",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 8823,
        "state": "closed",
        "title": "research(metal): map OSS Qwen3.8 decode hot paths into ablatable native experiments",
        "labels": [
          "metal",
          "performance"
        ],
        "created_at": "2026-08-25T03:36:14Z",
        "updated_at": "2026-08-25T04:21:08Z",
        "closed_at": "2026-08-25T04:21:08Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8823",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "multimodal_embedding",
          "loader_conversion",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 8833,
        "state": "open",
        "title": "perf(metal): batch Qwen full-attention Q8/Q4_K projections in one call-owned buffer",
        "labels": [
          "gpu",
          "metal",
          "model-support",
          "performance",
          "priority/P1",
          "gen/now",
          "tier/T1-optimal",
          "tier/T2-required",
          "class:dev"
        ],
        "created_at": "2026-08-25T04:23:10Z",
        "updated_at": "2026-08-26T20:36:44Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/8833",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 8844,
        "state": "closed",
        "title": "feat(metalgemm): expose scoped Metal execution event snapshots",
        "labels": [
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-25T07:11:38Z",
        "updated_at": "2026-08-25T17:41:59Z",
        "closed_at": "2026-08-25T17:41:59Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8844",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 8846,
        "state": "closed",
        "title": "perf(metal): retain Qwen GDN recurrent state on device",
        "labels": [
          "gpu",
          "metal",
          "model-support",
          "performance",
          "priority/P1",
          "gen/now",
          "tier/T1-optimal",
          "tier/T2-required",
          "class:dev"
        ],
        "created_at": "2026-08-25T07:32:46Z",
        "updated_at": "2026-08-26T19:27:40Z",
        "closed_at": "2026-08-26T19:27:40Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8846",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 8848,
        "state": "open",
        "title": "bench(qwen38): run witnessed 100-cell cross-node hill climb",
        "labels": [
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-25T16:04:50Z",
        "updated_at": "2026-08-25T19:06:45Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/8848",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 8922,
        "state": "open",
        "title": "audit(progress): make closed performance work visible and evidence-retained",
        "labels": [
          "class:frontdoor"
        ],
        "created_at": "2026-08-25T20:36:25Z",
        "updated_at": "2026-08-26T06:24:03Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/8922",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 8923,
        "state": "open",
        "title": "perf(qwen38): force the next baseline wave through a borrowed hot path",
        "labels": [
          "class:dev"
        ],
        "created_at": "2026-08-25T20:36:52Z",
        "updated_at": "2026-08-26T06:24:05Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/8923",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 8965,
        "state": "closed",
        "title": "perf(metalgemm): adapt pinned vectorized Q4_K P=1 GEMV",
        "labels": [
          "metal",
          "performance",
          "priority/P1",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-25T22:32:35Z",
        "updated_at": "2026-08-26T04:18:14Z",
        "closed_at": "2026-08-26T04:18:14Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8965",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 8966,
        "state": "closed",
        "title": "docs(sota): register the Qwen GDN Metal oracle",
        "labels": [
          "gpu",
          "metal",
          "performance",
          "priority/P1",
          "gen/now",
          "tier/T1-optimal",
          "tier/T2-required",
          "class:dev"
        ],
        "created_at": "2026-08-25T22:32:36Z",
        "updated_at": "2026-08-25T23:03:48Z",
        "closed_at": "2026-08-25T23:03:48Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8966",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 8968,
        "state": "closed",
        "title": "perf(model): capture the Qwen3.8 Metal control-token profile",
        "labels": [
          "metal",
          "model-support",
          "performance",
          "priority/P1",
          "gen/now",
          "tier/T1-optimal",
          "tier/T2-required",
          "class:dev"
        ],
        "created_at": "2026-08-25T22:32:39Z",
        "updated_at": "2026-08-26T05:10:11Z",
        "closed_at": "2026-08-26T05:10:11Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8968",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 8969,
        "state": "closed",
        "title": "feat(model): add a model-lifetime mapped Q4_K span owner",
        "labels": [
          "metal",
          "model-support",
          "performance",
          "priority/P1",
          "gen/now",
          "tier/T1-optimal",
          "tier/T2-required",
          "class:dev"
        ],
        "created_at": "2026-08-25T22:32:42Z",
        "updated_at": "2026-08-25T23:47:32Z",
        "closed_at": "2026-08-25T23:47:32Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8969",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 8972,
        "state": "open",
        "title": "bench(qwen38): capture a quality-complete matched P32/T64 Metal receipt",
        "labels": [
          "gpu",
          "metal",
          "performance",
          "priority/P1",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-25T22:32:47Z",
        "updated_at": "2026-08-26T21:42:50Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/8972",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 8973,
        "state": "closed",
        "title": "perf(metalgemm): encode mixed Q4_K/Q8 QKV in one caller-owned command buffer",
        "labels": [
          "gpu",
          "metal",
          "performance",
          "priority/P1",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-25T22:32:48Z",
        "updated_at": "2026-08-26T05:20:39Z",
        "closed_at": "2026-08-26T05:20:39Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/8973",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 9020,
        "state": "closed",
        "title": "feat(nativeperf): expose full-session Metal counters for Qwen3.8 profiles",
        "labels": [
          "metal",
          "model-support",
          "performance",
          "priority/P1",
          "gen/now",
          "tier/T1-required",
          "tier/T1-optimal",
          "class:dev"
        ],
        "created_at": "2026-08-25T23:32:54Z",
        "updated_at": "2026-08-26T04:34:11Z",
        "closed_at": "2026-08-26T04:34:11Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/9020",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 9026,
        "state": "closed",
        "title": "bug(qwen38quantrun): admit the pinned b9828 oracle revision",
        "labels": [
          "priority/P1",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-26T00:01:41Z",
        "updated_at": "2026-08-26T00:54:52Z",
        "closed_at": "2026-08-26T00:54:52Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/9026",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 9030,
        "state": "closed",
        "title": "fix(qwen38quantrun): bind campaign reports to execution engine",
        "labels": [
          "priority/P1",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-26T00:09:51Z",
        "updated_at": "2026-08-26T00:27:00Z",
        "closed_at": "2026-08-26T00:27:00Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/9030",
        "categories": [
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 9034,
        "state": "closed",
        "title": "perf(metalgemm): implement mixed Q4K Q8 QKV command buffer #8973",
        "labels": [],
        "created_at": "2026-08-26T00:24:09Z",
        "updated_at": "2026-08-26T00:24:11Z",
        "closed_at": "2026-08-26T00:24:11Z",
        "url": "https://github.com/anthony-chaudhary/fak/pull/9034",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing"
        ]
      },
      {
        "number": 9035,
        "state": "closed",
        "title": "fix(model): implement measured streamed Q4K capacity bound #8971",
        "labels": [],
        "created_at": "2026-08-26T00:29:29Z",
        "updated_at": "2026-08-26T00:29:32Z",
        "closed_at": "2026-08-26T00:29:32Z",
        "url": "https://github.com/anthony-chaudhary/fak/pull/9035",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 9043,
        "state": "closed",
        "title": "feat(qwen38quantrun): bind repeated P32T64 native identity in oracle receipts",
        "labels": [
          "metal",
          "performance",
          "priority/P1",
          "qwen",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-26T01:01:24Z",
        "updated_at": "2026-08-26T01:25:28Z",
        "closed_at": "2026-08-26T01:25:28Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/9043",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 9044,
        "state": "closed",
        "title": "perf(model): keep Qwen3.8 Q8 minority projections Metal-resident on 36 GiB",
        "labels": [
          "metal",
          "model-support",
          "performance",
          "priority/P1",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-26T01:07:24Z",
        "updated_at": "2026-08-26T03:32:32Z",
        "closed_at": "2026-08-26T03:32:32Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/9044",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 9050,
        "state": "open",
        "title": "epic(perf): ship the top ten fak-native Qwen laptop improvements",
        "labels": [
          "cuda",
          "epic",
          "gpu",
          "performance",
          "priority/P0",
          "qwen",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-26T05:32:58Z",
        "updated_at": "2026-08-26T15:07:26Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/9050",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 9059,
        "state": "closed",
        "title": "bench(qwen38): ablate streamed Q4_K load concurrency on M3 Pro",
        "labels": [
          "benchmark",
          "metal",
          "model-support",
          "performance",
          "priority/P1",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-26T06:15:38Z",
        "updated_at": "2026-08-26T12:17:55Z",
        "closed_at": "2026-08-26T12:17:27Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/9059",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 9060,
        "state": "closed",
        "title": "fix(serve): preserve the measured FREE_CPU streamed-Q4K envelope",
        "labels": [],
        "created_at": "2026-08-26T06:22:58Z",
        "updated_at": "2026-08-26T06:36:00Z",
        "closed_at": "2026-08-26T06:36:00Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/9060",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 9069,
        "state": "closed",
        "title": "fix(model): charge Qwen hybrid context plans for actual KV geometry",
        "labels": [
          "metal",
          "performance",
          "priority/P1",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-26T07:03:27Z",
        "updated_at": "2026-08-26T07:26:53Z",
        "closed_at": "2026-08-26T07:26:53Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/9069",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 9071,
        "state": "closed",
        "title": "feat(native): expose opt-in token-indexed decode traces",
        "labels": [
          "enhancement",
          "model-support",
          "performance",
          "priority/P1",
          "track/B-performance",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-26T07:05:07Z",
        "updated_at": "2026-08-26T12:36:57Z",
        "closed_at": "2026-08-26T12:36:56Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/9071",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 9074,
        "state": "open",
        "title": "perf(model): co-batch resident Qwen hybrid decode rows",
        "labels": [
          "batching",
          "metal",
          "model-support",
          "performance",
          "priority/P1",
          "qwen",
          "gen/next",
          "tier/T1-optimal",
          "tier/T2-required",
          "class:dev"
        ],
        "created_at": "2026-08-26T07:25:20Z",
        "updated_at": "2026-08-26T07:35:58Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/9074",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 9076,
        "state": "closed",
        "title": "perf(modelengine): preserve Qwen hybrid state across paged swap preemption",
        "labels": [
          "batching",
          "model-support",
          "performance",
          "priority/P1",
          "qwen",
          "gen/next",
          "tier/T1-optimal",
          "tier/T2-required",
          "class:dev"
        ],
        "created_at": "2026-08-26T07:32:29Z",
        "updated_at": "2026-08-26T12:00:48Z",
        "closed_at": "2026-08-26T12:00:48Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/9076",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 9079,
        "state": "closed",
        "title": "feat(serve): expose a bounded native admission token budget",
        "labels": [
          "enhancement",
          "agentic-serving",
          "metal",
          "performance",
          "priority/P1",
          "qwen",
          "gen/now",
          "tier/T1-optimal",
          "tier/T2-required",
          "class:dev"
        ],
        "created_at": "2026-08-26T08:10:55Z",
        "updated_at": "2026-08-26T08:38:00Z",
        "closed_at": "2026-08-26T08:26:08Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/9079",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 9080,
        "state": "closed",
        "title": "feat(serve): expose native scheduler token admission budget",
        "labels": [
          "enhancement",
          "agentic-serving",
          "performance",
          "priority/P1",
          "qwen",
          "gen/now",
          "tier/T1-optimal",
          "tier/T2-required",
          "class:dev"
        ],
        "created_at": "2026-08-26T08:12:12Z",
        "updated_at": "2026-08-26T08:18:44Z",
        "closed_at": "2026-08-26T08:18:44Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/9080",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 9083,
        "state": "closed",
        "title": "perf(agent): bound production Qwen Q4_K prefill panels",
        "labels": [
          "model-support",
          "performance",
          "priority/P1",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-26T09:08:19Z",
        "updated_at": "2026-08-26T09:44:45Z",
        "closed_at": "2026-08-26T09:32:32Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/9083",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 9089,
        "state": "closed",
        "title": "perf(agent): expose bounded Qwen Q4_K prefill chunk tuning",
        "labels": [
          "model-support",
          "performance",
          "priority/P1",
          "gen/next",
          "tier/T2-required",
          "tier/T2-optimal",
          "class:dev"
        ],
        "created_at": "2026-08-26T10:37:17Z",
        "updated_at": "2026-08-26T16:52:54Z",
        "closed_at": "2026-08-26T16:52:54Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/9089",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 9090,
        "state": "closed",
        "title": "feat(agent): bind Qwen prefill chunk width into native receipts",
        "labels": [
          "model-support",
          "performance",
          "priority/P1",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-26T10:37:45Z",
        "updated_at": "2026-08-26T10:56:01Z",
        "closed_at": "2026-08-26T10:55:35Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/9090",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 9093,
        "state": "open",
        "title": "perf(metal): run Qwen GDN sequence prefill as one resident operation",
        "labels": [
          "metal",
          "performance",
          "priority/P1",
          "gen/now",
          "tier/T1-optimal",
          "tier/T2-required",
          "class:dev"
        ],
        "created_at": "2026-08-26T12:18:03Z",
        "updated_at": "2026-08-26T18:11:26Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/9093",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 9094,
        "state": "closed",
        "title": "perf(qwen38): pre-reserve chunked full-attention KV on M3 Pro",
        "labels": [
          "metal",
          "model-support",
          "performance",
          "priority/P1",
          "gen/now",
          "tier/T1-optimal",
          "tier/T2-required",
          "class:dev"
        ],
        "created_at": "2026-08-26T12:24:12Z",
        "updated_at": "2026-08-26T13:38:58Z",
        "closed_at": "2026-08-26T13:38:58Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/9094",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 9097,
        "state": "closed",
        "title": "perf(model): wire resident Metal GDN into Q4_K prefill",
        "labels": [
          "metal",
          "performance",
          "priority/P1",
          "gen/now",
          "tier/T1-optimal",
          "tier/T2-required",
          "class:dev"
        ],
        "created_at": "2026-08-26T12:26:33Z",
        "updated_at": "2026-08-26T14:28:11Z",
        "closed_at": "2026-08-26T14:28:11Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/9097",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 9102,
        "state": "closed",
        "title": "perf(metalgemm): reuse Qwen Q4_K gate/up output slab across prefill",
        "labels": [
          "metal",
          "performance",
          "priority/P1",
          "generation",
          "gen/now",
          "tier/T1-optimal",
          "tier/T2-required",
          "class:dev"
        ],
        "created_at": "2026-08-26T14:01:52Z",
        "updated_at": "2026-08-26T15:40:32Z",
        "closed_at": "2026-08-26T15:39:22Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/9102",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 9103,
        "state": "closed",
        "title": "fix(metalgemm): drain Qwen GDN panel resources per call",
        "labels": [
          "metal",
          "performance",
          "priority/P1",
          "generation",
          "gen/now",
          "tier/T1-optimal",
          "tier/T2-required",
          "class:dev"
        ],
        "created_at": "2026-08-26T14:07:43Z",
        "updated_at": "2026-08-26T14:29:58Z",
        "closed_at": "2026-08-26T14:29:58Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/9103",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 9108,
        "state": "closed",
        "title": "bench(model): rerun resident GDN canary after panel drain",
        "labels": [
          "documentation",
          "benchmark",
          "metal",
          "performance",
          "priority/P1",
          "generation",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-26T14:33:58Z",
        "updated_at": "2026-08-26T15:09:19Z",
        "closed_at": "2026-08-26T15:09:19Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/9108",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 9117,
        "state": "closed",
        "title": "feat(model): own guarded Darwin native-canary execution",
        "labels": [
          "model",
          "performance",
          "priority/P1",
          "gen/now",
          "tier/T1-optimal",
          "tier/T2-required",
          "class:dev"
        ],
        "created_at": "2026-08-26T15:06:04Z",
        "updated_at": "2026-08-26T18:13:11Z",
        "closed_at": "2026-08-26T18:13:11Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/9117",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 9122,
        "state": "closed",
        "title": "research(qwen38): study Qwen3.8-Flash-Next architecture and fak opportunities",
        "labels": [
          "model-support",
          "priority/P1",
          "gen/now"
        ],
        "created_at": "2026-08-26T15:27:20Z",
        "updated_at": "2026-08-26T17:41:32Z",
        "closed_at": "2026-08-26T15:45:12Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/9122",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 9123,
        "state": "closed",
        "title": "feat(qwen4exp): build native four-layer correctness oracle",
        "labels": [
          "model-support",
          "priority/P1",
          "qwen",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-26T15:38:30Z",
        "updated_at": "2026-08-26T17:38:59Z",
        "closed_at": "2026-08-26T17:21:51Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/9123",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 9131,
        "state": "closed",
        "title": "product(local): define the embeddable Mac/Metal inference layer for shipped desktop apps",
        "labels": [
          "gen/now"
        ],
        "created_at": "2026-08-26T15:48:59Z",
        "updated_at": "2026-08-26T16:35:17Z",
        "closed_at": "2026-08-26T16:26:54Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/9131",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 9149,
        "state": "open",
        "title": "spine(localapp): ship signed Mac helper + Swift SDK job-apply drop-in",
        "labels": [
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-26T16:01:18Z",
        "updated_at": "2026-08-26T17:07:51Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/9149",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 9192,
        "state": "open",
        "title": "perf(metalgemm): fuse Qwen Q4_K/Q6_K prefill MLP on Metal",
        "labels": [
          "gpu",
          "metal",
          "performance",
          "priority/P1",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-26T16:42:41Z",
        "updated_at": "2026-08-26T17:12:33Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/9192",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 9202,
        "state": "closed",
        "title": "perf(metalgemm): specialize Q4_K simdgroup MMA for P32",
        "labels": [
          "gpu",
          "metal",
          "performance",
          "priority/P1",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-26T17:19:04Z",
        "updated_at": "2026-08-26T18:56:37Z",
        "closed_at": "2026-08-26T18:56:37Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/9202",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 9204,
        "state": "open",
        "title": "epic(model): QWEN38-FLASH-NEXT fak-native support",
        "labels": [
          "epic",
          "model-support",
          "priority/P0",
          "gen/now"
        ],
        "created_at": "2026-08-26T17:38:40Z",
        "updated_at": "2026-08-26T17:46:47Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/9204",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 9210,
        "state": "open",
        "title": "perf(qwen4exp): ship exact Metal kernels and unified-memory plan",
        "labels": [
          "model-support",
          "performance",
          "priority/P0",
          "qwen",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-26T17:41:01Z",
        "updated_at": "2026-08-26T17:41:01Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/9210",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 9212,
        "state": "open",
        "title": "bench(qwen4exp): publish matched CUDA Metal and runtime frontier",
        "labels": [
          "gpu",
          "model-support",
          "performance",
          "priority/P0",
          "qwen",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-26T17:41:06Z",
        "updated_at": "2026-08-26T17:41:07Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/9212",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing"
        ]
      },
      {
        "number": 9214,
        "state": "closed",
        "title": "docs(qwen4exp): publish support matrix rollback and upstream watch",
        "labels": [
          "documentation",
          "model-support",
          "priority/P1",
          "qwen",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-26T17:41:11Z",
        "updated_at": "2026-08-26T18:21:31Z",
        "closed_at": "2026-08-26T18:21:31Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/9214",
        "categories": [
          "native_backends",
          "models_architectures",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 9216,
        "state": "open",
        "title": "perf(metal): fuse Qwen linear-attention Q8 projections with resident GDN",
        "labels": [],
        "created_at": "2026-08-26T17:52:25Z",
        "updated_at": "2026-08-26T17:52:25Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/9216",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 9224,
        "state": "open",
        "title": "refactor(slop): retire model-adjacent clean-file duplication cohort",
        "labels": [
          "priority/P1",
          "current",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-26T18:18:32Z",
        "updated_at": "2026-08-26T21:45:21Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/9224",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 9230,
        "state": "open",
        "title": "perf(metalgraph): eliminate per-op waits and host materialization from Qwen prefill",
        "labels": [
          "model",
          "performance",
          "priority/P1",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-26T18:24:09Z",
        "updated_at": "2026-08-26T20:51:15Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/9230",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 9231,
        "state": "open",
        "title": "feat(model): land session-owned Qwen3.5 Metal GDN decode state",
        "labels": [
          "gpu",
          "metal",
          "model-support"
        ],
        "created_at": "2026-08-26T18:25:27Z",
        "updated_at": "2026-08-26T21:13:11Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/9231",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 9236,
        "state": "closed",
        "title": "bench(metalprof): decompose the 25x Qwen prefill gap against llama.cpp",
        "labels": [
          "model",
          "performance",
          "priority/P1",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-26T18:34:54Z",
        "updated_at": "2026-08-26T20:42:10Z",
        "closed_at": "2026-08-26T20:42:10Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/9236",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "build_platform_ops"
        ]
      },
      {
        "number": 9238,
        "state": "closed",
        "title": "perf(metalgemm): adapt llama.cpp Q6_K tiled MM for P32 prefill",
        "labels": [
          "model",
          "performance",
          "priority/P1",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-26T18:55:30Z",
        "updated_at": "2026-08-26T20:53:06Z",
        "closed_at": "2026-08-26T20:53:06Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/9238",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 9246,
        "state": "closed",
        "title": "fix(metalgemm): restore issue8833 Q4_K encoder link contract",
        "labels": [
          "bug",
          "metal",
          "priority/P1",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-26T19:54:04Z",
        "updated_at": "2026-08-26T20:06:41Z",
        "closed_at": "2026-08-26T20:06:41Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/9246",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 9249,
        "state": "closed",
        "title": "fix(metalgemm): bind Q8 output geometry in mixed-QKV encoder",
        "labels": [
          "bug",
          "metal",
          "priority/P1",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-26T20:02:13Z",
        "updated_at": "2026-08-26T20:24:03Z",
        "closed_at": "2026-08-26T20:24:03Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/9249",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops"
        ]
      },
      {
        "number": 9257,
        "state": "open",
        "title": "feat(metal): implement Qwen sequence-prefill backend contract",
        "labels": [
          "metal",
          "model",
          "performance",
          "priority/P1",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-26T20:30:26Z",
        "updated_at": "2026-08-26T21:49:46Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/9257",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "api_ui_docs"
        ]
      },
      {
        "number": 9259,
        "state": "closed",
        "title": "refactor(compute): add caller-owned Metal command encoder",
        "labels": [
          "metal",
          "performance",
          "priority/P1",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-26T20:51:03Z",
        "updated_at": "2026-08-26T21:30:45Z",
        "closed_at": "2026-08-26T21:30:45Z",
        "url": "https://github.com/anthony-chaudhary/fak/issues/9259",
        "categories": [
          "native_backends",
          "models_architectures",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs"
        ]
      },
      {
        "number": 9268,
        "state": "open",
        "title": "research(vllm): exhaustively index upstream work and prioritize FAK tickets",
        "labels": [
          "priority/P1",
          "research",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-26T21:52:49Z",
        "updated_at": "2026-08-26T22:02:50Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/9268",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "multimodal_embedding",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 9270,
        "state": "open",
        "title": "research(llamacpp): exhaustively index upstream and prioritize FAK tickets",
        "labels": [
          "priority/P1",
          "research",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-26T21:54:10Z",
        "updated_at": "2026-08-26T21:54:10Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/9270",
        "categories": [
          "native_backends",
          "models_architectures",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "api_ui_docs",
          "training_adapters"
        ]
      },
      {
        "number": 9273,
        "state": "open",
        "title": "epic(architecture): scale fak around typed composition, shared resource lifecycle, and causal evidence",
        "labels": [
          "agentic-serving",
          "epic",
          "model-arch",
          "observability",
          "priority/P1",
          "track/G-foundation",
          "context-engineering",
          "gen/next"
        ],
        "created_at": "2026-08-26T22:02:06Z",
        "updated_at": "2026-08-26T22:02:06Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/9273",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 9275,
        "state": "open",
        "title": "research(runtime): declare and index the related-system adjacency set",
        "labels": [
          "priority/P1",
          "research",
          "gen/now",
          "class:dev"
        ],
        "created_at": "2026-08-26T22:02:08Z",
        "updated_at": "2026-08-26T22:07:57Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/9275",
        "categories": [
          "models_architectures",
          "multimodal_embedding",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 9279,
        "state": "open",
        "title": "epic(architecture): scale fak around typed composition, shared resource lifecycle, and causal evidence",
        "labels": [
          "agentic-serving",
          "epic",
          "model-arch",
          "observability",
          "priority/P1",
          "track/G-foundation",
          "context-engineering",
          "gen/next"
        ],
        "created_at": "2026-08-26T22:03:11Z",
        "updated_at": "2026-08-26T22:05:20Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/9279",
        "categories": [
          "native_backends",
          "quantization",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "multimodal_embedding",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 9280,
        "state": "open",
        "title": "feat(architecture): ship the first validated composition snapshot and receipt",
        "labels": [
          "agentic-serving",
          "model-arch",
          "observability",
          "priority/P1",
          "testing",
          "track/G-foundation",
          "gen/now"
        ],
        "created_at": "2026-08-26T22:03:27Z",
        "updated_at": "2026-08-26T22:03:27Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/9280",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 9292,
        "state": "open",
        "title": "perf(specdecode): adapt draft depth from acceptance and evaluate DFlash2 selection",
        "labels": [
          "performance",
          "priority/P1",
          "speculative-decoding",
          "gen/next",
          "class:dev"
        ],
        "created_at": "2026-08-26T22:12:49Z",
        "updated_at": "2026-08-26T22:12:49Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/9292",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "speculative_generation",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 9293,
        "state": "open",
        "title": "perf(grammar): profile and remove native constrained-sampling bottlenecks",
        "labels": [
          "performance",
          "priority/P1",
          "gen/next",
          "class:dev"
        ],
        "created_at": "2026-08-26T22:12:51Z",
        "updated_at": "2026-08-26T22:12:51Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/9293",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "structured_generation_tools",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops",
          "training_adapters"
        ]
      },
      {
        "number": 9294,
        "state": "open",
        "title": "feat(kvcache): attach token lineage to native KV cells for exact reuse witnesses",
        "labels": [
          "model",
          "performance",
          "priority/P1",
          "gen/next",
          "class:dev"
        ],
        "created_at": "2026-08-26T22:12:53Z",
        "updated_at": "2026-08-26T22:12:53Z",
        "closed_at": null,
        "url": "https://github.com/anthony-chaudhary/fak/issues/9294",
        "categories": [
          "native_backends",
          "models_architectures",
          "memory_kv_context",
          "scheduling_serving",
          "loader_conversion",
          "correctness_testing",
          "build_platform_ops"
        ]
      }
    ]
  },
  "priority_queue": [
    {
      "rank": 1,
      "disposition": "execute",
      "fak_issue": 9204,
      "priority": "P0",
      "reason": "Qwen3.8-Flash-Next is the active native model spine; upstream #27742 is reference-only and is already decomposed into #9205-#9213.",
      "upstream": [
        27742
      ],
      "fak_issue_title": "epic(model): QWEN38-FLASH-NEXT fak-native support",
      "fak_issue_state": "open"
    },
    {
      "rank": 2,
      "disposition": "execute",
      "fak_issue": 9292,
      "priority": "P1",
      "reason": "New gap: adaptive MTP depth and DFlash2 selector evaluation were not dispatchably covered by #23/#4202.",
      "upstream": [
        27210,
        27342
      ],
      "fak_issue_title": null,
      "fak_issue_state": null
    },
    {
      "rank": 3,
      "disposition": "execute",
      "fak_issue": 9293,
      "priority": "P1",
      "reason": "New gap: correctness seams existed, but no open grammar-sampling performance ticket matched upstream roadmap #4218.",
      "upstream": [
        4218
      ],
      "fak_issue_title": null,
      "fak_issue_state": null
    },
    {
      "rank": 4,
      "disposition": "execute",
      "fak_issue": 9294,
      "priority": "P1",
      "reason": "New gap: add diagnostic token lineage to native KV ownership, building on RAC address/lifecycle issues #8464/#8468.",
      "upstream": [
        27762
      ],
      "fak_issue_title": null,
      "fak_issue_state": null
    },
    {
      "rank": 5,
      "disposition": "existing",
      "fak_issue": 8394,
      "priority": "P1",
      "reason": "Existing borrow-one-profiled-Metal-hot-path ticket covers compile-time specialization candidate #4085.",
      "upstream": [
        4085
      ],
      "fak_issue_title": "perf(qwen38): copy or adapt one profiled Metal hot path from llama.cpp or MLX-LM",
      "fak_issue_state": "open"
    },
    {
      "rank": 6,
      "disposition": "existing",
      "fak_issue": 50,
      "priority": "P0",
      "reason": "Existing disaggregated-serving epic covers upstream roadmap #21266; no duplicate child created.",
      "upstream": [
        21266
      ],
      "fak_issue_title": "epic(serving): large-scale disaggregated serving  -  dual-track (RIDE + NATIVE) over one shared spine",
      "fak_issue_state": "open"
    },
    {
      "rank": 7,
      "disposition": "existing",
      "fak_issue": 5127,
      "priority": "unlabeled",
      "reason": "Existing per-NUMA weight-replication ticket covers upstream PR #16000.",
      "upstream": [
        16000
      ],
      "fak_issue_title": "perf(model): per-NUMA-node Q4_K weight replication + barrier-free per-node decode schedule (Qwen3.6 CPU decode)",
      "fak_issue_state": "open"
    },
    {
      "rank": 8,
      "disposition": "existing",
      "fak_issue": 4867,
      "priority": "unlabeled",
      "reason": "Existing ternary Qwen/GGUF epic is the nearest product lane for upstream TQ2_0 kernels #11183.",
      "upstream": [
        11183
      ],
      "fak_issue_title": "epic: Support prism-ml Ternary-Bonsai-27B (Qwen3.6 ternary GGUF) — VLM, hybrid attention, DSpark spec-decode",
      "fak_issue_state": "open"
    },
    {
      "rank": 9,
      "disposition": "watch",
      "fak_issue": null,
      "priority": "watch",
      "reason": "Windows/Vulkan checkpoint crash #27560 is outside the current native Metal/CUDA critical path; retain in index until a matching FAK Vulkan envelope exists.",
      "upstream": [
        27560
      ]
    },
    {
      "rank": 10,
      "disposition": "exclude",
      "fak_issue": null,
      "priority": "exclude",
      "reason": "Generic upstream memory leak report #27725 lacks a reproducible FAK-native failure or matching envelope.",
      "upstream": [
        27725
      ]
    }
  ],
  "coverage_receipts": {
    "recent_commit_count": 119,
    "recent_commit_range": [
      "2026-08-20T03:40:19Z",
      "2026-08-26T21:34:28Z"
    ],
    "release_count": 6958,
    "latest_release": {
      "tag": "b10642",
      "published_at": "2026-08-26T22:00:33Z"
    },
    "raw_artifacts": {
      "upstream_open_pages_sha256": "4855ab1dcc94b884dffcb437da4df87892c4821811b88029a376add8eb757510",
      "upstream_tree_sha256": "e1d73eb4eb422c626a78e93ca057f232ad37a3ea52f22028395ef62ce9565f48",
      "fak_search_union_sha256": "49a4d4707d3ee6cd5d225c148633da588a6ea392c9d732eb34d90fe78cae567d"
    }
  }
}