xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-10 20:35:01 +08:00
Code Issues Packages Projects Releases Wiki Activity

allow disable flashinfer prefill (#25276)

Browse Source 
Signed-off-by: Lu Fang <fanglu@fb.com>
This commit is contained in:
Lucia Fang 2025-09-19 15:59:41 -07:00 committed by GitHub
parent 431535b522
commit ee7a66dd9a
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 5 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
vllm/envs.py
View File

@ -32,6 +32,7 @@ if TYPE_CHECKING:
VLLM_CONFIG_ROOT: str = os.path.expanduser("~/.config/vllm") VLLM_CONFIG_ROOT: str = os.path.expanduser("~/.config/vllm")
VLLM_USAGE_STATS_SERVER: str = "https://stats.vllm.ai" VLLM_USAGE_STATS_SERVER: str = "https://stats.vllm.ai"
VLLM_NO_USAGE_STATS: bool = False VLLM_NO_USAGE_STATS: bool = False
VLLM_DISABLE_FLASHINFER_PREFILL: bool = False
VLLM_DO_NOT_TRACK: bool = False VLLM_DO_NOT_TRACK: bool = False
VLLM_USAGE_SOURCE: str = "" VLLM_USAGE_SOURCE: str = ""
VLLM_CONFIGURE_LOGGING: int = 1 VLLM_CONFIGURE_LOGGING: int = 1
@ -479,6 +480,8 @@ environment_variables: dict[str, Callable[[], Any]] = {
lambda: os.environ.get("VLLM_USAGE_STATS_SERVER", "https://stats.vllm.ai"), lambda: os.environ.get("VLLM_USAGE_STATS_SERVER", "https://stats.vllm.ai"),
"VLLM_NO_USAGE_STATS": "VLLM_NO_USAGE_STATS":
lambda: os.environ.get("VLLM_NO_USAGE_STATS", "0") == "1", lambda: os.environ.get("VLLM_NO_USAGE_STATS", "0") == "1",
"VLLM_DISABLE_FLASHINFER_PREFILL":
lambda: os.environ.get("VLLM_DISABLE_FLASHINFER_PREFILL", "0") == "1",
"VLLM_DO_NOT_TRACK": "VLLM_DO_NOT_TRACK":
lambda: (os.environ.get("VLLM_DO_NOT_TRACK", None) or os.environ.get( lambda: (os.environ.get("VLLM_DO_NOT_TRACK", None) or os.environ.get(
"DO_NOT_TRACK", None) or "0") == "1", "DO_NOT_TRACK", None) or "0") == "1",

3
vllm/v1/attention/backends/mla/common.py
View File

@ -412,7 +412,8 @@ M = TypeVar("M", bound=MLACommonMetadata)
def use_flashinfer_prefill() -> bool: def use_flashinfer_prefill() -> bool:
# For blackwell default to flashinfer prefill if it's available since # For blackwell default to flashinfer prefill if it's available since
# it is faster than FA2. # it is faster than FA2.
return (flashinfer_available and not envs.VLLM_USE_CUDNN_PREFILL return (not envs.VLLM_DISABLE_FLASHINFER_PREFILL and flashinfer_available
and not envs.VLLM_USE_CUDNN_PREFILL
and current_platform.is_device_capability(100)) and current_platform.is_device_capability(100))