Fix assigning tensor with requires_grad as constant in export #137997
Conversation
When we insert cojstants into unlifted graph, we need to detach them if they require grad. Differential Revision: [D64406859](https://our.internmc.facebook.com/intern/diff/D64406859/) [ghstack-poisoned]
tugsbayasgalan
requested review from
angelayi,
avikchaudhuri,
ydwu4 and
zhxchen17
as code owners
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/137997
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit a9a5da7 with merge base f173623 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
|
This pull request was exported from Phabricator. Differential Revision: D64406859 |
When we insert cojstants into unlifted graph, we need to detach them if they require grad. Differential Revision: [D64406859](https://our.internmc.facebook.com/intern/diff/D64406859/) ghstack-source-id: 248115756 Pull Request resolved: #137997
…ort" When we insert cojstants into unlifted graph, we need to detach them if they require grad. Differential Revision: [D64406859](https://our.internmc.facebook.com/intern/diff/D64406859/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D64406859 |
Pull Request resolved: #137997 When we insert cojstants into unlifted graph, we need to detach them if they require grad. Differential Revision: [D64406859](https://our.internmc.facebook.com/intern/diff/D64406859/) ghstack-source-id: 248221115
…ort" When we insert cojstants into unlifted graph, we need to detach them if they require grad. Differential Revision: [D64406859](https://our.internmc.facebook.com/intern/diff/D64406859/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D64406859 |
avikchaudhuri
left a comment
avikchaudhuri
left a comment
There was a problem hiding this comment.
Rebase and check CI before landing...
torch/export/_unlift.py
Outdated
| if isinstance(value, torch.Tensor): | ||
| if value.requires_grad: | ||
| warnings.warn( | ||
| f"A model attribute `{const_name}` requires gradient. " |
| if value in original_tensor_to_detached_tensor: | ||
| value = original_tensor_to_detached_tensor[value] | ||
| else: | ||
| detached_value = value.detach() |
There was a problem hiding this comment.
maybe this code should be refactored into a util?
There was a problem hiding this comment.
It is bit hard to make it a util because this is only relevant in this specific location.
…ort" When we insert cojstants into unlifted graph, we need to detach them if they require grad BUT when we detach we need to preserve the original aliasing information. Differential Revision: [D64406859](https://our.internmc.facebook.com/intern/diff/D64406859/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D64406859 |
…ort" When we insert cojstants into unlifted graph, we need to detach them if they require grad BUT when we detach we need to preserve the original aliasing information. Differential Revision: [D64406859](https://our.internmc.facebook.com/intern/diff/D64406859/) [ghstack-poisoned]
Pull Request resolved: #137997 When we insert cojstants into unlifted graph, we need to detach them if they require grad. ghstack-source-id: b218468 Differential Revision: [D64406859](https://our.internmc.facebook.com/intern/diff/D64406859/)
|
@tugsbayasgalan has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
|
@pytorchbot merge (Initiating merge automatically since Phabricator Diff has merged) |
Merge startedYour change will be merged once all checks pass (ETA 0-4 Hours). Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
4 participants
Stack from ghstack (oldest at bottom):
When we insert cojstants into unlifted graph, we need to detach them if they require grad BUT when we detach we need to preserve the original aliasing information.
Differential Revision: D64406859