Add Dynamic Output Shape Tag For data-dependent ops, handle in FakeTensor#79170
Closed
eellison wants to merge 4 commits intogh/eellison/307/basefrom
Closed
Add Dynamic Output Shape Tag For data-dependent ops, handle in FakeTensor#79170eellison wants to merge 4 commits intogh/eellison/307/basefrom
eellison wants to merge 4 commits intogh/eellison/307/basefrom
Conversation
[ghstack-poisoned]
Contributor
🔗 Helpful links
❌ 1 New Failures, 1 PendingAs of commit 90a4bc6 (more details on the Dr. CI page): Expand to see more
🕵️ 1 new failure recognized by patternsThe following CI failures do not appear to be due to upstream breakages
|
… in FakeTensor" [ghstack-poisoned]
…e in FakeTensor" [ghstack-poisoned]
ezyang
reviewed
Jun 9, 2022
ezyang
reviewed
Jun 9, 2022
|
|
||
|
|
||
| # TODO: use tags when available | ||
| # operators whose output shape depends on input tensor data |
ezyang
approved these changes
Jun 9, 2022
…e in FakeTensor" Adds a tag for operators which have an output whose shape depends on input Tensor data. This tag is tested through the fake_tensor tests. If an operator is not tagged appropriately, the fake tensor test will fallback to cpu to run the operator, and then output metadata will diverge. This helped me find `linalg_lstsq` among others. I included in the tags CompositeImplicitAutograd ops, although I'm not sure what the general policy is for these operators. One difficulty wrt/fake tensors was [aten::one_hot](https://codebrowser.bddppq.com/pytorch/pytorch/aten/src/ATen/native/Onehot.cpp.html#23). Because it is a `CompositeImplicitAutograd`, it won't be seen by `__torch_dispatch__`. It would be possible to handle this case either a) always throwing with `local_scalar_dense`, but this would add a huge amount of false negatives where we think operators have data-dependent output shapes but are actually not. b) Maybe I should also be defining `__torch_function__` for fake_tensor to error on specific composites. (this didn't work when I tried it, didn't get invoked, so maybe there's something I'm missing) Feel free to suggest a different name than `dynamic_output_shape`. [ghstack-poisoned]
Contributor
Author
|
@pytorchbot merge |
Collaborator
|
@pytorchbot successfully started a merge job. Check the current status here |
Contributor
|
Hey @eellison. |
facebook-github-bot
pushed a commit
that referenced
this pull request
Jun 10, 2022
…sor (#79170) Summary: Pull Request resolved: #79170 Approved by: https://github.com/ezyang Test Plan: contbuild & OSS CI, see https://hud.pytorch.org/commit/pytorch/pytorch/13a8867c01dd1191f21139915a8c821fa120cf58 Reviewed By: osalpekar Differential Revision: D37059412 Pulled By: osalpekar fbshipit-source-id: ee660b8792a03f1984216810c87f4975e9eac08e
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
4 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.
Stack from ghstack (oldest at bottom):
Adds a tag for operators which have an output whose shape depends on input Tensor data. This tag is tested through the fake_tensor tests. If an operator is not tagged appropriately, the fake tensor test will fallback to cpu to run the operator, and then output metadata will diverge. This helped me find
linalg_lstsqamong others.I included in the tags CompositeImplicitAutograd ops, although I'm not sure what the general policy is for these operators.
One difficulty wrt/fake tensors was aten::one_hot. Because it is a
CompositeImplicitAutograd, it won't be seen by__torch_dispatch__. It would be possible to handle this case eithera) always throwing with
local_scalar_dense, but this would add a huge amount of false negatives where we think operators have data-dependent output shapes but are actually not.b) Maybe I should also be defining
__torch_function__for fake_tensor to error on specific composites. (this didn't work when I tried it, didn't get invoked, so maybe there's something I'm missing)Feel free to suggest a different name than
dynamic_output_shape.