PyTorch autograd updates #2808
Open
PyTorch autograd updates #2808
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
…d nightly PyTorch by adjusting args_tensor_mask handling. Update tests to utilize a helper function for kwargs creation, ensuring seamless integration across versions.
for more information, see https://pre-commit.ci
for more information, see https://pre-commit.ci
…to handle ctx argument based on args_tensor_mask presence. Update tests to accommodate changes in forward and backward module definitions for both stable and nightly PyTorch versions.
for more information, see https://pre-commit.ci
Collaborator
Author
|
I see another type of failure on PT nightly: which seems to be unrelated to the |
for more information, see https://pre-commit.ci
for more information, see https://pre-commit.ci
mattteochen
requested review from
KaelanDt,
lantiga and
mruberry
as code owners
mattteochen
commented
Dec 18, 2025
|
|
||
| # Detect once at module load time whether PyTorch uses args_tensor_mask. | ||
| # This must be done outside the JIT-traced function to avoid interpreter issues. | ||
| def _detect_has_args_tensor_mask(): |
Collaborator
Author
There was a problem hiding this comment.
If you have better ideas...
crcrpar
approved these changes
Dec 19, 2025
IvanYashchuk
approved these changes
Dec 19, 2025
Contributor
There was a problem hiding this comment.
Pull request overview
This PR enables torch autograd support for both stable and nightly PyTorch versions by handling the API change where nightly PyTorch removed the args_tensor_mask parameter and the ctx argument requirement from autograd function signatures.
Key Changes:
- Made
args_tensor_maskparameter optional in autograd functions and added conditional logic to handle both stable PyTorch (with ctx) and nightly PyTorch (without ctx) - Implemented runtime detection mechanism to determine whether PyTorch version uses
args_tensor_mask - Removed
xfail_if_args_tensor_mask_removeddecorators from tests that were previously expected to fail
Reviewed changes
Copilot reviewed 5 out of 5 changed files in this pull request and generated 2 comments.
Show a summary per file
| File | Description |
|---|---|
| thunder/torch/init.py | Made args_tensor_mask optional and added conditional logic in autograd_function_apply, augmented_forward_autograd_function_apply, and backward_autograd_function_apply to handle forward/backward calls with or without ctx argument |
| thunder/core/jit_ext.py | Updated lookaside function to conditionally handle args_tensor_mask presence, adjusted argument construction for forward/backward calls, and modified gradient input handling based on PyTorch version |
| thunder/tests/test_jit_general.py | Added _detect_has_args_tensor_mask() detection function using source inspection, created helper function _autograd_function_apply_kwargs() for version-agnostic kwargs, and adapted test cases to define forward/backward functions conditionally based on PyTorch version |
| thunder/tests/test_update_aliases.py | Removed xfail_if_args_tensor_mask_removed import and decorator from test |
| thunder/tests/test_dynamo.py | Removed xfail_if_args_tensor_mask_removed decorator from test |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
Fixed comment Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
Collaborator
Author
|
@KaelanDt for review. Thanks. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
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.
This PR enables torch autograd support for both stable and nightly PyTorch versions. The target is to remove all the expected failures introduced by #2805.
See #2803.