Repository navigation

Enable gfx950 FP8 HD128/256 fused-attn ASM forward from aiter - #741

Open

yaomingamd wants to merge 21 commits into

ROCm:devfrom

yaomingamd:ymmu/fp8_hd128_256_fwd_attns

Open

Enable gfx950 FP8 HD128/256 fused-attn ASM forward from aiter#741
yaomingamd wants to merge 21 commits into
ROCm:devfrom
yaomingamd:ymmu/fp8_hd128_256_fwd_attns

Commits on Sep 22, 2026
- Enable gfx950 FP8 HD128/256 fused-attn ASM forward.
  yaomingamd
  committed
  ced226c
Commits on Oct 6, 2026
- FP8 scale chagne to support real scal-factor for input tensors
  yaomingamd
  committed
  f6024b4
- modify FFI for support FP8 attention with scale factor for FP8 inputs
  yaomingamd
  committed
  708a31c
Commits on Oct 7, 2026
- Document why FP8 FMHA fails closed instead of falling through to CK.
  
  yaomingamd
  and
  cursoragent
  committed
  d5570f6
- Clarify that FP8 E4M3 is an input dtype and BF16 output is a separate constraint.
  
  yaomingamd
  and
  cursoragent
  committed
  e9de099
- Describe FP8 descales as one FP32 inverse scale per tensor.
  
  yaomingamd
  and
  cursoragent
  committed
  03bd6b9
- Note that "fp8bf16" is the AITER kernel tag, not the TE dtype name.
  
  yaomingamd
  and
  cursoragent
  committed
  ac9ec8b
- Describe the CK dtype filter as a Q/KV type check.
  
  yaomingamd
  and
  cursoragent
  committed
  990083d
- Bind FP8 scale_inv after the unused segment tensors in the JAX FFI.
  
  yaomingamd
  and
  cursoragent
  committed
  bb896a7
Commits on Oct 8, 2026
- Merge remote-tracking branch 'upstream/dev' into ymmu/fp8_hd128_256_fwd_attns.
  
  yaomingamd
  and
  cursoragent
  committed
  9c70b62
Commits on Oct 18, 2026
- Merge remote-tracking branch 'rocm/dev' into ymmu/fp8_hd128_256_fwd_attns.
  
  yaomingamd
  and
  cursoragent
  committed
  369e177
- Use one named skip marker for the gfx950 FP8 attention tests.
  
  yaomingamd
  and
  cursoragent
  committed
  5fbcf68
- Rename the CK fused-attention FP8 enum value to kFloat8.
  
  yaomingamd
  and
  cursoragent
  committed
  9e76ec8
- Clarify that the CK data-type string is an AITER kernel tag.
  
  yaomingamd
  and
  cursoragent
  committed
  4463bee
- Skip FP8 fused-attn kernel tests when AITER is disabled.
  
  yaomingamd
  and
  cursoragent
  committed
  28c623f
- Pass batch_size into the keyword-only fused attention helper.
  
  yaomingamd
  and
  cursoragent
  committed
  86d781d
Commits on Oct 20, 2026
- Select no fused-attn backend for FP8 when AITER v3 is disabled.
  
  yaomingamd
  and
  cursoragent
  committed
  bc0a7eb
Commits on Oct 21, 2026
- Treat FP8 scale inverses as primal fused-attn inputs under custom_vjp.
  
  yaomingamd
  and
  cursoragent
  committed
  cb22dea
- Keep the FP8 ASM probe from launching sink or split-KV kernels.
  
  yaomingamd
  and
  cursoragent
  committed
  96f66df
Commits on Oct 23, 2026
- Add the AMD copyright header on multi-head attention.
  
  yaomingamd
  and
  cursoragent
  committed
  ccfd30f
Commits on Oct 27, 2026
- Reuse the existing FP8 input check for the fused-attn output dtype.
  
  yaomingamd
  and
  cursoragent
  committed
  0e1ba6c

Commits

Commits on Sep 22, 2026

Commits on Oct 6, 2026

Commits on Oct 7, 2026

Commits on Oct 8, 2026

Commits on Oct 18, 2026

Commits on Oct 20, 2026

Commits on Oct 21, 2026

Commits on Oct 23, 2026

Commits on Oct 27, 2026