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
[Bug] Fix DeepGEMM Env Control (#23591)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Wentao Ye committed
Aug 25, 2025 at 21:41 UTC
56dcf4e7e965e34043acf20ca4e4aceda21d41ec
1 file changed
+4
-7
vllm/utils/deep_gemm.py
+4
-7
@@ -27,7 +27,7 @@ def is_deep_gemm_supported() -> bool:
27
is_supported_arch = current_platform.is_cuda() and (
28
current_platform.is_device_capability(90)
29
or current_platform.is_device_capability(100))
30
- return has_deep_gemm() and is_supported_arch
30
+ return envs.VLLM_USE_DEEP_GEMM and has_deep_gemm() and is_supported_arch
31
32
33
@functools.cache
@@ -35,12 +35,9 @@ def is_blackwell_deep_gemm_e8m0_used() -> bool:
35
"""Return ``True`` if vLLM is configured to use DeepGEMM "
36
"E8M0 scale on a Blackwell-class GPU.
37
"""
38
- if not (envs.VLLM_USE_DEEP_GEMM):
39
- logger.debug_once("DeepGEMM E8M0 disabled: VLLM_USE_DEEP_GEMM=0.")
40
- return False
41
-
42
- if not has_deep_gemm():
43
- logger.debug_once("DeepGEMM E8M0 disabled: DeepGEMM backend missing.")
38
+ if not is_deep_gemm_supported():
39
+ logger.debug_once(
40
+ "DeepGEMM E8M0 disabled: DeepGEMM not supported on this system.")
41
return False
42
43
if not envs.VLLM_USE_DEEP_GEMM_E8M0: