trunk/4144bea4b7c2f67da9d4840f4659f977bbd754e5: [MPS] Remove svd_kernel_mps small-input CPU gate (#195950)
- PyTorch: 590 events in the last 90 days
- PyTorch: 584th Release in the last 90 days
- Previous: earlier the same day · trunk/13948faea12f1eb3a6df72f181a51f3e65dbf309: [MPS] Orthonormalize null-space U columns in svd_jacobi (#196139)
What happened
svd_kernel_mps routed inputs with numel infos.any().item(), a mandatory device->host sync on every call regardless of backend, so the GPU path does not pipeline and there is no async barrier for the gate to avoid. The CPU fallback then adds two data transfers (A to host, U/S/Vh back) on top of that same sync, making it strictly slower. Measured on M5 Pro, native GPU path vs the exact work the gate did (A.to(cpu) ->…
Summary assembled by rule from the sources below