A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
allow disable flashinfer prefill (#25276)
Signed-off-by: Lu Fang <fanglu@fb.com>
Lucia Fang committed
Sep 19, 2025 at 15:59 UTC
ee7a66dd9a5ead46f062502af33766f45076f05d
2 files changed
+5
-1
vllm/envs.py
+3
@@ -32,6 +32,7 @@ if TYPE_CHECKING:
32
VLLM_CONFIG_ROOT: str = os.path.expanduser("~/.config/vllm")
33
VLLM_USAGE_STATS_SERVER: str = "https://stats.vllm.ai"
34
VLLM_NO_USAGE_STATS: bool = False
35
+ VLLM_DISABLE_FLASHINFER_PREFILL: bool = False
36
VLLM_DO_NOT_TRACK: bool = False
37
VLLM_USAGE_SOURCE: str = ""
38
VLLM_CONFIGURE_LOGGING: int = 1
@@ -479,6 +480,8 @@ environment_variables: dict[str, Callable[[], Any]] = {
480
lambda: os.environ.get("VLLM_USAGE_STATS_SERVER", "https://stats.vllm.ai"),
481
"VLLM_NO_USAGE_STATS":
482
lambda: os.environ.get("VLLM_NO_USAGE_STATS", "0") == "1",
483
+ "VLLM_DISABLE_FLASHINFER_PREFILL":
484
+ lambda: os.environ.get("VLLM_DISABLE_FLASHINFER_PREFILL", "0") == "1",
485
"VLLM_DO_NOT_TRACK":
486
lambda: (os.environ.get("VLLM_DO_NOT_TRACK", None) or os.environ.get(
487
"DO_NOT_TRACK", None) or "0") == "1",
vllm/v1/attention/backends/mla/common.py
+2
-1
@@ -412,7 +412,8 @@ M = TypeVar("M", bound=MLACommonMetadata)
412
def use_flashinfer_prefill() -> bool:
413
# For blackwell default to flashinfer prefill if it's available since
414
# it is faster than FA2.
415
- return (flashinfer_available and not envs.VLLM_USE_CUDNN_PREFILL
415
+ return (not envs.VLLM_DISABLE_FLASHINFER_PREFILL and flashinfer_available
416
+ and not envs.VLLM_USE_CUDNN_PREFILL
417
and current_platform.is_device_capability(100))
418
419