biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Perf] Enable FlashInfer DeepGEMM swapAB on SM90 by default (#34924)

Browse Source 
Signed-off-by: mgoin <mgoin64@gmail.com>
Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
This commit is contained in:
Michael Goin
2026-02-23 20:34:41 -05:00
committed by GitHub
parent 3ef9fd0f98
commit a4bd661fb3
2 changed files with 7 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
vllm/envs.py
View File

@@ -159,7 +159,7 @@ if TYPE_CHECKING:
"relax",
] = "relax"
VLLM_USE_FUSED_MOE_GROUPED_TOPK: bool = True
VLLM_BLOCKSCALE_FP8_GEMM_FLASHINFER: bool = False
VLLM_BLOCKSCALE_FP8_GEMM_FLASHINFER: bool = True
VLLM_USE_FLASHINFER_MOE_FP16: bool = False
VLLM_USE_FLASHINFER_MOE_FP8: bool = False
VLLM_USE_FLASHINFER_MOE_FP4: bool = False
@@ -1198,7 +1198,7 @@ environment_variables: dict[str, Callable[[], Any]] = {
# Allow use of FlashInfer FP8 block-scale GEMM for linear layers.
# This uses TensorRT-LLM kernels and requires SM90+ (Hopper).
"VLLM_BLOCKSCALE_FP8_GEMM_FLASHINFER": lambda: bool(
int(os.getenv("VLLM_BLOCKSCALE_FP8_GEMM_FLASHINFER", "0"))
int(os.getenv("VLLM_BLOCKSCALE_FP8_GEMM_FLASHINFER", "1"))
),
# Allow use of FlashInfer BF16 MoE kernels for fused moe ops.
"VLLM_USE_FLASHINFER_MOE_FP16": lambda: bool(