xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-10 02:05:01 +08:00
Code Issues Packages Projects Releases Wiki Activity

[ROCm] Avoid using the default stream on ROCm (#13238)

Browse Source 
Signed-off-by: Gregory Shtrasberg <Gregory.Shtrasberg@amd.com>
This commit is contained in:
Gregory Shtrasberg 2025-02-13 20:29:26 -05:00 committed by GitHub
parent e38be640e6
commit 410886950a
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 6 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
vllm/utils.py
View File

@ -942,11 +942,16 @@ def current_stream() -> torch.cuda.Stream:
the underlying hypothesis is that we do not call `torch._C._cuda_setStream`
from C/C++ code.
"""
from vllm.platforms import current_platform
global _current_stream
if _current_stream is None:
# when this function is called before any stream is set,
# we return the default stream.
_current_stream = torch.cuda.current_stream()
# On ROCm using the default 0 stream in combination with RCCL
# is hurting performance. Therefore creating a dedicated stream
# per process
_current_stream = torch.cuda.Stream() if current_platform.is_rocm(
) else torch.cuda.current_stream()
return _current_stream