GPU acceleration

MSL doesn’t support GPU acceleration yet. Programs in a distribution can’t use the Mac’s GPU, so CUDA, ROCm, Vulkan and OpenCL workloads that need a GPU don’t run with hardware acceleration.

GPU support for machine learning workloads is in progress, and tracked in #13.

Alternatives

• Run GPU workloads on macOS directly, with tools that use Metal, such as PyTorch’s MPS backend or Apple’s MLX. Keep the rest of your toolchain in the distribution, and share files through /mnt/macos (see Working across file systems).
• Run a model server on macOS and call it from the distribution over the network. From a distribution, host.internal points at macOS (see Networking).