Repository navigation
fix(moe): keep precision-overridden grouped experts off the op-fuser path - #7212
Merged
fanshiqing merged 1 commit intoSep 10, 2026
Conversation
…path - Bug: the op-fuser's grouped-MLP kernels are FP8/NVFP4-only and take their recipe from the global autocast state, so a --te-precision-config-file override never reached them. The layer quantized anyway: silently under plain TE, fatally under GTP, whose backward hands the kernel an unquantized weight (AttributeError on '_columnwise_data'). - Fix: gate _with_fused_impl on both linears' will_execute_quantized(). Fusion spans fc1 and fc2, so either one opting out disables it. Signed-off-by: Shiqing Fan <shiqingf@nvidia.com>
svcnvidia-nemo-ci
marked this pull request as draft
September 10, 2026 15:14
Contributor
|
This PR has been automatically converted to draft because all PRs must start as drafts. When you are ready for review, click Ready for Review to begin the review process. This will:
See the contribution guide for more details. |
fanshiqing
marked this pull request as ready for review
September 10, 2026 15:14
Member
Author
|
/ok to test 9f0564b |
Member
Author
|
/claude review |
deepakn94
approved these changes
Sep 10, 2026
This comment was marked as low quality.
This comment was marked as low quality.
kvareddy
approved these changes
Sep 10, 2026
Mellonta
approved these changes
Sep 10, 2026
fanshiqing
enabled auto-merge
September 10, 2026 16:42
Phlip79
approved these changes
Sep 10, 2026
|
🔄 Merge queue validation started! You can track the progress here: https://github.com/NVIDIA/Megatron-LM/actions/runs/34522758319 |
fanshiqing
deleted the
shiqingf/fix_gmm_fusion_with_precision_override
branch
September 10, 2026 21:09
fanshiqing
changed the title
Closed
Closed
4 of 7 tasks
Closed
xuwchen
pushed a commit
to xuwchen/Megatron-LM
that referenced
this pull request
Oct 14, 2026
NVIDIA#7212) Signed-off-by: Shiqing Fan <shiqingf@nvidia.com> (cherry picked from commit 8962453) Signed-off-by: xuwenc <xuwenc@nvidia.com>
xuwchen
pushed a commit
to xuwchen/Megatron-LM
that referenced
this pull request
Oct 14, 2026
NVIDIA#7212) Signed-off-by: Shiqing Fan <shiqingf@nvidia.com> (cherry picked from commit 8962453) Signed-off-by: xuwenc <xuwenc@nvidia.com>
This branch was previously deployed
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.
What does this PR do?
fix(moe): keep precision-overridden grouped experts off the op-fuser path
FP8/NVFP4-onlyand take their recipefrom the global autocast state, so a
--te-precision-config-fileoverride neverreached them.
GTP, whose backward hands the kernel an unquantized weight (AttributeError on
'_columnwise_data').
_with_fused_implon both linears' will_execute_quantized(). Fusion spansfc1 and fc2, so either one opting out disables it.
Before the fix: casting+mxfp8-gmm-fusion
After the fix: fall back to unfused gmm in bf16
Contribution process
Pre-checks