[Experimental] Breakable CUDA graph (#42304)
Signed-off-by: zjy0516 <riverclouds.zhu@qq.com>
This commit is contained in:
@@ -145,6 +145,7 @@ if TYPE_CHECKING:
|
||||
VLLM_DP_SIZE: int = 1
|
||||
VLLM_USE_STANDALONE_COMPILE: bool = True
|
||||
VLLM_ENABLE_PREGRAD_PASSES: bool = True
|
||||
VLLM_USE_BREAKABLE_CUDAGRAPH: bool = False
|
||||
VLLM_DP_MASTER_IP: str = ""
|
||||
VLLM_DP_MASTER_PORT: int = 0
|
||||
VLLM_RANDOMIZE_DP_DUMMY_INPUTS: bool = False
|
||||
@@ -637,6 +638,10 @@ environment_variables: dict[str, Callable[[], Any]] = {
|
||||
"VLLM_ENABLE_PREGRAD_PASSES": lambda: (
|
||||
os.environ.get("VLLM_ENABLE_PREGRAD_PASSES", "1") == "1"
|
||||
),
|
||||
# Experimental: breakable cudagraph does not rely on torch.compile
|
||||
"VLLM_USE_BREAKABLE_CUDAGRAPH": lambda: (
|
||||
os.environ.get("VLLM_USE_BREAKABLE_CUDAGRAPH", "0") == "1"
|
||||
),
|
||||
# Debug pattern matching inside custom passes.
|
||||
# Should be set to the fx.Node name (e.g. 'getitem_34' or 'scaled_mm_3').
|
||||
"VLLM_PATTERN_MATCH_DEBUG": lambda: os.environ.get(
|
||||
|
||||
Reference in New Issue
Block a user