Repository navigation

perf(0.69.4): batch: qwen35 persistent embed buffer (#4316) + hot-path allocs + stale q8 (#4258) - #4322

Closed

noahgift wants to merge 11 commits into

mainfrom

perf/0694-qwen35-persistent-embed

Closed

perf(0.69.4): batch: qwen35 persistent embed buffer (#4316) + hot-path allocs + stale q8 (#4258)#4322
noahgift wants to merge 11 commits into
mainfrom
perf/0694-qwen35-persistent-embed

Commits on Sep 24, 2026
- perf(qwen35-cuda): one resident residual buffer, filled in stream order — no per-token GpuBuffer::from_host (#4316)
  
  noahgift
  and
  claude
  committed
  02729ec
- PMAT-4316: implementation receipt (GPU run pending)
  
  noahgift
  and
  claude
  committed
  635bd67
- PMAT-4316: roadmap kind:code label
  
  noahgift
  and
  claude
  committed
  bba6b09
- PMAT-4316: quorum receipt — AGREED 3/3 PASS on bba6b0924
  
  noahgift
  and
  claude
  committed
  be6acf6
- PMAT-4316: GPU evidence — qwen35_cuda 17/17 on RTX 4090
  
  noahgift
  and
  claude
  committed
  86498a9
- fix(cuda): a DP4A GEMV requantizes when its input buffer or contents change (Closes #4258)
  
  noahgift
  and
  claude
  committed
  263958f
- test(cuda): pre-allocate every input so no staleness case rides a reused address; document the (ptr,len) key's limit (Refs #4258)
  
  noahgift
  and
  claude
  committed
  56c9f91
- docs(roadmap): register PMAT-4258 (Refs #4258)
  
  noahgift
  and
  claude
  committed
  4d4a784
- test(qwen35-cuda): the DP4A falsifier becomes a parity guard — its 'catastrophic' reading was #4258
  
  noahgift
  and
  claude
  committed
  555a677
- PMAT-4258: impl receipt — carried from #4302, GPU green on 555a6774d
  
  noahgift
  and
  claude
  committed
  36f1ea7
- PMAT-4258: quorum AGREED 3/3 PASS on 36f1ea707 + stale falsifier name in parity thresholds (comment)
  
  noahgift
  and
  claude
  committed
  3b23f45