Skip to main content

Mojo module

mha_rdna

RDNA-specific MHA kernel configurations and entry points.

This module provides attention computation configurations optimized for AMD RDNA consumer GPUs (Radeon RX 7000/8000 series, gfx11xx/gfx12xx).

Key features:

  • Wave32 execution model
  • 16x16x16 WMMA shape (only supported shape for RDNA)
  • Optimized shared memory management with K/P reuse

Structs

Was this page helpful?