viable/strict/1788439121: [MPS] migrate log softmax to metal (#195441)
- PyTorch: 2653 events in the last 90 days
- PyTorch: 2612th Release in the last 90 days
- Previous: earlier the same day · trunk/56ff81dade97f30294ac89a786b1f3516ad993bb: [xpu] Add non split pool support on xpu caching allcoator (#192033)
What happened
Following #195368 of PRs for reducing memory usage for the ops which are mainly on modern model inference/training side and use MPS Graph, next one being log_softmax. To briefly explain the kernels, there are 4: log_softmax_row - used when dim we are reducing on is relatively small ( =128) to keep gpu saturated (reduction dim should be contiguous) log_softmax_partial and log_softmax_finalize - Regular 2 pass kernel,…
Summary assembled by rule from the sources below