[ROCm] remove VLLM_ROCM_USE_AITER_FP4_ASM_GEMM environment variable; make w4a4 use the preshuffle triton+asm by default (#53141)

Signed-off-by: Andy Friedrich <afriedri@amd.com>
Signed-off-by: afriedri <afriedri@amd.com>
Co-authored-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
afriedri
2026-08-28 18:59:53 +00:00
committed by GitHub
co-authored by Andreas Karatzas
parent 31c579503d
commit 2f1cba799e
4 changed files with 9 additions and 27 deletions
@@ -35,7 +35,7 @@ if is_aiter_found_and_supported():
) -> torch.Tensor:
from aiter.ops.triton.gemm_afp4wfp4 import (
gemm_afp4wfp4,
gemm_afp4wfp4_preshuffled_weight_scales,
gemm_afp4wfp4_preshuffle,
)
from aiter.ops.triton.quant import dynamic_mxfp4_quant
@@ -63,7 +63,7 @@ if is_aiter_found_and_supported():
x_s = x_s[:M, ...].view(torch.uint8)
y = torch.empty(M, N, device=x_q.device, dtype=out_dtype)
gemm_afp4wfp4_preshuffled_weight_scales(
gemm_afp4wfp4_preshuffle(
x_q.view(torch.uint8),
weight.view(torch.uint8).view(weight.shape[0] // 16, -1),
x_s,