attn-hip / README.md

Add kernel card (library_name: kernels)

603e8e5 verified 18 days ago

610 Bytes

library_name: kernels
tags:
  - kernels
  - rocm

Flash-Attention Prefill (RDNA4/gfx1201)

Dense flash-attention prefill (causal + sliding-window + additive mask) as a native HIP + rocWMMA kernel.

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

from kernels import get_kernel
kernel = get_kernel("pat883/attn-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.