Repository navigation

RDNA4 phase 4: FlashAttention on gfx12 WMMA (bf16, FP8 QK^T) - #397

Open

morousg wants to merge 8 commits into

mainfrom

rdna4/phase-4-wmma

Open

RDNA4 phase 4: FlashAttention on gfx12 WMMA (bf16, FP8 QK^T)#397
morousg wants to merge 8 commits into
mainfrom
rdna4/phase-4-wmma

Commits on Oct 18, 2026
- HIP: build gfx1200 and gfx1201 by default, add attention wave helpers
  
  albertandaluz
  and
  claude
  committed
  2e25ffb
- automatically detected host gpu arch by default
  Albert Andaluz
  committed
  1cf881b
- Attention: test FP8 KV, flash decode and the mma kernel on CUDA
  
  albertandaluz
  and
  claude
  committed
  a1df596
- Merge branch 'rdna4/phase-1-hip-targets' into rdna4/phase-3-hip-attention
  albertandaluz
  committed
  e53bfaa
- Attention: scalar FlashAttention, FP8/int8 KV, decode and softmax on HIP
  
  albertandaluz
  and
  claude
  committed
  7c78320
- HIP arch docs: the detected GPU is the default, HIP_ARCH seeds the first configure
  
  albertandaluz
  and
  claude
  committed
  7ce748d
- Merge branch 'rdna4/phase-1-hip-targets' into rdna4/phase-3-hip-attention
  albertandaluz
  committed
  d185619
- Attention: FlashAttention on RDNA4 WMMA (bf16, FP8 QK^T) for HIP
  
  albertandaluz
  and
  claude
  committed
  7f1d2ea