Fix BMM style MoE export in fp8_pc_pt recipe #515
Conversation
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
Codecov Report✅ All modified and coverable lines are covered by tests. Additional details and impacted files@@ Coverage Diff @@
## main #515 +/- ##
==========================================
+ Coverage 73.44% 73.46% +0.02%
==========================================
Files 180 180
Lines 18147 18161 +14
==========================================
+ Hits 13328 13342 +14
Misses 4819 4819 ☔ View full report in Codecov by Sentry. 🚀 New features to boost your workflow:
|
Signed-off-by: Zhiyu Cheng <[email protected]>
Edwardf0t1
force-pushed
the
zhiyu/fix-bmm-moe-fp8-pc-pt-export
branch
from
4bb85d1 to
fbd5417
Compare
Signed-off-by: Zhiyu Cheng <[email protected]>
cjluo-nv
reviewed
Nov 12, 2025
| if weights_scaling_factor.shape[-1] == weight.shape[-1]: | ||
| # (num_experts, input_dim) -> (num_experts, 1, input_dim), BMM-style | ||
| return (weight / weights_scaling_factor.unsqueeze(-2)).to(torch.float8_e4m3fn) | ||
| elif weights_scaling_factor.shape[-1] == weight.shape[-2]: |
Collaborator
There was a problem hiding this comment.
do we also need to check weights_scaling_factor.shape[0] == weight.shape[0]
Collaborator
|
Could you share what models have you tested? |
meenchen
reviewed
Nov 12, 2025
| # Per-expert scaling only: (num_experts,) -> (num_experts, 1, 1) | ||
| return (weight / weights_scaling_factor[:, None, None]).to(torch.float8_e4m3fn) | ||
| elif weights_scaling_factor.dim() == 2: | ||
| # Per-channel scaling: check which dimension matches |
Contributor
There was a problem hiding this comment.
Can we add a check here to make sure weight.shape[-1] != weight.shape[-2]?
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?
Type of change: Bug fix
Overview: The Llama-4-Scout-17B-16E-Instruct model uses Llama4TextExperts, which stores expert weights in a BMM (batch matrix multiply) layout: (num_experts, input_dim, output_dim). This is different from standard MoE models. The FP8_PC_PT (FP8 per-channel per-token) quantization code didn't handle this layout properly, causing shape mismatches.
Usage
Testing
Before your PR is "Ready for review"
Additional Information