xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-09 11:45:59 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Perf] Increase default max splits for FA3 full cudagraphs (#25495)

Browse Source 
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
This commit is contained in:
Lucas Wilkinson 2025-09-23 19:53:34 -04:00 committed by GitHub
parent bde2a1a8a4
commit e0b24ea030
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 2 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
vllm/envs.py
View File

@ -119,7 +119,7 @@ if TYPE_CHECKING:
VLLM_SERVER_DEV_MODE: bool = False
VLLM_V1_OUTPUT_PROC_CHUNK_SIZE: int = 128
VLLM_MLA_DISABLE: bool = False
VLLM_FLASH_ATTN_MAX_NUM_SPLITS_FOR_CUDA_GRAPH: int = 16
VLLM_FLASH_ATTN_MAX_NUM_SPLITS_FOR_CUDA_GRAPH: int = 32
VLLM_RAY_PER_WORKER_GPUS: float = 1.0
VLLM_RAY_BUNDLE_INDICES: str = ""
VLLM_CUDART_SO_PATH: Optional[str] = None
@ -1017,7 +1017,7 @@ environment_variables: dict[str, Callable[[], Any]] = {
# max number splits for cuda graph decode
"VLLM_FLASH_ATTN_MAX_NUM_SPLITS_FOR_CUDA_GRAPH":
lambda: int(os.getenv("VLLM_FLASH_ATTN_MAX_NUM_SPLITS_FOR_CUDA_GRAPH",
"16")),
"32")),
# Number of GPUs per worker in Ray, if it is set to be a fraction,
# it allows ray to schedule multiple actors on a single GPU,