Repository navigation
perf(serve): decode GEMV precision on Qwen3.5 (RCA-SRV-001 fix 3) #4485
Description
Residual decode is 1.37-1.82x llama after #4273 (GEMV + recurrence). Apply the same precision-policy split as fix 1 to decode GEMV, keeping the test-isolation tolerance out of production.
Source: RCA-SRV-001 (docs/findings/rca-srv-001-five-whys.md, branch rca/srv-001 @f2b57c8bd, aprender-76, measured on RTX 4090). Epic: 0.71 E2 Verbs Are Fast (#3598). Acceptance: a measured delta on the PRM-S1 v2 set against llama.cpp, with the pinned binary (version+sha) and an e2e quality falsifier (greedy text identical, or a stated tolerance).
Activity
test(#4485): the DP4A graph replays exactly; the graph step's ~5e-6 r…
feat(#4485): log the decode GEMV pair a Qwen3.5 CUDA model armed, and…
docs(#4485): PRM-S1 v2 receipt — DP4A holds verdict mix, +4% p75 over…
noahgift commentedon Oct 12, 2026
park proposal: parked, decode GEMV precision work (RCA-SRV-001 fix 3) has no commit on main and no activity for 8 days; valid, but out of the current train's budget.
noahgift commentedon Oct 14, 2026
merged into #5305 (master ticket for epic #3598; triage rule 7b, 2026-10-14). It is a checklist row there; reopen this issue when work on it starts or its PR opens.
noahgift commentedon Oct 14, 2026
Moved: this sub-ticket is now a row in #5320 (Qwen3.5 CUDA decode speed: close the GEMV and DeltaNet gaps to llama.cpp), split out of #5305 by theme.