Skip to the content.

SGLang upstream study — 2026-08-27

Verdict: the pinned tree is complete, every declared source class and all 12 candidates are accounted for, and no unduplicated implementation ticket was created. The forge corpus is partial, not complete: both capture attempts paginated the six endpoint classes, but fak study-forge rejected cross-endpoint PR identity consistency.

Pin and capture receipt

The complete receipt is gzip-compressed for transport (33.8 MB; 165.2 MB expanded): ../inventory/sgl-project-sglang-corpus.jsonl.gz. Full class accounting is in source-completeness.json.

Candidate ledger

Candidate Disposition Existing fak work Representative upstream evidence
scheduler-admission ADAPT #35, #36, #8395 https://github.com/sgl-project/sglang/pull/17026
radix-prefix-reuse COPY #3889, #3890, #3891, #41, #8395 https://github.com/sgl-project/sglang/issues/26618
chunked-prefill-continuous-batching ADAPT #36, #282, #8395 https://github.com/sgl-project/sglang/pull/35300
prefill-decode-disaggregation INTEGRATE #28, #29, #37, #53, #79 https://github.com/sgl-project/sglang/pull/35224
speculative-decoding WATCH #23 https://github.com/sgl-project/sglang/pull/21272
structured-generation INTEGRATE #26, #8382 https://github.com/sgl-project/sglang/pull/28804
multimodal-moe WATCH #25, #290, #399, #4875, #5777 https://github.com/sgl-project/sglang/pull/27602
distributed-execution ADAPT #25 https://github.com/sgl-project/sglang/issues/22084
observability INTEGRATE #216, #5629, #5631, #6801 https://github.com/sgl-project/sglang/pull/23169
kernel-runtime-integration REJECT #8395 https://github.com/sgl-project/sglang/issues/26715
failure-reliability ADAPT #8382, #8395 https://github.com/sgl-project/sglang/pull/32118
compatibility-benchmark INTEGRATE #39, #44, #6473, #6474 https://github.com/sgl-project/sglang/tree/536f570e6692eec0656ef9689db7591ca1d0e0a7

Dispositions mean: COPY a narrow proven mechanism while retaining fak ownership; ADAPT its invariant to a fak seam; INTEGRATE only through an explicit governed adapter; WATCH until the operating envelope justifies work; REJECT wholesale adoption or silent fallback. SGLang never becomes an automatic recovery path for native inference.

The machine ledger is candidate-ledger.json. It records operating envelope, next-best alternative, evidence, and ticket mapping for every candidate.

Dedup result

dedup-map.json accounts for all 271 issues returned by the repository-wide sglang tracker query and explicitly maps every candidate to existing work. Direct anchors include #23–#44, #216/#282/#290/#322, #3889–#3891, #5629/#5631, #6473/#6474, #6801, #8382, and #8395.

New issues created: none. The study found no genuine unduplicated gap that cleared the existing serving, compatibility, parser, cache, distributed, multimodal, observability, or benchmark tickets.

Completeness boundary

Deterministic validation

Run:

gzip -dc docs/research/inventory/sgl-project-sglang-corpus.jsonl.gz > $TMPDIR/sglang-corpus.json`npython -m json.tool $TMPDIR/sglang-corpus.json > /dev/null
Get-ChildItem docs/research/sglang-2026-08-27/*.json | ForEach-Object { python -m json.tool $_.FullName > $null }
fak study-forge validate --receipt $TMPDIR/sglang-corpus.json  # -> valid

See validation-report.json for the executed checks and checksums.