mps-events
Metal shared events inside PyTorch's MPS stream (GPU-side waits a host thread signals) and CPU access to shared MPS tensors, for Apple Silicon with torch >= 2.15 (MPS tensors in unified memory). Used by transformers' streamed MoE experts: the GPU waits on an event while a host thread installs the experts a layer missed, instead of draining the stream at every layer.
Built for torch 2.15 / Python 3.12 and 3.13 (build_ext.py, source mps_events.mm). Interim home until a
kernel-builder build or a PyTorch API covers it.
- Downloads last month
- -
kernels
metal
apache-2.0
Supported hardwares new
Metal
- OS
- macos
- Arch
- aarch64


