Use object identity for deepcopy memo #126089
Closed
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
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/126089
Note: Links to docs will display an error until the docs builds have been completed. ❌ 24 New Failures, 1 PendingAs of commit 1c3724f with merge base 96a5698 ( NEW FAILURES - The following jobs have failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
ezyang
added a commit
that referenced
this pull request
May 13, 2024
Signed-off-by: Edward Z. Yang <ezyang@meta.com> ghstack-source-id: 622d69516328977309a5e30a820507b68f0964a0 Pull Request resolved: #126089
github-actions
bot
requested review from
albanD,
antoniojkim,
bdhirsh,
miladm and
SherlockNoMad
|
@jerryzh168 to add the test |
zdevito
approved these changes
May 13, 2024
|
assuming there is nothing suspicious in internal tests, since this is changing (buggy) behavior |
|
thanks @davidberard98 for helping with finishing this |
davidberard98
added a commit
that referenced
this pull request
May 14, 2024
Copy of #126089, with some fixes & tests (TODO) [TODO description] [ghstack-poisoned]
davidberard98
added a commit
that referenced
this pull request
May 14, 2024
Copy of #126089, with some fixes & tests (TODO) [TODO description] [ghstack-poisoned]
malfet
approved these changes
May 14, 2024
davidberard98
added a commit
that referenced
this pull request
May 14, 2024
Copy of #126089, with some fixes & tests (TODO) [TODO description] Differential Revision: [D57340612](https://our.internmc.facebook.com/intern/diff/D57340612) [ghstack-poisoned]
davidberard98
added a commit
that referenced
this pull request
May 15, 2024
Copy of #126089, with some additional fixes & tests Partial fix for #125635: previously, the deepcopy implementation would group together any tensors with any aliasing relationship and assign them to the same tensor. This was sort of good if you have two tensors `b = a.detach()`, because then if you deepcopy `list = [a, b]` to `list2 = list.deepcopy()`, then writes to `list2[0]` will also modify `list2[1]`. But for the most part, it's bad; (1) if you have `b = a.as_strided((4, 4), (16, 1), 16)`, then it'll make `b == a` in the deepcopied implementation, which is completely wrong; and (2) even if you have `b = a.detach()`, these are still initially two different tensors which become the same tensor after the old deepcopy implementation. The new implementation only groups together tensors that have the same identity. This is a partial fix, but it's more reasonable. What changes: * (becomes more correct): different views of the same base tensor will no longer all become equal after deepcopying * (still kind of wrong): views won't actually alias each other after deepcopying. * (arguably a minor regression): equivalent views of the same tensor will no longer be copied to the same tensor - so they won't alias. BC breaking: C++ deepcopy interface changes from accepting `IValue::HashAliasedIValueMap memo` to accepting `IValue::HashIdentityIValueMap memo`. If there are objections, we can keep the old API. However, it seems likely that users generally won't try to deepcopy from C++. Differential Revision: [D57340612](https://our.internmc.facebook.com/intern/diff/D57340612) cc ezyang gchanan [ghstack-poisoned]
davidberard98
added a commit
that referenced
this pull request
May 15, 2024
Copy of #126089, with some additional fixes & tests Partial fix for #125635: previously, the deepcopy implementation would group together any tensors with any aliasing relationship and assign them to the same tensor. This was sort of good if you have two tensors `b = a.detach()`, because then if you deepcopy `list = [a, b]` to `list2 = list.deepcopy()`, then writes to `list2[0]` will also modify `list2[1]`. But for the most part, it's bad; (1) if you have `b = a.as_strided((4, 4), (16, 1), 16)`, then it'll make `b == a` in the deepcopied implementation, which is completely wrong; and (2) even if you have `b = a.detach()`, these are still initially two different tensors which become the same tensor after the old deepcopy implementation. The new implementation only groups together tensors that have the same identity. This is a partial fix, but it's more reasonable. What changes: * (becomes more correct): different views of the same base tensor will no longer all become equal after deepcopying * (still kind of wrong): views won't actually alias each other after deepcopying. * (arguably a minor regression): equivalent views of the same tensor will no longer be copied to the same tensor - so they won't alias. BC breaking: C++ deepcopy interface changes from accepting `IValue::HashAliasedIValueMap memo` to accepting `IValue::HashIdentityIValueMap memo`. If there are objections, we can keep the old API. However, it seems likely that users generally won't try to deepcopy from C++. Differential Revision: [D57340612](https://our.internmc.facebook.com/intern/diff/D57340612) cc ezyang gchanan [ghstack-poisoned]
pytorchmergebot
pushed a commit
that referenced
this pull request
May 17, 2024
Copy of #126089, with some additional fixes & tests Partial fix for #125635: previously, the deepcopy implementation would group together any tensors with any aliasing relationship and assign them to the same tensor. This was sort of good if you have two tensors `b = a.detach()`, because then if you deepcopy `list = [a, b]` to `list2 = list.deepcopy()`, then writes to `list2[0]` will also modify `list2[1]`. But for the most part, it's bad; (1) if you have `b = a.as_strided((4, 4), (16, 1), 16)`, then it'll make `b == a` in the deepcopied implementation, which is completely wrong; and (2) even if you have `b = a.detach()`, these are still initially two different tensors which become the same tensor after the old deepcopy implementation. The new implementation only groups together tensors that have the same identity. This is a partial fix, but it's more reasonable. What changes: * (becomes more correct): different views of the same base tensor will no longer all become equal after deepcopying * (still kind of wrong): views won't actually alias each other after deepcopying. * (arguably a minor regression): equivalent views of the same tensor will no longer be copied to the same tensor - so they won't alias. BC breaking: C++ deepcopy interface changes from accepting `IValue::HashAliasedIValueMap memo` to accepting `IValue::HashIdentityIValueMap memo`. If there are objections, we can keep the old API. However, it seems likely that users generally won't try to deepcopy from C++. Differential Revision: [D57406306](https://our.internmc.facebook.com/intern/diff/D57406306) Pull Request resolved: #126126 Approved by: https://github.com/ezyang
ZelboK
pushed a commit
to ZelboK/pytorch
that referenced
this pull request
May 19, 2024
Copy of pytorch#126089, with some additional fixes & tests Partial fix for pytorch#125635: previously, the deepcopy implementation would group together any tensors with any aliasing relationship and assign them to the same tensor. This was sort of good if you have two tensors `b = a.detach()`, because then if you deepcopy `list = [a, b]` to `list2 = list.deepcopy()`, then writes to `list2[0]` will also modify `list2[1]`. But for the most part, it's bad; (1) if you have `b = a.as_strided((4, 4), (16, 1), 16)`, then it'll make `b == a` in the deepcopied implementation, which is completely wrong; and (2) even if you have `b = a.detach()`, these are still initially two different tensors which become the same tensor after the old deepcopy implementation. The new implementation only groups together tensors that have the same identity. This is a partial fix, but it's more reasonable. What changes: * (becomes more correct): different views of the same base tensor will no longer all become equal after deepcopying * (still kind of wrong): views won't actually alias each other after deepcopying. * (arguably a minor regression): equivalent views of the same tensor will no longer be copied to the same tensor - so they won't alias. BC breaking: C++ deepcopy interface changes from accepting `IValue::HashAliasedIValueMap memo` to accepting `IValue::HashIdentityIValueMap memo`. If there are objections, we can keep the old API. However, it seems likely that users generally won't try to deepcopy from C++. Differential Revision: [D57406306](https://our.internmc.facebook.com/intern/diff/D57406306) Pull Request resolved: pytorch#126126 Approved by: https://github.com/ezyang
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 (oldest at bottom):
Internal xref: https://fb.workplace.com/groups/1405155842844877/posts/8379274608766264
Signed-off-by: Edward Z. Yang ezyang@meta.com