Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/22958
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit d4699dc with merge base bb2683b ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
Bot
added
the
CLA Signed
Contributor
|
@ethansfng has exported this pull request. If you are a Meta employee, you can view the originating Diff in D120615213. |
This PR needs a
|
meta-codesync
Bot
changed the title
ethansfng
added a commit
to ethansfng/executorch
that referenced
this pull request
Sep 21, 2026
Summary: `test_emg_cascade_detector_fp32` failed four consecutive times because `RemovePermutesAroundElementwiseOps` generated an invalid view for a lower-rank broadcast constant. The affected graph permutes an input shaped `[1, 256, 1, 1]` to channel-last layout, multiplies it by a `[256]` vector, and then permutes the result back. While removing the surrounding permutes, the pass padded the vector shape to `[1, 1, 1, 256]`, applied the output permutation, and then blindly discarded the three dimensions originally added for broadcast alignment. Because the permutation moved the non-singleton `256` dimension into that discarded prefix, the pass derived `[1]` and attempted to reshape a 256-element layer-norm weight to it. Export then failed with `RuntimeError: shape [1] is invalid for input of size 256`. This change computes the prospective constant view before rewriting and verifies that every dimension which would be discarded remains a singleton after permutation. If a non-singleton dimension moved into that prefix, the optimization is skipped and the original permutes are preserved. Safe scalar and lower-rank constant rewrites continue to be optimized. A regression test covers the failing vector-broadcast pattern. Differential Revision: D120615213
ethansfng
force-pushed
the
export-D120615213
branch
from
September 21, 2026 16:41
fd65085 to
6e96da7
Compare
ethansfng
added a commit
to ethansfng/executorch
that referenced
this pull request
Sep 21, 2026
Summary: `test_emg_cascade_detector_fp32` failed four consecutive times because `RemovePermutesAroundElementwiseOps` generated an invalid view for a lower-rank broadcast constant. The affected graph permutes an input shaped `[1, 256, 1, 1]` to channel-last layout, multiplies it by a `[256]` vector, and then permutes the result back. While removing the surrounding permutes, the pass padded the vector shape to `[1, 1, 1, 256]`, applied the output permutation, and then blindly discarded the three dimensions originally added for broadcast alignment. Because the permutation moved the non-singleton `256` dimension into that discarded prefix, the pass derived `[1]` and attempted to reshape a 256-element layer-norm weight to it. Export then failed with `RuntimeError: shape [1] is invalid for input of size 256`. This change computes the prospective constant view before rewriting and verifies that every dimension which would be discarded remains a singleton after permutation. If a non-singleton dimension moved into that prefix, the optimization is skipped and the original permutes are preserved. Safe scalar and lower-rank constant rewrites continue to be optimized. A regression test covers the failing vector-broadcast pattern. Differential Revision: D120615213
ethansfng
force-pushed
the
export-D120615213
branch
from
September 21, 2026 16:44
6e96da7 to
92827c4
Compare
ethansfng
added a commit
to ethansfng/executorch
that referenced
this pull request
Sep 21, 2026
Summary: Pull Request resolved: pytorch#22958 `test_emg_cascade_detector_fp32` failed four consecutive times because `RemovePermutesAroundElementwiseOps` generated an invalid view for a lower-rank broadcast constant. The affected graph permutes an input shaped `[1, 256, 1, 1]` to channel-last layout, multiplies it by a `[256]` vector, and then permutes the result back. While removing the surrounding permutes, the pass padded the vector shape to `[1, 1, 1, 256]`, applied the output permutation, and then blindly discarded the three dimensions originally added for broadcast alignment. Because the permutation moved the non-singleton `256` dimension into that discarded prefix, the pass derived `[1]` and attempted to reshape a 256-element layer-norm weight to it. Export then failed with `RuntimeError: shape [1] is invalid for input of size 256`. This change computes the prospective constant view before rewriting and verifies that every dimension which would be discarded remains a singleton after permutation. If a non-singleton dimension moved into that prefix, the optimization is skipped and the original permutes are preserved. Safe scalar and lower-rank constant rewrites continue to be optimized. A regression test covers the failing vector-broadcast pattern. Differential Revision: D120615213
ethansfng
force-pushed
the
export-D120615213
branch
from
September 21, 2026 16:45
92827c4 to
0081db9
Compare
ethansfng
added a commit
to ethansfng/executorch
that referenced
this pull request
Sep 21, 2026
Summary: `test_emg_cascade_detector_fp32` failed four consecutive times because `RemovePermutesAroundElementwiseOps` generated an invalid view for a lower-rank broadcast constant. The affected graph permutes an input shaped `[1, 256, 1, 1]` to channel-last layout, multiplies it by a `[256]` vector, and then permutes the result back. While removing the surrounding permutes, the pass padded the vector shape to `[1, 1, 1, 256]`, applied the output permutation, and then blindly discarded the three dimensions originally added for broadcast alignment. Because the permutation moved the non-singleton `256` dimension into that discarded prefix, the pass derived `[1]` and attempted to reshape a 256-element layer-norm weight to it. Export then failed with `RuntimeError: shape [1] is invalid for input of size 256`. This change computes the prospective constant view before rewriting and verifies that every dimension which would be discarded remains a singleton after permutation. If a non-singleton dimension moved into that prefix, the optimization is skipped and the original permutes are preserved. Safe scalar and lower-rank constant rewrites continue to be optimized. A regression test covers the failing vector-broadcast pattern. Differential Revision: D120615213
ethansfng
force-pushed
the
export-D120615213
branch
from
September 21, 2026 16:49
0081db9 to
9028f77
Compare
Summary: Pull Request resolved: pytorch#22958 `test_emg_cascade_detector_fp32` failed four consecutive times because `RemovePermutesAroundElementwiseOps` generated an invalid view for a lower-rank broadcast constant. The affected graph permutes an input shaped `[1, 256, 1, 1]` to channel-last layout, multiplies it by a `[256]` vector, and then permutes the result back. While removing the surrounding permutes, the pass padded the vector shape to `[1, 1, 1, 256]`, applied the output permutation, and then blindly discarded the three dimensions originally added for broadcast alignment. Because the permutation moved the non-singleton `256` dimension into that discarded prefix, the pass derived `[1]` and attempted to reshape a 256-element layer-norm weight to it. Export then failed with `RuntimeError: shape [1] is invalid for input of size 256`. This change computes the prospective constant view before rewriting and verifies that every dimension which would be discarded remains a singleton after permutation. If a non-singleton dimension moved into that prefix, the optimization is skipped and the original permutes are preserved. Safe scalar and lower-rank constant rewrites continue to be optimized. A regression test covers the failing vector-broadcast pattern. Differential Revision: D120615213
ethansfng
force-pushed
the
export-D120615213
branch
from
September 21, 2026 16:53
9028f77 to
d4699dc
Compare
This branch has not been deployed
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.
Summary:
The affected graph permutes an input shaped
[1, 256, 1, 1]to channel-last layout, multiplies it by a[256]vector, and then permutes the result back. While removing the surrounding permutes, the pass padded the vector shape to[1, 1, 1, 256], applied the output permutation, and then blindly discarded the three dimensions originally added for broadcast alignment. Because the permutation moved the non-singleton256dimension into that discarded prefix, the pass derived[1]and attempted to reshape a 256-element layer-norm weight to it. Export then failed withRuntimeError: shape [1] is invalid for input of size 256.This change computes the prospective constant view before rewriting and verifies that every dimension which would be discarded remains a singleton after permutation. If a non-singleton dimension moved into that prefix, the optimization is skipped and the original permutes are preserved. Safe scalar and lower-rank constant rewrites continue to be optimized. A regression test covers the failing vector-broadcast pattern.
Differential Revision: D120615213