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

[Feature] Add VLLM_USE_DEEP_GEMM_E8M0 Env to Control E8M0 Scale (#21968)

Browse Source 
Signed-off-by: yewentao256 <zhyanwentao@126.com>
This commit is contained in:
Wentao Ye
2025-08-11 12:39:08 -04:00
committed by GitHub
parent 8e13d9fe6d
commit f7dcce7a4a
9 changed files with 65 additions and 39 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
vllm/envs.py
View File

@@ -127,6 +127,7 @@ if TYPE_CHECKING:
VLLM_TPU_MOST_MODEL_LEN: Optional[int] = None
VLLM_TPU_USING_PATHWAYS: bool = False
VLLM_USE_DEEP_GEMM: bool = False
VLLM_USE_DEEP_GEMM_E8M0: bool = True
VLLM_SKIP_DEEP_GEMM_WARMUP: bool = False
VLLM_USE_FLASHINFER_MOE_FP8: bool = False
VLLM_USE_FLASHINFER_MOE_FP4: bool = False
@@ -925,6 +926,10 @@ environment_variables: dict[str, Callable[[], Any]] = {
"VLLM_USE_DEEP_GEMM":
lambda: bool(int(os.getenv("VLLM_USE_DEEP_GEMM", "0"))),
# Whether to use E8M0 scaling when DeepGEMM is used on Blackwell GPUs.
# E8M0 is faster on B200 but may reduce accuracy.
"VLLM_USE_DEEP_GEMM_E8M0":
lambda: bool(int(os.getenv("VLLM_USE_DEEP_GEMM_E8M0", "1"))),
# DeepGemm JITs the kernels on-demand. The warmup attempts to make DeepGemm
# JIT all the required kernels before model execution so there is no
# JIT'ing in the hot-path. However, this warmup increases the engine