Skip to the content.

Dedicated cache-warming alternatives — 2026-08-10

Verdict

INCOMPLETE. The local packet executes fak’s dedicated-warm planner/readback accounting and a demand-only baseline. Anthropic, Gemini, OpenAI, LMCache, Mooncake, vLLM, and SGLang retain zero measurements until their real APIs and cache stores run the common fanout; #6162 tracks those witnesses.

Same-workload contract

Every arm receives five cases over the same 1,024-token/4,096-byte fingerprint: profitable Anthropic explicit warm, fingerprint mismatch, below-break-even fanout, OpenAI automatic-cache ordering, and unsupported active capability. Fak also reconciles one closed two-call fanout with 2,048 observed cache-read tokens. The oracle requires one dedicated confirmed warm, correct refusal reasons, and no fabricated provider result.

Complete provider/engine runs replay the same continuation fanout and report decision quality, useful and wasted warms, TTFT and planning latency, cache-write/read/input tokens, bytes, CPU/RSS/network/storage, throughput, and total cost.

Arm Class Local availability Honest result
fak native dedicated warm planner and accounting native yes exact five-case oracle; one confirmed warm
demand-only fills without dedicated warming tuned baseline yes no warm spend and no pre-fanout cache entry
fak + Anthropic prompt caching first-class integration no real API usage/readback required
fak + Gemini CachedContent first-class integration no real explicit cache lifecycle required
fak + OpenAI automatic prefix caching first-class integration no real API usage/readback required
fak + LMCache first-class integration no real engine required
fak + Mooncake first-class integration no real engine required
vLLM automatic prefix caching external no zero measurements
SGLang HiCache external no zero measurements

Adapters, fake provider responses, and local cache-shaped stores are not provider/engine witnesses.

Local native witness

go test ./internal/vcachewarm -bench BenchmarkPlanAndReconcileWarm -benchmem -run '^$' -count=5

Windows/amd64, AMD Ryzen 9 9950X: 96.47, 95.07, 100.4, 106.6, 107.9 ns/op. Median: 100.4 ns/five-case planning plus reconciliation, 0 B/op, 0 allocs/op. This is pure planner overhead, not provider TTFT or cache-fill latency.

Reproduce

go test ./internal/vcachewarm -run TestCompareLocalKeepsWarmingAlternativesExplicit
go test ./internal/vcachewarm -bench BenchmarkPlanAndReconcileWarm -benchmem -run '^$' -count=5
go test ./internal/nativebench