Update to pytorch 2.6 #8944
Open
Update to pytorch 2.6 #8944
Conversation
This file contains 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
tengyifei
approved these changes
Apr 5, 2025
| [tool.pytest.ini_options] | ||
| addopts="-n auto" | ||
| #[tool.pytest.ini_options] | ||
| #addopts="-n auto" |
There was a problem hiding this comment.
Is this intended change?
| @@ -129,6 +129,7 @@ def run_export_and_compare(testcase, | |||
| equal_nan=True, | |||
| ignore_indices=False): | |||
| atol, rtol = (1e-3, 1e-5) | |||
| #breakpoint() | |||
There was a problem hiding this comment.
Is this intended change?
|
I wonder if it's possible to eliminate all autograd-related dispatch keys when calling these torch ops, since we just need decompositions in the forward direction and can rely on Jax autograd, I think.. It maybe runs into other PyTorch problems |
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Outdated suggestions cannot be applied.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.
Pytorch 2.6 introduced usage of decompositions via CompositeImplicitAutograd key: https://github.com/pytorch/pytorch/blob/60a45eb862d5e8b4ba2dd435d34ef04ae231e885/torch/_export/utils.py#L1249
These decompositions currently does not play nice with
torch_dispatchbased decomposition, resulting infinite recursion.Likely we are using it wrong, I made a post here soliciting the right way to use decomposition https://dev-discuss.pytorch.org/t/what-is-the-right-way-to-use-decompositions-in-dispatch-mode/2888 soliciting advice on the topic.
Meanwhile the workaround is to explicitly list out decompositions we like without the CompositeImplicitAutograd decompositions. The explicit list is produced via keys in
core_aten_decompositions()call usingtorch 2.5.1.