Repository navigation
fix(utils): restore torch cross-entropy state in revert helpers (#1587) - #1611
Open
YuvrajSinghBhadoria2 wants to merge 2 commits into
Open
YuvrajSinghBhadoria2 wants to merge 2 commits into
YuvrajSinghBhadoria2 wants to merge 2 commits into
Conversation
) Skip the torch.cuda.device()/torch.xpu.device() context-manager guard when torch.is_compiling() is true. Dynamo emits incorrect device-restore nodes during tracing, which leads to wrong output or illegal memory access (linkedin#1590). Use getattr(torch, 'is_compiling', lambda: False)() so the check is robust on torch versions that do not define torch.is_compiling. Add CPU-runnable unit tests covering the compile-skip path, the non-compile path, CPU device, and backends without a .device attribute.
…edin#1587) In transformers >= 4.46.1, CrossEntropyLoss is patched into module-level references and torch.nn.functional.cross_entropy. The revert helper functions previously only reloaded modeling modules, leaving F.cross_entropy and nn.CrossEntropyLoss pointing to LigerCrossEntropyLoss. Consequently, multi-model convergence tests compared two Liger-patched models against each other after the first test case. - Add _ORIGINAL_TORCH_CROSS_ENTROPY capturing the original F.cross_entropy before any Liger patching - Add revert_liger_cross_entropy_patches() to restore F.cross_entropy, reload torch.nn, and remove any lingering CrossEntropyLoss attribute on modeling modules bound to LigerCrossEntropyLoss - Call revert_liger_cross_entropy_patches() from all revert_liger_kernel_to_* helpers in test/utils.py - Add test_revert_liger_cross_entropy_patches unit test Fixes linkedin#1587
This branch has not been deployed
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
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 #1587
In transformers >= 4.46.1, CrossEntropyLoss is patched into module-level
references and torch.nn.functional.cross_entropy. The revert helper
functions previously only reloaded modeling modules, leaving
F.cross_entropy and nn.CrossEntropyLoss pointing to LigerCrossEntropyLoss.
Consequently, multi-model convergence tests compared two Liger-patched
models against each other after the first test case.
before any Liger patching
reload torch.nn, and remove any lingering CrossEntropyLoss attribute
on modeling modules bound to LigerCrossEntropyLoss
helpers in test/utils.py
All tests verified passing in a CPU-only Docker container (no GPU required).