kernelport

Bring CUDA-only kernels to Mac, AMD, browsers and CPU, with proof they are correct. Your coding agent writes the ports; kernelport runs each one on real hardware and checks it against a reference that was itself validated against the original CUDA kernel.

pip install kernelport

Check the ports on your GPU →Runs 12 WebGPU ports here in your browser and compares every output with the validated reference. Source and docs →README, install guide, the agent skill and the MCP server. Real code: llm.c →9 unmodified llm.c kernels ported to Triton, Metal and WebGPU, and a bug it found upstream.

What it looks like

Terminal recording: kernelport validates a softmax reference against the original CUDA kernel, a first Triton port fails with NaN, the fixed port passes 24 of 24 cases on an RTX 5060

Backends

Triton (NVIDIA, AMD): verified on an RTX 5060
Metal via MLX (Apple Silicon): verified on Apple M1 in CI on every push
WebGPU (WGSL, any GPU and browsers): verified on NVIDIA, Apple, Intel and software adapters
CPU via Numba: verified
HIP (AMD): ports written, not yet run on AMD hardware

A port is reported as PASS only if its current code ran on real hardware and matched.