fp8 tensor core support on h100

[Da Chen]

trition version: 2.3.0

It was mentioned in the issue #3156 that FP8 tensor core can already be used on h100. But when i ran the same example and use ncu to check mma instructions, I found the mma layout was 16816 which does not support fp8 as operands.

• ncu command: ncu --launch-count 1 --print-source=sass --page source --kernel-name regex:"_attn_fwd_0d1d2d34d5d6de7de8d*" "python3" triton_fused_attention.py
• output screenshot:
  

[Thomas Raoux]

you need to use nightly wheel or build top of tree triton from source.

[Da Chen]

it sees both nighly wheel and v2.2.0 can run fp8 tensor core correctly