xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-05-23 07:11:19 +08:00
Code Issues Packages Projects Releases Wiki Activity

Remove global CUDA graph pool

Browse Source 
Signed-off-by: Woosuk Kwon <woosuk@thinkingmachines.ai>
This commit is contained in:
Woosuk Kwon 2025-09-19 23:27:51 +00:00
parent ee7a66dd9a
commit 20098c10d9
3 changed files with 3 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
vllm/compilation/cuda_graph.py
View File

@ -82,7 +82,7 @@ class CUDAGraphWrapper:
# TODO: in the future, if we want to use multiple # TODO: in the future, if we want to use multiple
# streams, it might not be safe to share a global pool. # streams, it might not be safe to share a global pool.
# only investigate this when we use multiple streams # only investigate this when we use multiple streams
self.graph_pool = current_platform.get_global_graph_pool() self.graph_pool = current_platform.graph_pool_handle()
if cudagraph_options is None: if cudagraph_options is None:
cudagraph_options = CUDAGraphOptions() cudagraph_options = CUDAGraphOptions()

7
vllm/platforms/interface.py
View File

@ -140,8 +140,6 @@ class Platform:
additional_env_vars: list[str] = [] additional_env_vars: list[str] = []
_global_graph_pool: Optional[Any] = None
@property @property
def supported_dtypes(self) -> list[torch.dtype]: def supported_dtypes(self) -> list[torch.dtype]:
"""Returns the supported dtypes for the current platform.""" """Returns the supported dtypes for the current platform."""
@ -539,10 +537,7 @@ class Platform:
""" """
Return the global graph pool for this platform. Return the global graph pool for this platform.
""" """
cls = self.__class__ return self.graph_pool_handle()
if cls._global_graph_pool is None:
cls._global_graph_pool = self.graph_pool_handle()
return cls._global_graph_pool
@classmethod @classmethod
def get_cu_count(cls, device_id: int = 0) -> int: def get_cu_count(cls, device_id: int = 0) -> int:

2
vllm/v1/worker/gpu_ubatch_wrapper.py
View File

@ -54,7 +54,7 @@ class UBatchWrapper:
if runtime_mode is not CUDAGraphMode.NONE: if runtime_mode is not CUDAGraphMode.NONE:
self.cudagraph_wrapper = CUDAGraphWrapper( self.cudagraph_wrapper = CUDAGraphWrapper(
runnable, vllm_config, runtime_mode=runtime_mode) runnable, vllm_config, runtime_mode=runtime_mode)
self.graph_pool = current_platform.get_global_graph_pool() self.graph_pool = current_platform.graph_pool_handle()
def __getattr__(self, key: str): def __getattr__(self, key: str):
# allow accessing the attributes of the runnable. # allow accessing the attributes of the runnable.