{{ message }}
Repository navigation
[ROCm][Bugfix] Keep zero MiniMax MXFP8 activation blocks finite - #58089
Merged
shen-shanshan merged 2 commits intoSep 23, 2026
Merged
Conversation
Use finite reciprocal scaling to prevent all-zero MXFP8 activation blocks from generating NaNs on CDNA. The existing regression fails without the fix; all 68 AMD MiniMax kernel tests pass with it. Co-authored-by: OpenAI Codex Signed-off-by: Andreas Karatzas <akaratza@amd.com>
AndreasKaratzas
marked this pull request as ready for review
September 22, 2026 17:34
AndreasKaratzas
requested review from
dllehr-amd,
hongxiayang and
tjtanaa
as code owners
September 22, 2026 17:34
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
Member
Author
|
✅ Triggered Buildkite CI #90463 for commit |
Fangzhou-Ai
approved these changes
Sep 22, 2026
Collaborator
|
LGTM! |
Comment on lines
+118
to
+119
| rescale = tl.exp2(127.0 - sb) | ||
| aq = (act * rescale[:, None]).to(aq_ptr.dtype.element_ty) |
Contributor
There was a problem hiding this comment.
If this doesn't influence the model accuracy, then LGTM.
Member
Author
There was a problem hiding this comment.
mm .. let me then do a quick lm eval
Member
Author
There was a problem hiding this comment.
How does this look?
Contributor
There was a problem hiding this comment.
Run Source commit Correct Accuracy
base74370a0e30441256/1319 95.2237%
pra10595a429791252/1319 94.9204%
base-repeat74370a0e30441256/1319 95.2237%
pr-repeata10595a429791252/1319 94.9204%
How does this look?
LGTM.
shen-shanshan
approved these changes
Sep 23, 2026
michael-gpu
pushed a commit
to practical-inference/vllm-ampere-extended
that referenced
this pull request
Oct 20, 2026
…-project#58089) Signed-off-by: Andreas Karatzas <akaratza@amd.com>
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.
0/0when the byte-zero E8M0 divisor flushes to zero on CDNA, while preserving the encoded scale bytes.The MI355 parity run exposed NaNs in an all-zero activation block. Merged #53674 corrected the fused kernel’s scale formula but retained division by the decoded scale. The generic quantizer already uses reciprocal multiplication from merged #53110, and this change applies the same arithmetic to the fused path. Fresh searches and the related PR review found no open change covering this remaining fused-kernel defect.
AI assistance: OpenAI Codex assisted with diagnosis, implementation and validation.