MPS

MPS (Metal Performance Shaders) is Apple’s framework of hand-tuned compute kernels built on top of Metal — the GPU primitives for things like matrix multiply, convolution, and other neural-net ops, plus a higher-level graph API (MPS Graph) that schedules and fuses them. It’s roughly Apple’s analogue to NVIDIA’s cuDNN / cuBLAS: the optimized building blocks that frameworks call rather than writing raw GPU code themselves. In practice you meet MPS as PyTorch’s mps device (torch.device("mps")), which is how PyTorch runs on Apple-silicon GPUs instead of falling back to CPU; Apple’s MLX also leans on the same Metal foundation. Note the name collision: “MPS” here means Metal Performance Shaders, not “model-parallel” anything.