Drop aten.transpose.Dimname from arm quantizer op list#20074
Drop aten.transpose.Dimname from arm quantizer op list#20074ashutoshsameer wants to merge 2 commits into
Conversation
Summary: PyTorch removed named tensor support in D-equivalent of pytorch/pytorch#173895 ("Remove named tensor", commit 05c965e7f8a7), which deleted all `Dimname`-based op overloads from `native_functions.yaml`, including `aten::transpose.Dimname`. Builds picked up after that change no longer register the `Dimname` overload. `executorch/backends/arm/quantizer/quantization_annotator.py` references `torch.ops.aten.transpose.Dimname` in a module-level constant list. As soon as that module is imported, evaluating the constant triggers `OpOverloadPacket.__getattr__('Dimname')` in `caffe2/torch/_ops.py`, which raises: ``` AttributeError: The underlying op of 'aten.transpose' has no overload name 'Dimname' ``` This breaks the PTQ step of the EMG release Ceres2 CC pipeline (FBLearner Flow run f1093420800) and any other workflow that imports the arm quantizer. Fix: drop the `transpose.Dimname` entry. The same list already contains `transpose.int`, which is the only `transpose` overload still registered after the named-tensor removal. No other `.Dimname` references remain anywhere under `fbcode/executorch/`. Differential Revision: D107675466
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/20074
Note: Links to docs will display an error until the docs builds have been completed. ❌ 2 New Failures, 1 Cancelled JobAs of commit 40d41e3 with merge base 502fdbe ( NEW FAILURES - The following jobs have failed:
CANCELLED JOB - The following job was cancelled. Please retry:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
|
meta-cla
Bot
added
the
CLA Signed
|
|
|
@ashutoshsameer has exported this pull request. If you are a Meta employee, you can view the originating Diff in D107675466. |
This PR needs a
|
digantdesai
left a comment
digantdesai
left a comment
There was a problem hiding this comment.
Review automatically exported from Phabricator review in Meta.
digantdesai
left a comment
digantdesai
left a comment
There was a problem hiding this comment.
Review automatically exported from Phabricator review in Meta.
2 participants
Summary:
PyTorch removed named tensor support in D-equivalent of pytorch/pytorch#173895 ("Remove named tensor", commit 05c965e7f8a7), which deleted all
Dimname-based op overloads fromnative_functions.yaml, includingaten::transpose.Dimname. Builds picked up after that change no longer register theDimnameoverload.executorch/backends/arm/quantizer/quantization_annotator.pyreferencestorch.ops.aten.transpose.Dimnamein a module-level constant list. As soon as that module is imported, evaluating the constant triggersOpOverloadPacket.__getattr__('Dimname')incaffe2/torch/_ops.py, which raises:This breaks the PTQ step of the EMG release Ceres2 CC pipeline (FBLearner Flow run f1093420800) and any other workflow that imports the arm quantizer.
Fix: drop the
transpose.Dimnameentry. The same list already containstranspose.int, which is the onlytransposeoverload still registered after the named-tensor removal. No other.Dimnamereferences remain anywhere underfbcode/executorch/.Differential Revision: D107675466
cc @digantdesai @freddan80 @per @zingo @oscarandersson8218 @mansnils @Sebastian-Larsson @robell @rascani