xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-13 22:35:32 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Bugfix] DeepGEMM is not enabled on B200 due to _lazy_init() (#21472)

Browse Source 
Signed-off-by: Clayton Coleman <smarterclayton@gmail.com>
Co-authored-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
Clayton Coleman 2025-07-28 16:51:22 -04:00 committed by GitHub
parent b18b417fbf
commit c6f36cfa26
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 9 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
vllm/utils/deep_gemm.py
View File

@ -13,7 +13,8 @@ from typing import Any, Callable, NoReturn
import torch import torch
import vllm.envs as envs import vllm.envs as envs
from vllm.utils import cuda_get_device_properties, has_deep_gemm from vllm.platforms import current_platform
from vllm.utils import has_deep_gemm
@functools.cache @functools.cache
@ -21,12 +22,15 @@ def is_blackwell_deep_gemm_used() -> bool:
"""Return ``True`` if vLLM is configured to use DeepGEMM on a """Return ``True`` if vLLM is configured to use DeepGEMM on a
Blackwell-class GPU. Blackwell-class GPU.
""" """
if not (envs.VLLM_USE_DEEP_GEMM and has_deep_gemm()):
if not (envs.VLLM_USE_DEEP_GEMM and has_deep_gemm()
and _per_block_cast_impl is not None):
return False return False
return cuda_get_device_properties(0, ("major", ))[0] == 10 _lazy_init()
if _per_block_cast_impl is None:
return False
return (current_platform.is_cuda()
and current_platform.is_device_capability(100))
def _missing(*_: Any, **__: Any) -> NoReturn: def _missing(*_: Any, **__: Any) -> NoReturn: