Preserve DeviceMesh identity in expand_rule's op_schema deepcopy#488
Open
fmassa wants to merge 5 commits into
Open
Preserve DeviceMesh identity in expand_rule's op_schema deepcopy#488fmassa wants to merge 5 commits into
fmassa wants to merge 5 commits into
Conversation
expand_rule used copy.deepcopy(op_schema_) to snapshot the schema before mutating it. DeviceMesh has no __deepcopy__, so deepcopy went through __getstate__/__setstate__ and produced a fresh DeviceMesh with the same logical content but an empty _flatten_mapping cache. The DTensorSpecs returned from expand_rule carried these duplicates, which propagated into the sharding solution. apply_placement's pre-warming in _apply_placement_common only populates the user mesh's cache. When _optimize_same_nd_sharding_as_1d inside make_fx hit a duplicate mesh, _flatten() cache-missed and dispatched as_strided on the real rank_map — failing FakeTensorMode's non-fake-input check. Which solution the solver picked depended on the cost model, so the failure surfaced on g5/A10G CI but not on H100. Fix: _deepcopy_preserving_mesh pre-seeds copy.deepcopy's memo with DeviceMesh identity mappings so duplicates aren't produced. Adds a regression test that asserts every spec mesh's root has a warm _flatten cache after apply_placement. Authored with Claude.
meta-cla
Bot
added
the
CLA Signed
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
expand_rule used copy.deepcopy(op_schema_) to snapshot the schema before mutating it. DeviceMesh has no deepcopy, so deepcopy went through getstate/setstate and produced a fresh DeviceMesh with the same logical content but an empty _flatten_mapping cache. The DTensorSpecs returned from expand_rule carried these duplicates, which propagated into the sharding solution.
apply_placement's pre-warming in _apply_placement_common only populates the user mesh's cache. When _optimize_same_nd_sharding_as_1d inside make_fx hit a duplicate mesh, _flatten() cache-missed and dispatched as_strided on the real rank_map — failing FakeTensorMode's non-fake-input check. Which solution the solver picked depended on the cost model, so the failure surfaced on g5/A10G CI but not on H100.
Fix: _deepcopy_preserving_mesh pre-seeds copy.deepcopy's memo with DeviceMesh identity mappings so duplicates aren't produced. Adds a regression test that asserts every spec mesh's root has a warm _flatten cache after apply_placement.
Authored with Claude.