[Kernel][SM70] Integrate concurrent QPN and verifier paths (#476) - #554
Merged
yangzhuxinyzx merged 12 commits intoSep 7, 2026
Merged
Conversation
Signed-off-by: yangzhuxinyzx <153831768+yangzhuxinyzx@users.noreply.github.com>
Signed-off-by: yangzhuxinyzx <153831768+yangzhuxinyzx@users.noreply.github.com>
Signed-off-by: yangzhuxinyzx <153831768+yangzhuxinyzx@users.noreply.github.com>
Signed-off-by: yangzhuxinyzx <153831768+yangzhuxinyzx@users.noreply.github.com>
Signed-off-by: yangzhuxinyzx <153831768+yangzhuxinyzx@users.noreply.github.com>
Signed-off-by: yangzhuxinyzx <153831768+yangzhuxinyzx@users.noreply.github.com>
Signed-off-by: yangzhuxinyzx <153831768+yangzhuxinyzx@users.noreply.github.com>
…p4-dflash2-concurrency-20260903-131310 Signed-off-by: yangzhuxinyzx <153831768+yangzhuxinyzx@users.noreply.github.com> # Conflicts: # docs/design/sm70_v100_migration_control.md
Signed-off-by: yangzhuxinyzx <153831768+yangzhuxinyzx@users.noreply.github.com>
This reverts commit 8df5cc7.
…vements Merge #476 preserving latest-main collectives and E4M3 paths; keep HC storage disjoint; default measured QPN M16/M32 paths on with compatible split-K fallback. Optional batched verifier and enlarged collective remain opt-in. Co-authored-by: OpenAI Codex <noreply@openai.com> Signed-off-by: yangzhuxinyzx <153831768+yangzhuxinyzx@users.noreply.github.com>
Retain distinct request-major and FP32 workspaces, E4M3 row-length semantics, and both native capability entries. Co-authored-by: OpenAI Codex <noreply@openai.com> Signed-off-by: yangzhuxinyzx <153831768+yangzhuxinyzx@users.noreply.github.com>
Contributor
Author
|
Final composition check: the reviewed #554 head combined with #551 produces tree ff08c8b912526dfd7eed88b35b034b7526638043. That exact local composite passed 176 additional targeted CPU/GPU tests, including new Bailing state/graph tests, fresh QPN native oracle tests, shared FP8/MXFP4 configuration, default environment policies and collective dispatch. Combined fresh Flash-V100 previously passed 323 tests. No E2E suite was run. Merge remains gated on this published head’s hosted checks. |
yangzhuxinyzx
marked this pull request as ready for review
September 7, 2026 10:55
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
AI-assisted maintainer-requested integration of #476 onto latest main 378a93a, preserving #548/#549/#550 and original PR ancestry.
Default-on: measured channel-FP8 QPN8 M16/native M32 dense plus M16-chunked gated projections, and NVFP4 QPN2 M16. Unsupported split-K falls back instead of throwing. Existing large-row/block-FP8 routes retained. Batched grouped verification and larger push-allreduce remain opt-in where evidence is workload-dependent.
Merge repairs preserve role-aware defaults, HC-disjoint collective storage, and E4M3 explicit row lengths. Request-major and FP32 partial workspaces are independently keyed.
Tests: 247 initial native/attention/rejection checks; 266 policy checks passed with 13 GPU skips in the CPU run; fresh combined Flash-V100 after #550 passed 323 tests. Fresh native TP4 push/sum2: 7 sizes x 7 graph replays with HC canaries. QPN tests use fresh sidecar and independent FP64 oracle. Actual mixed-checkpoint graph timings: FP8 down M16 185.290→43.277 us, M32 189.456→63.818 us; NVFP4 gate/up M16 67.222→65.363 us, down 34.979→30.573 us, NVFP4 bitwise unchanged. FP8 max observed relative L2 versus prior path 0.06636%. No new model-E2E throughput claim; original endpoint/text-health evidence retained separately. See docs/design/sm70_concurrency_integration_audit.md. Draft pending hosted checks.