Expand description
Accelerator selection for the tensorized EML forward pass.
phop has four forward-eval backends: NVIDIA CUDA (gpu-cuda, exact f64, via oxicuda), native
Apple Metal (gpu-metal, f32, via oxicuda-metal, macOS only), a portable
WebGPU/Metal/Vulkan/DX12 path (gpu-wgpu, f32, via crate::wgpu_forward), and the
always-available CPU path (f64). gpu_backend picks the best one present at runtime, in the
order CUDA → Metal → wgpu → CPU — CUDA first for its f64 precision and existing tuned
kernels, native Metal next on Apple hardware, wgpu for portability elsewhere, CPU as the
universal fallback.
Enums§
- GpuBackend
- A compute backend for the EML forward pass.
Functions§
- gpu_
backend - Select the best available forward-eval backend at runtime: CUDA → Metal → wgpu → CPU.