A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[Kernels] Enable DeepGEMM by default (#24462)
Signed-off-by: Bill Nell <bnell@redhat.com>
bnellnm committed
Sep 17, 2025 at 23:19 UTC
4ac510f4844ae2ab168c2dbac545e3dd28a0a1b9
1 file changed
+2
-2
vllm/envs.py
+2
-2
@@ -135,7 +135,7 @@ if TYPE_CHECKING:
135
VLLM_TPU_BUCKET_PADDING_GAP: int = 0
136
VLLM_TPU_MOST_MODEL_LEN: Optional[int] = None
137
VLLM_TPU_USING_PATHWAYS: bool = False
138
- VLLM_USE_DEEP_GEMM: bool = False
138
+ VLLM_USE_DEEP_GEMM: bool = True
139
VLLM_USE_DEEP_GEMM_E8M0: bool = True
140
VLLM_USE_DEEP_GEMM_E8M0_HOPPER: bool = False
141
VLLM_SKIP_DEEP_GEMM_WARMUP: bool = False
@@ -1044,7 +1044,7 @@ environment_variables: dict[str, Callable[[], Any]] = {
1044
1045
# Allow use of DeepGemm kernels for fused moe ops.
1046
"VLLM_USE_DEEP_GEMM":
1047
- lambda: bool(int(os.getenv("VLLM_USE_DEEP_GEMM", "0"))),
1047
+ lambda: bool(int(os.getenv("VLLM_USE_DEEP_GEMM", "1"))),
1048
1049
# Whether to use E8M0 scaling when DeepGEMM is used on Blackwell GPUs.
1050
"VLLM_USE_DEEP_GEMM_E8M0":