Improve error message when view of intermediate is returned from autograd.Function and marked dirty #149543
Conversation
…d.Function and marked dirty [ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/149543
Note: Links to docs will display an error until the docs builds have been completed. ✅ You can merge normally! (2 Unrelated Failures)As of commit 3455c3b with merge base e57cdb8 ( FLAKY - The following job failed but was likely due to flakiness present on trunk:
UNSTABLE - The following job is marked as unstable, possibly due to flakiness on trunk:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
soulitzer
added
release notes: autograd
soulitzer
changed the title
…d from autograd.Function and marked dirty" Fixes #149252 [ghstack-poisoned]
…d from autograd.Function and marked dirty" Fixes #149252 [ghstack-poisoned]
| "Only input Tensors that have been mutated should be passed to " | ||
| "ctx.mark_dirty()."; | ||
| // We reach this path in the view of intermediate case | ||
| TORCH_CHECK(!var.is_view(), mark_dirty_error_msg); |
There was a problem hiding this comment.
How does the intermediate case result in var being a view?
There was a problem hiding this comment.
a view of a intermediate tensor is still a view because we don't disable ADInplaceOrView
There was a problem hiding this comment.
What prevents an input from being a view?
There was a problem hiding this comment.
Inputs can be views and that is fine, we should do the usual rebase history stuff. If you mean to ask what happens if you are a view of the input, your view will be a leaf that require_grad being modified inplace and that is handled else-where (the previous branch of the if-else).
This PR is mainly targeting the case where the view was created from an intermediate, in which we are leaf with requires_grad=False and will immediately error (with an internal assert) because this autograd.Function is marked dirty and will call rebase_history on it.
|
@pytorchbot merge |
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 |
Merge failedReason: 1 jobs have failed, first few of them are: trunk / cuda12.1-py3.10-gcc9-sm80 / build Details for Dev Infra teamRaised by workflow job |
|
@pytorchbot merge |
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 |
…grad.Function and marked dirty (pytorch#149543) Fixes pytorch#149252 Pull Request resolved: pytorch#149543 Approved by: https://github.com/zou3519 ghstack dependencies: pytorch#149220
3 participants
Stack from ghstack (oldest at bottom):
Fixes #149252