trunk/f6585f7f189d6d457fcec520b399a39140a8ba8a: [MPS] Migrate attention from mps graph to metal fully (#201745)
- PyTorch: 2749 events in the last 90 days
- PyTorch: 2706th Release in the last 90 days
- Previous: earlier the same day · trunk/199499c48eb5f76febc1fb3dd5c59a6facfd449b: Use vectorized kernels for index_copy (#201814)
What happened
Migrates attention from MPS Graph to Metal. Look at the table below for PR2 to look at the speedups for various shapes. Most of the shapes with this PR is faster than previous MPS Graph, though some of the PRs in the stack might be worth merging to squeeze the most performance. But discarding those PRs, this PR already achieves a nice geomean on M5 Pro and M3 ultra machines so I think it's worth to merge this initia…
Summary assembled by rule from the sources below