perf(qwen35-cuda): DP4A requant staleness (#4258) + argmax first-of-ties (#4259) + allocation-free decode plumbing (#4215) — replaces #4322 #4332 by noahgift · Pull Request #4439 · paiml/aprender

fix(cuda): a DP4A GEMV requantizes when its input buffer or contents change (Closes #4258)

noahgift

and

claude

committed

a3019f2

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

ac4dac3

docs(roadmap): register PMAT-4258 (Refs #4258)

noahgift

and

claude

committed

3c6e02e

test(qwen35-cuda): the DP4A falsifier becomes a parity guard — its 'catastrophic' reading was #4258

noahgift

and

claude

committed

ed67017

PMAT-4258: impl receipt — carried from #4302, GPU green on 555a6774d

noahgift

and

claude

committed

73b2902

PMAT-4258: quorum AGREED 3/3 PASS on 36f1ea707 + stale falsifier name in parity thresholds (comment)

noahgift

and

claude

committed

876c884

GPU argmax returned the LAST of two exactly-equal maxima; the CPU argmax it replaces returns the FIRST (Closes #4259)

noahgift

and

claude

committed

fdf3d1b

[WIP] perf(qwen35): allocation-free greedy decode token (Refs #4215)

noahgift

and

claude

committed

c3c8aaa

test(qwen35): decode tok/s at 850 bench, ignored (Refs #4215)

noahgift

and

claude

committed

f8cecc3

perf(qwen35): a greedy decode token allocates nothing on host or device and downloads 4 bytes (Refs #4215)

noahgift

and

claude

committed

a2e46a4

style: rustfmt the #4215 driver re-exports after the cherry-pick onto main

noahgift

and

claude

committed

1562b73

fix(cuda): CudaExecutor::new inits the #4215/#4258 fields with Default — the complexity ratchet read cognitive 40 -> 41

noahgift

and

claude

committed

cafbd6c

test(qwen35-cuda): the #4215 zero-host-alloc token measures the release lookup, not main's #3759 test-build key proof

noahgift

and

claude

committed

f1448ff

ci: re-run §6 R-2 against the keep-open body (#3759 #4215 are Refs, reasons stated)

noahgift

and

claude

committed

b5c1041

fix(guard): keep par-062's baselined claim line at 257 — the #4215 comment had shifted it by 2

noahgift

and

claude

committed

cb6c2ff

Merge branch 'main' into perf/qwen35-dp4a-argmax-decode-replaces-4322-4332

noahgift

authored

0f3b6a2

fix(guard-tree test): row 55 read an overlap as 'alone' when the serial guard ran inside the pool guards

noahgift

and

claude

committed

c3546d3

Navigation Menu

Repository navigation

perf(qwen35-cuda): DP4A requant staleness (#4258) + argmax first-of-ties (#4259) + allocation-free decode plumbing (#4215) — replaces #4322 #4332 - #4439

perf(qwen35-cuda): DP4A requant staleness (#4258) + argmax first-of-ties (#4259) + allocation-free decode plumbing (#4215) — replaces #4322 #4332#4439
noahgift merged 17 commits into
mainfrom
perf/qwen35-dp4a-argmax-decode-replaces-4322-4332

Commits on Sep 25, 2026

fix(cuda): a DP4A GEMV requantizes when its input buffer or contents change (Closes #4258)

test(cuda): pre-allocate every input so no staleness case rides a reused address; document the (ptr,len) key's limit (Refs #4258)

docs(roadmap): register PMAT-4258 (Refs #4258)

test(qwen35-cuda): the DP4A falsifier becomes a parity guard — its 'catastrophic' reading was #4258

PMAT-4258: impl receipt — carried from #4302, GPU green on 555a6774d

PMAT-4258: quorum AGREED 3/3 PASS on 36f1ea707 + stale falsifier name in parity thresholds (comment)

GPU argmax returned the LAST of two exactly-equal maxima; the CPU argmax it replaces returns the FIRST (Closes #4259)

[WIP] perf(qwen35): allocation-free greedy decode token (Refs #4215)

test(qwen35): decode tok/s at 850 bench, ignored (Refs #4215)

perf(qwen35): a greedy decode token allocates nothing on host or device and downloads 4 bytes (Refs #4215)

style: rustfmt the #4215 driver re-exports after the cherry-pick onto main

fix(cuda): CudaExecutor::new inits the #4215/#4258 fields with Default — the complexity ratchet read cognitive 40 -> 41

test(qwen35-cuda): the #4215 zero-host-alloc token measures the release lookup, not main's #3759 test-build key proof

ci: re-run §6 R-2 against the keep-open body (#3759 #4215 are Refs, reasons stated)

fix(guard): keep par-062's baselined claim line at 257 — the #4215 comment had shifted it by 2

Merge branch 'main' into perf/qwen35-dp4a-argmax-decode-replaces-4322-4332

fix(guard-tree test): row 55 read an overlap as 'alone' when the serial guard ran inside the pool guards