← all projects

Kernelbench

Inference kernel experiments

source code ↗

RMSNorm and softmax experiments with reproducible measurement tooling.

Python · PyTorch · Triton

notes from the build

  • Masked kernels, numerical-error reports, eager/compiled comparisons, and distributed gradient checks.
  • Two-process CPU/Gloo validation ran on Windows. CUDA kernel execution and multi-GPU NCCL validation remain pending; no GPU speedup is claimed.