Fix torch.export compatibility for Mixtral MoE models #40114
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
- Replace data-dependent .nonzero() operation with static expert loop - Resolves GuardOnDataDependentSymNode error during torch.export - Maintains identical functionality while enabling export compatibility - Fixes issue introduced in PR huggingface#32429 - Add tests for torch.export compatibility
akacmazz
changed the title
- Auto-generate modeling_mixtral.py with the same fix - Apply black formatting - Fix repository consistency check
akacmazz
force-pushed
the
fix-mixtral-torch-export-compatibility
branch
3 times, most recently
from
9b41625 to
c3e3c5e
Compare
Thx for the feedback, i will work on this |
…ility - Training path: Keep efficient .nonzero() for performance - Inference path: Use static loop for torch.export compatibility - Add conditional check to skip empty experts in inference - Update tests to validate inference mode export - Addresses maintainer feedback on performance concerns
akacmazz
force-pushed
the
fix-mixtral-torch-export-compatibility
branch
from
952a181 to
0aa9de7
Compare
- Apply black formatting to fix code style - Fix import sorting with isort - Address CI code quality checks
- Fix import organization in modeling_mixtral.py - Fix import organization in modular_mixtral.py - Address ruff I001 import sorting warnings
- Remove manually edited modeling_mixtral.py - Auto-generate from modular_mixtral.py using proper tool - Ensure consistency between modular and generated files - Fix check_repository_consistency CI failure
- Remove 'if top_x.shape[0] == 0: continue' check that causes GuardOnDataDependentSymNode error - Empty expert tensors naturally contribute 0, no explicit check needed - Update test error message for clarity - Fixes tests_processors CI failure Co-authored-by: ArthurZucker <[email protected]>
|
[For maintainers] Suggested jobs to run (before merge) run-slow: mixtral |
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?
This PR fixes a torch.export compatibility issue #38518 with Mixtral MoE models that was introduced in PR #32429.
Problem
The optimization in PR #32429 introduced a .nonzero() operation that creates data-dependent tensor shapes, causing torch.export to fail with:
torch.fx.experimental.symbolic_shapes.GuardOnDataDependentSymNode: Could not extract specialized integer from data-dependent expression
Solution
Replace the dynamic expert selection loop:
expert_hit = torch.greater(expert_mask.sum(dim=(-1, -2)), 0).nonzero()
for expert_idx in expert_hit:
With a static loop over all experts:
for expert_idx in range(self.num_experts):
Impact
Testing
Fixes torch.export compatibility issues reported for Mixtral-8x7B models.
Before submitting
Pull Request section?
to it if that's the case.
documentation guidelines, and
here are tips on formatting docstrings.
Who can review?
Anyone in the community is free to review the PR once the tests have passed. Feel free to tag members/contributors who may be interested in your PR.
@Cyrilvallez
@ArthurZucker
@gante