moe-splitk-hip / README.md
pat883's picture
Add kernel card (library_name: kernels)
c842129 verified
|
Raw
History Blame Contribute Delete
560 Bytes
metadata
library_name: kernels
tags:
  - kernels
  - rocm

Split-K W4A8 MoE GEMM (RDNA4/gfx1201)

Split-K W4A8 fp8-WMMA MoE GEMM with atomic scatter.

Built with kernel-builder for AMD RDNA4 (gfx1201). Load with the kernels library:

from kernels import get_kernel
kernel = get_kernel("pat883/moe-splitk-hip")

Requires a ROCm PyTorch build (torch 2.10 / ROCm 7.x) on an RDNA4 card. Built variants: torch210-cxx11-rocm70 and torch210-cxx11-rocm71.