[ROCm][CI] Fall back to lossless Kimi K3 MXFP4 emulation on gfx942 - #50516
Merged
Conversation
Signed-off-by: Andreas Karatzas <Andreas.Karatzas@amd.com>
AndreasKaratzas
marked this pull request as ready for review
July 31, 2026 03:57
AndreasKaratzas
requested review from
mgoin,
pavanimajety,
robertgshaw2-redhat,
tlrmchlsmth,
yewentao256 and
zyongye
as code owners
July 31, 2026 03:57
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
AndreasKaratzas
added
the
ready
mgoin
approved these changes
Jul 31, 2026
This was referenced Aug 3, 2026
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
2 participants
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
git bisectidentifiedaeeb36b1f171, vLLM #50000, as the first revision with both Kimi K3 failures.Motivation:
PR #50000 defaults Kimi K3 to an MXFP4 RoutedExperts configuration. The model-specific native path handles gfx950, while gfx942 reaches the generic MXFP4 MoE oracle. That oracle offered only
AITER_MXFP4_BF16on ROCm. AITER rejects this SiTU/gfx942 combination, leaving no selectable backend and raisingNotImplementedErrorduring model construction. The tensor-schema test fails for the same reason before reaching schema processing. The emulation leaf also treated the generic configuration as W4A4. Kimi's checkpoint is weight-only W4A16, so the fallback must keep activations in BF16 and consume the checkpoint's packed MXFP4 layout without conversion.