Repository navigation
[determinism] Preserve IEEE FP32 dot inputs - #117
Merged
Merged
Conversation
Use explicit IEEE input precision in batch-invariant matmul kernels and cover FP32 precision and exact batch/permutation/CUDA-graph behavior.
This was referenced Nov 6, 2026
yonromai
approved these changes
Nov 6, 2026
Member
There was a problem hiding this comment.
Non-blocking: do we have before/after timings for representative FP32 router projections, or serving throughput?
Member
Author
|
🤖 No before/after router timings or serving-throughput measurements yet. The recorded GPU results cover numerical correctness on the earlier 02872a7 base with the same corrected module bytes; they do not measure the performance cost on this PR’s full base. I’ll keep that limit explicit rather than infer performance from the correctness tests. |
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Use IEEE FP32 input precision in the persistent matmul, descriptor matmul and batched matmul kernels used by batch-invariant arithmetic. Triton's default dot precision can round FP32 inputs even when the caller expects the full FP32 mantissa. Add one-hot precision regressions with an independent exact answer, plus exact singleton, batched, permuted and CUDA-graph replay comparisons.
The corrected kernel file is byte-identical to the candidate exercised in the saved H100 qualification: two original-precision controls failed as expected, then all 47 operator/attention cases passed without skips. Four later full-length Grug DP8/EP8 evaluations independently graded all 100 tasks each and matched exactly under the tested deterministic compiler policy. Those measurements used vLLM
02872a7plus this kernel change, not the proposedf6ae232ff3747569233ce9fce99ec4e05dbec6b6base, which includes newer FP8 support. On that base, required lint, focused Ruff checks, syntax and exact source/AST checks pass; GPU regressions have not been rerun. The historical GPU results neither establish universal determinism nor isolate the cause of prior full-model variation.This is a source fix. Wheel publication and consumer adoption retain their separate release and application gates. The change lets the reproducibility note in marin-community/marin#10760 reference maintained source instead of a patch attachment.