Don't allocate result Tensors in out overloads: Reduction Ops #53218
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
Contributor
💊 CI failures summary and remediationsAs of commit 6a2c8a3 (more details on the Dr. CI page): 💚 💚 Looks good so far! There are no failures yet. 💚 💚 This comment was automatically generated by Dr. CI (expand for details).Follow this link to opt-out of these comments for your Pull Requests.Please report bugs/suggestions to the (internal) Dr. CI Users group. |
This was referenced Mar 4, 2021
Closed
ezyang
reviewed
Mar 4, 2021
Contributor
|
It looks like the duplicated logic doesn't perfectly reflect the original. If you want I can give you an accept (because I'm generally pro this PR) but it seems like there need to be substantive changes. |
…Ops" We have some operators that previously allowed you to pass in an undefined tensor to the out argument, and then would go on to allocate that for you. This behavior is broken and doesn't work in JIT when things are converted to/from IValues. Because of this, it blocks backend fallbacks because they force going through IValue. This PR removes that behavior and forces out arguments to be defined tensors. It only looks at reduction ops for now, there's likely more PRs coming for other ops. BC Breaking: This breaks BC since those ops previously allowed calling with undefined tensors and that isn't allowed anymore. Differential Revision: [D26795461](https://our.internmc.facebook.com/intern/diff/D26795461/) [ghstack-poisoned]
This was referenced Mar 8, 2021
Contributor
Well, this PR is failing tests. So definitionally, something wrong has happened. |
…Ops" We have some operators that previously allowed you to pass in an undefined tensor to the out argument, and then would go on to allocate that for you. This behavior is broken and doesn't work in JIT when things are converted to/from IValues. Because of this, it blocks backend fallbacks because they force going through IValue. This PR removes that behavior and forces out arguments to be defined tensors. It only looks at reduction ops for now, there's likely more PRs coming for other ops. BC Breaking: This breaks BC since those ops previously allowed calling with undefined tensors and that isn't allowed anymore. Differential Revision: [D26795461](https://our.internmc.facebook.com/intern/diff/D26795461/) [ghstack-poisoned]
…Ops" We have some operators that previously allowed you to pass in an undefined tensor to the out argument, and then would go on to allocate that for you. This behavior is broken and doesn't work in JIT when things are converted to/from IValues. Because of this, it blocks backend fallbacks because they force going through IValue. This PR removes that behavior and forces out arguments to be defined tensors. It only looks at reduction ops for now, there's likely more PRs coming for other ops. BC Breaking: This breaks BC since those ops previously allowed calling with undefined tensors and that isn't allowed anymore. Differential Revision: [D26795461](https://our.internmc.facebook.com/intern/diff/D26795461/) [ghstack-poisoned]
This was referenced Mar 9, 2021
…Ops" We have some operators that previously allowed you to pass in an undefined tensor to the out argument, and then would go on to allocate that for you. This behavior is broken and doesn't work in JIT when things are converted to/from IValues. Because of this, it blocks backend fallbacks because they force going through IValue. This PR removes that behavior and forces out arguments to be defined tensors. It only looks at reduction ops for now, there's likely more PRs coming for other ops. BC Breaking: This breaks BC since those ops previously allowed calling with undefined tensors and that isn't allowed anymore. Differential Revision: [D26795461](https://our.internmc.facebook.com/intern/diff/D26795461/) [ghstack-poisoned]
…Ops" We have some operators that previously allowed you to pass in an undefined tensor to the out argument, and then would go on to allocate that for you. This behavior is broken and doesn't work in JIT when things are converted to/from IValues. Because of this, it blocks backend fallbacks because they force going through IValue. This PR removes that behavior and forces out arguments to be defined tensors. It only looks at reduction ops for now, there's likely more PRs coming for other ops. BC Breaking: This breaks BC of the C++ frontend API since those ops previously allowed calling with undefined tensors and that isn't allowed anymore. Differential Revision: [D26795461](https://our.internmc.facebook.com/intern/diff/D26795461/) [ghstack-poisoned]
…Ops" We have some operators that previously allowed you to pass in an undefined tensor to the out argument, and then would go on to allocate that for you. This behavior is broken and doesn't work in JIT when things are converted to/from IValues. Because of this, it blocks backend fallbacks because they force going through IValue. This PR removes that behavior and forces out arguments to be defined tensors. It only looks at reduction ops for now, there's likely more PRs coming for other ops. BC Breaking: This breaks BC of the C++ frontend API since those ops previously allowed calling with undefined tensors and that isn't allowed anymore. Differential Revision: [D26795461](https://our.internmc.facebook.com/intern/diff/D26795461/) [ghstack-poisoned]
…Ops" We have some operators that previously allowed you to pass in an undefined tensor to the out argument, and then would go on to allocate that for you. This behavior is broken and doesn't work in JIT when things are converted to/from IValues. Because of this, it blocks backend fallbacks because they force going through IValue. This PR removes that behavior and forces out arguments to be defined tensors. It only looks at reduction ops for now, there's likely more PRs coming for other ops. BC Breaking: This breaks BC of the C++ frontend API since those ops previously allowed calling with undefined tensors and that isn't allowed anymore. Differential Revision: [D26795461](https://our.internmc.facebook.com/intern/diff/D26795461/) [ghstack-poisoned]
Contributor
Author
…Ops" We have some operators that previously allowed you to pass in an undefined tensor to the out argument, and then would go on to allocate that for you. This behavior is broken and doesn't work in JIT when things are converted to/from IValues. Because of this, it blocks backend fallbacks because they force going through IValue. This PR removes that behavior and forces out arguments to be defined tensors. It only looks at reduction ops for now, there's likely more PRs coming for other ops. BC Breaking: This breaks BC of the C++ frontend API since those ops previously allowed calling with undefined tensors and that isn't allowed anymore. Differential Revision: [D26795461](https://our.internmc.facebook.com/intern/diff/D26795461/) [ghstack-poisoned]
Contributor
Author
|
nvm, rebasing broke them again |
ezyang
reviewed
Mar 21, 2021
| ScalarType dtype) | ||
| { | ||
| auto shape = shape_from_dim_mask(self, mask, keepdim); | ||
| TORCH_CHECK(result.defined(), "Cannot create a new tensor inside a reduction op. You likely tried to call an operator with an out argument but the out argument was an undefined tensor."); |
Contributor
Author
There was a problem hiding this comment.
This can be triggered by user code without a bug in PyTorch when they call an out overload from the C++ API with an undefined tensor. So this should stay a TORCH_CHECK.
ezyang
reviewed
Mar 21, 2021
ezyang
reviewed
Mar 22, 2021
ezyang
approved these changes
Mar 22, 2021
…Ops" We have some operators that previously allowed you to pass in an undefined tensor to the out argument, and then would go on to allocate that for you. This behavior is broken and doesn't work in JIT when things are converted to/from IValues. Because of this, it blocks backend fallbacks because they force going through IValue. This PR removes that behavior and forces out arguments to be defined tensors. It only looks at reduction ops for now, there's likely more PRs coming for other ops. BC Breaking: This breaks BC of the C++ frontend API since those ops previously allowed calling with undefined tensors and that isn't allowed anymore. Differential Revision: [D26795461](https://our.internmc.facebook.com/intern/diff/D26795461/) [ghstack-poisoned]
…Ops" We have some operators that previously allowed you to pass in an undefined tensor to the out argument, and then would go on to allocate that for you. This behavior is broken and doesn't work in JIT when things are converted to/from IValues. Because of this, it blocks backend fallbacks because they force going through IValue. This PR removes that behavior and forces out arguments to be defined tensors. It only looks at reduction ops for now, there's likely more PRs coming for other ops. BC Breaking: This breaks BC of the C++ frontend API since those ops previously allowed calling with undefined tensors and that isn't allowed anymore. Differential Revision: [D26795461](https://our.internmc.facebook.com/intern/diff/D26795461/) [ghstack-poisoned]
…Ops" We have some operators that previously allowed you to pass in an undefined tensor to the out argument, and then would go on to allocate that for you. This behavior is broken and doesn't work in JIT when things are converted to/from IValues. Because of this, it blocks backend fallbacks because they force going through IValue. This PR removes that behavior and forces out arguments to be defined tensors. It only looks at reduction ops for now, there's likely more PRs coming for other ops. BC Breaking: This breaks BC of the C++ frontend API since those ops previously allowed calling with undefined tensors and that isn't allowed anymore. Differential Revision: [D26795461](https://our.internmc.facebook.com/intern/diff/D26795461/) [ghstack-poisoned]
…Ops" We have some operators that previously allowed you to pass in an undefined tensor to the out argument, and then would go on to allocate that for you. This behavior is broken and doesn't work in JIT when things are converted to/from IValues. Because of this, it blocks backend fallbacks because they force going through IValue. This PR removes that behavior and forces out arguments to be defined tensors. It only looks at reduction ops for now, there's likely more PRs coming for other ops. BC Breaking: This breaks BC of the C++ frontend API since those ops previously allowed calling with undefined tensors and that isn't allowed anymore. Differential Revision: [D26795461](https://our.internmc.facebook.com/intern/diff/D26795461/) [ghstack-poisoned]
…Ops" We have some operators that previously allowed you to pass in an undefined tensor to the out argument, and then would go on to allocate that for you. This behavior is broken and doesn't work in JIT when things are converted to/from IValues. Because of this, it blocks backend fallbacks because they force going through IValue. This PR removes that behavior and forces out arguments to be defined tensors. It only looks at reduction ops for now, there's likely more PRs coming for other ops. BC Breaking: This breaks BC of the C++ frontend API since those ops previously allowed calling with undefined tensors and that isn't allowed anymore. Differential Revision: [D26795461](https://our.internmc.facebook.com/intern/diff/D26795461/) [ghstack-poisoned]
…Ops" We have some operators that previously allowed you to pass in an undefined tensor to the out argument, and then would go on to allocate that for you. This behavior is broken and doesn't work in JIT when things are converted to/from IValues. Because of this, it blocks backend fallbacks because they force going through IValue. This PR removes that behavior and forces out arguments to be defined tensors. It only looks at reduction ops for now, there's likely more PRs coming for other ops. BC Breaking: This breaks BC of the C++ frontend API since those ops previously allowed calling with undefined tensors and that isn't allowed anymore. Differential Revision: [D26795461](https://our.internmc.facebook.com/intern/diff/D26795461/) [ghstack-poisoned]
…Ops" We have some operators that previously allowed you to pass in an undefined tensor to the out argument, and then would go on to allocate that for you. This behavior is broken and doesn't work in JIT when things are converted to/from IValues. Because of this, it blocks backend fallbacks because they force going through IValue. This PR removes that behavior and forces out arguments to be defined tensors. It only looks at reduction ops for now, there's likely more PRs coming for other ops. BC Breaking: This breaks BC of the C++ frontend API since those ops previously allowed calling with undefined tensors and that isn't allowed anymore. Differential Revision: [D26795461](https://our.internmc.facebook.com/intern/diff/D26795461/) [ghstack-poisoned]
…Ops" We have some operators that previously allowed you to pass in an undefined tensor to the out argument, and then would go on to allocate that for you. This behavior is broken and doesn't work in JIT when things are converted to/from IValues. Because of this, it blocks backend fallbacks because they force going through IValue. This PR removes that behavior and forces out arguments to be defined tensors. It only looks at reduction ops for now, there's likely more PRs coming for other ops. BC Breaking: This breaks BC of the C++ frontend API since those ops previously allowed calling with undefined tensors and that isn't allowed anymore. Differential Revision: [D26795461](https://our.internmc.facebook.com/intern/diff/D26795461/) [ghstack-poisoned]
…Ops" We have some operators that previously allowed you to pass in an undefined tensor to the out argument, and then would go on to allocate that for you. This behavior is broken and doesn't work in JIT when things are converted to/from IValues. Because of this, it blocks backend fallbacks because they force going through IValue. This PR removes that behavior and forces out arguments to be defined tensors. It only looks at reduction ops for now, there's likely more PRs coming for other ops. BC Breaking: This breaks BC of the C++ frontend API since those ops previously allowed calling with undefined tensors and that isn't allowed anymore. Differential Revision: [D26795461](https://our.internmc.facebook.com/intern/diff/D26795461/) [ghstack-poisoned]
Contributor
|
This pull request has been merged in db75eba. |
peterbell10
added a commit
that referenced
this pull request
Apr 8, 2021
Mean is broken for complex types, since #53218 it's now allocating the result as a real tensor which discards the imaginary component. This wasn't picked up in testing because `_test_dim_ops` tests are defined as closures inside of `_test_dim_ops` instead of as methods on the test class. The result is, they never get run. For best results, view diff with "Hide whitespace changes". [ghstack-poisoned]
peterbell10
added a commit
that referenced
this pull request
Apr 8, 2021
Mean is broken for complex types, since #53218 it's now allocating the result as a real tensor which discards the imaginary component. This wasn't picked up in testing because `_test_dim_ops` tests are defined as closures inside of `_test_dim_ops` instead of as methods on the test class. The result is, they never get run. For best results, view diff with "Hide whitespace changes". [ghstack-poisoned]
facebook-github-bot
pushed a commit
that referenced
this pull request
Apr 9, 2021
Summary: Pull Request resolved: #55640 Mean is broken for complex types, since #53218 it's now allocating the result as a real tensor which discards the imaginary component. This wasn't picked up in testing because `_test_dim_ops` tests are defined as closures inside of `_test_dim_ops` instead of as methods on the test class. The result is, they never get run. For best results, view diff with "Hide whitespace changes". Test Plan: Imported from OSS Reviewed By: ngimel Differential Revision: D27671127 Pulled By: mruberry fbshipit-source-id: 4a1f6fea1048919fda7339c867ee78e88f2d7bd2
krshrimali
pushed a commit
to krshrimali/pytorch
that referenced
this pull request
May 19, 2021
Summary: Pull Request resolved: pytorch#55640 Mean is broken for complex types, since pytorch#53218 it's now allocating the result as a real tensor which discards the imaginary component. This wasn't picked up in testing because `_test_dim_ops` tests are defined as closures inside of `_test_dim_ops` instead of as methods on the test class. The result is, they never get run. For best results, view diff with "Hide whitespace changes". Test Plan: Imported from OSS Reviewed By: ngimel Differential Revision: D27671127 Pulled By: mruberry fbshipit-source-id: 4a1f6fea1048919fda7339c867ee78e88f2d7bd2
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.
Stack from ghstack:
We have some operators that previously allowed you to pass in an undefined tensor to the out argument,
and then would go on to allocate that for you. This behavior is broken and doesn't work in JIT when things
are converted to/from IValues. Because of this, it blocks backend fallbacks because they force going
through IValue.
This PR removes that behavior and forces out arguments to be defined tensors.
It only looks at reduction ops for now, there's likely more PRs coming for other ops.
BC Breaking: This breaks BC of the C++ frontend API since those ops previously allowed calling with undefined tensors and that isn't allowed anymore.
Differential Revision: D26795461