[ROCm] Add UE8M0 scale packing for Triton silu_mul_quant (#37835)

Signed-off-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
Andreas Karatzas
2026-08-19 10:20:05 -05:00
committed by GitHub
parent 160f7f0840
commit 17dbd42930
2 changed files with 44 additions and 13 deletions
@@ -239,12 +239,8 @@ def test_silu_mul_fp8_quant_deep_gemm(E: int, T: int, H: int, fp8_type: torch.dt
scale_fmts = [
DeepGemmQuantScaleFMT.FLOAT32,
DeepGemmQuantScaleFMT.FLOAT32_CEIL_UE8M0,
DeepGemmQuantScaleFMT.UE8M0,
]
# UE8M0 (int32 packed) scales require the C++ kernel which is
# not available on ROCm (#ifndef USE_ROCM).
# https://github.com/ROCm/aiter/issues/2420
if current_platform.is_cuda():
scale_fmts.append(DeepGemmQuantScaleFMT.UE8M0)
# Run the SiLU V2 kernel
for scale_fmt in scale_fmts: