[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
+6 -14
View File
@@ -147,26 +147,22 @@ def test_fp4_env_defaults():
"""ROCm FP4 env defaults should stay stable for the AITER gates."""
import vllm.envs as envs
assert envs.VLLM_ROCM_USE_AITER_FP4_ASM_GEMM is False
assert envs.VLLM_ROCM_USE_AITER_FP4BMM is True
@pytest.mark.parametrize(
(
"use_aiter",
"use_fp4_asm_gemm",
"use_fp4bmm",
),
[
(True, True, True),
(True, True, False),
(True, False, True),
(False, True, True),
(True, True),
(True, False),
(False, True),
],
)
def test_rocm_aiter_fp4_enablement_follows_env_and_arch(
use_aiter,
use_fp4_asm_gemm,
use_fp4bmm,
monkeypatch,
):
@@ -182,15 +178,11 @@ def test_rocm_aiter_fp4_enablement_follows_env_and_arch(
_assert_aiter_supported()
on_gfx950_value = on_gfx950()
expected_asm_gemm = use_aiter and use_fp4_asm_gemm and on_gfx950_value
expected_asm_gemm = use_aiter and on_gfx950_value
expected_fp4bmm = use_aiter and use_fp4bmm and on_gfx950_value
with monkeypatch.context() as mp:
mp.setenv("VLLM_ROCM_USE_AITER", "1" if use_aiter else "0")
mp.setenv(
"VLLM_ROCM_USE_AITER_FP4_ASM_GEMM",
"1" if use_fp4_asm_gemm else "0",
)
mp.setenv("VLLM_ROCM_USE_AITER_FP4BMM", "1" if use_fp4bmm else "0")
_reload_envs()
rocm_aiter_ops.refresh_env_variables()
@@ -410,7 +402,7 @@ def test_aiter_fp4_gemm_preshuffled_tuned_shapes(shape):
from aiter import per_1x32_f4_quant_hip
from aiter.ops.shuffle import shuffle_weight
from aiter.ops.triton.gemm_afp4wfp4 import (
gemm_afp4wfp4_preshuffled_weight_scales,
gemm_afp4wfp4_preshuffle,
)
from aiter.ops.triton.quant import dynamic_mxfp4_quant
@@ -446,7 +438,7 @@ def test_aiter_fp4_gemm_preshuffled_tuned_shapes(shape):
def run_preshuffled() -> torch.Tensor:
y = torch.empty(M, N, device="cuda", dtype=torch.bfloat16)
return gemm_afp4wfp4_preshuffled_weight_scales(
return gemm_afp4wfp4_preshuffle(
A_q.contiguous().view(torch.uint8),
B_fp4.contiguous().view(torch.uint8).reshape(B_fp4.shape[0] // 16, -1),
A_s,