pytorch
213ada85 - [pyper] to + lengths_to_offsets with nnpi shape inference support (#5931)

Commit
2 years ago
[pyper] to + lengths_to_offsets with nnpi shape inference support (#5931) Summary: X-link: https://github.com/pytorch/glow/pull/5931 Revert: D35214444 (https://github.com/pytorch/pytorch/commit/91a72e9021c73b19b2db2a72e65e53e514c8abe0) Revert commit changeset: f2f8d96918f3 Original diff: D34696255 (https://github.com/pytorch/pytorch/commit/f2ca4341c9008472022227b43d943af3e5e46e81) Test Plan: In addition to the test plan of D34696255 (https://github.com/pytorch/pytorch/commit/f2ca4341c9008472022227b43d943af3e5e46e81), ran the following for glow shape inference support: `buck test glow/fb/torch_glow/tests/model_tests:basic_shape_tests -- ShapeInferenceTest.TestToLengthsToOffsets` `servicelab create cogwheel_pyper_inference_fullsync_ads_inline_cvr_post_imp -a D35226685` https://www.internalfb.com/intern/servicelab/1002759356/ `servicelab create cogwheel_pyper_inference_fullsync_ads_10x_ctr_mbl_feed_non_mimo -a D35226685` https://www.internalfb.com/intern/servicelab/1002759361/ Model published: f335674524. Inference check job finished: f335693514. Inference test tw tasks: https://fburl.com/tupperware/8pb3ep45. I checked that replayer and storage tier task logs look good. Nnpi tier logs (https://fburl.com/tupperware/4grcwoil, P493812547) have various printouts related other longstanding operators missing shape inference inference support. I'll ask Oleg to sanity check these. Reviewed By: khabinov Differential Revision: D35226685 Pull Request resolved: https://github.com/pytorch/pytorch/pull/75609 Approved by: https://github.com/mikeiovine
Author
Committer
Parents
Loading