fix: disable xformers for custom attention bias in GRPO training #489
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
xformers with GQA (Grouped Query Attention) switches to 5D tensor format during gradient checkpointing when requires_grad=False. The cutlass backend doesn't support custom attention bias with 5D tensors, causing: "RuntimeError: Bias expected in BMHK format" Solution: Temporarily disable xformers during training to force the SDPA path, which always uses 4D tensors and properly supports custom attention bias for trajectory group/parent masking. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <[email protected]>
Add type: ignore comments and explicit int() casts for model config attributes to pass pyright type checking in CI. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <[email protected]>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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Outdated suggestions cannot be applied.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.
Summary
RuntimeError: Bias expected in BMHK formatwhen using custom attention bias with GQA models (e.g., Mistral-7B) during GRPO trainingrequires_grad=False), but the cutlass backend doesn't support custom bias with 5D tensorsTest plan
loss=-0, grad_norm=43.1, policy_loss=-3.41e-8, entropy=6.31🤖 Generated with Claude Code