Add @torch.no_grad() to cache layer update methods #43041
+2
−0
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
Decorates all cache layer update() methods with @torch.no_grad() to prevent PyTorch autograd from complaining about tensor version changes when computing gradients with respect to model inputs. This follows the same pattern used by optimizer.step() methods and is safe because cache updates are only used during inference/generation, never during training. Methods decorated: - DynamicLayer.update() - DynamicSlidingWindowLayer.update() - StaticLayer.update() - StaticSlidingWindowLayer.update() - QuantizedLayer.update() Fixes huggingface#43010
Remove @torch.no_grad() from DynamicLayer, DynamicSlidingWindowLayer, and QuantizedLayer since they use torch.cat() (not in-place) and need gradient flow preserved. Keep @torch.no_grad() only on StaticLayer and StaticSlidingWindowLayer which use index_copy_() and copy_() (in-place operations) that cause torch.func.grad to fail.
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Outdated suggestions cannot be applied.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.
Fixes #43010
What does this PR do?
Adds
@torch.no_grad()decorator to cache layerupdate()methods that use in-place operations, preventingtorch.func.gradfrom failing with "in-place operation would mutate a captured Tensor" errors.Why only StaticLayer and StaticSlidingWindowLayer?
After investigation, I found that only these two classes need the decorator:
@torch.no_grad()?DynamicLayertorch.cat()DynamicSlidingWindowLayertorch.cat()StaticLayerindex_copy_()StaticSlidingWindowLayercopy_(),index_copy_()QuantizedLayertorch.cat()Adding the decorator to
DynamicLayer(and subclasses) would break gradient flow becausetorch.cat()creates new tensors that participate in the computation graph. Models like T5 useDynamicCacheand need gradients to flow through cached key/values.Changes
Added
@torch.no_grad()decorator to:StaticLayer.update()StaticSlidingWindowLayer.update()Testing
torch.func.gradworks with StaticCache