[DTensor] make expand_to_full_mesh_op_strategy filter incompatible out= strategies#172420
[DTensor] make expand_to_full_mesh_op_strategy filter incompatible out= strategies#172420wconstab wants to merge 8 commits intogh/wconstab/497/basefrom
Conversation
…t= strategies [ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/172420
Note: Links to docs will display an error until the docs builds have been completed. ❌ 7 New Failures, 1 Pending, 1 Unrelated FailureAs of commit 2647e22 with merge base 879fd46 ( NEW FAILURES - The following jobs have failed:
FLAKY - The following job failed but was likely due to flakiness present on trunk:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
…mpatible out= strategies" [ghstack-poisoned]
…mpatible out= strategies" [ghstack-poisoned]
| if ( | ||
| op_schema.is_out_variant_op() | ||
| and "out" in op_schema.kwargs_schema | ||
| and isinstance(op_schema.kwargs_schema["out"], OpStrategy) |
There was a problem hiding this comment.
nit: is there a TupleStrategy case we need to consider?
There was a problem hiding this comment.
are you talking about a case where out= provides a list of output tensors instead of a single tensor? I am not sure if such an op exists, trying to find one
…mpatible out= strategies" [ghstack-poisoned]
…mpatible out= strategies" [ghstack-poisoned]
…mpatible out= strategies" [ghstack-poisoned]
|
@pytorchbot merge |
Merge failedReason: PR 172477 is out of sync with the corresponding revision 27ce290 on branch gh/wconstab/499/orig that would be merged into main. This usually happens because there is a non ghstack change in the PR. Please sync them and try again (ex. make the changes on origin/gh/wconstab/497/orig and run ghstack). Details for Dev Infra teamRaised by workflow job |
…t= strategies ghstack-source-id: 1a0ef00 Pull Request resolved: pytorch/pytorch#172420
…mpatible out= strategies" [ghstack-poisoned]
|
Starting merge as part of PR stack under #172477 |
…mpatible out= strategies" [ghstack-poisoned]
|
Starting merge as part of PR stack under #172477 |
1 similar comment
|
Starting merge as part of PR stack under #172477 |
This enables the inplace filtering logic that skips strategies with incompatible input placements. Previously, inplace ops were able to generate expanded strategies incompatible with the current input placements, which aren't allowed to be modified by redistribution. Pull Request resolved: #172477 Approved by: https://github.com/zpcore ghstack dependencies: #172420
…t= strategies ghstack-source-id: cec0535 Pull Request resolved: pytorch/pytorch#172420
…t= strategies ghstack-source-id: 59a1306 Pull Request resolved: pytorch/pytorch#172420
Stack from ghstack (oldest at bottom):