xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-01-08 04:18:43 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Kernels] Enable DeepGEMM by default (#24462)

Browse Source 
Signed-off-by: Bill Nell <bnell@redhat.com>
This commit is contained in:
bnellnm 2025-09-17 23:19:52 -04:00 committed by GitHub
parent 7fb2a5be28
commit 4ac510f484
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 2 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
vllm/envs.py
View File

@ -135,7 +135,7 @@ if TYPE_CHECKING:
VLLM_TPU_BUCKET_PADDING_GAP: int = 0
VLLM_TPU_MOST_MODEL_LEN: Optional[int] = None
VLLM_TPU_USING_PATHWAYS: bool = False
VLLM_USE_DEEP_GEMM: bool = False
VLLM_USE_DEEP_GEMM: bool = True
VLLM_USE_DEEP_GEMM_E8M0: bool = True
VLLM_USE_DEEP_GEMM_E8M0_HOPPER: bool = False
VLLM_SKIP_DEEP_GEMM_WARMUP: bool = False
@ -1044,7 +1044,7 @@ environment_variables: dict[str, Callable[[], Any]] = {
# Allow use of DeepGemm kernels for fused moe ops.
"VLLM_USE_DEEP_GEMM":
lambda: bool(int(os.getenv("VLLM_USE_DEEP_GEMM", "0"))),
lambda: bool(int(os.getenv("VLLM_USE_DEEP_GEMM", "1"))),
# Whether to use E8M0 scaling when DeepGEMM is used on Blackwell GPUs.
"VLLM_USE_DEEP_GEMM_E8M0":