fix(utils): call empty_cache() after fp16→fp32 casts in prepare_model_for_kbit_training#3293
Open
umutonuryasar wants to merge 1 commit into
Open
Conversation
…_for_kbit_training The bulk param.data = param.data.to(torch.float32) loop creates temporary tensors that PyTorch's CUDA allocator keeps cached even after they are no longer referenced, resulting in ~1 GB of reserved-but-unused CUDA memory on return. This breaks training on 8 GB unified-memory devices. Fix: add a single torch.cuda.empty_cache() call (guarded by torch.cuda.is_available()) after the cast loop so the allocator releases those blocks back to the driver immediately. Fixes huggingface#3265
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
The bulk param.data = param.data.to(torch.float32) loop creates temporary
tensors that PyTorch's CUDA allocator keeps cached even after they are no
longer referenced, resulting in ~1 GB of reserved-but-unused CUDA memory
on return. This breaks training on 8 GB unified-memory devices.
Fix: add a single torch.cuda.empty_cache() call (guarded by
torch.cuda.is_available()) after the cast loop so the allocator releases
those blocks back to the driver immediately.
Fixes #3265