[ROCm] Restore VLLM_ROCM_USE_AITER_FP4_ASM_GEMM and default w4a4 ASM GEMM back to off (#57055)

Signed-off-by: Andy Friedrich <afriedri@amd.com>
This commit is contained in:
afriedri
2026-09-17 13:39:52 -05:00
committed by GitHub
parent 6a344d8c85
commit a9a7e45f31
3 changed files with 23 additions and 5 deletions
+6
View File
@@ -143,6 +143,7 @@ if TYPE_CHECKING:
VLLM_ROCM_USE_AITER_MLA: bool = True
VLLM_ROCM_AITER_MLA_ASM_PADDING: Literal["auto", "gluon", "asm"] = "auto"
VLLM_ROCM_USE_AITER_MHA: bool = True
VLLM_ROCM_USE_AITER_FP4_ASM_GEMM: bool = False
VLLM_ROCM_USE_AITER_TRITON_ROPE: bool = False
VLLM_ROCM_USE_AITER_FP8BMM: bool = True
VLLM_ROCM_USE_AITER_FP4BMM: bool = True
@@ -1305,6 +1306,11 @@ environment_variables: dict[str, Callable[[], Any]] = {
"VLLM_ROCM_USE_AITER_MHA": lambda: (
os.getenv("VLLM_ROCM_USE_AITER_MHA", "True").lower() in ("true", "1")
),
# Whether to use aiter fp4 gemm asm.
# By default is disabled.
"VLLM_ROCM_USE_AITER_FP4_ASM_GEMM": lambda: (
os.getenv("VLLM_ROCM_USE_AITER_FP4_ASM_GEMM", "False").lower() in ("true", "1")
),
# Whether to use aiter rope.
# By default is disabled.
"VLLM_ROCM_USE_AITER_TRITON_ROPE": lambda: (