Skip to content

vllm_omni.platforms.npu.graph_tools

Reusable exact-signature NPUGraph capture and replay helpers.

logger module-attribute 

logger = init_logger(__name__)

CapturedDeviceGraph dataclass

graph instance-attribute 

graph: _ReplayGraph

static_inputs instance-attribute 

static_inputs: tuple[Tensor, ...]

static_outputs instance-attribute 

static_outputs: tuple[Tensor, ...]

replay 

replay(inputs: tuple[Tensor, ...]) -> tuple[Tensor, ...]

NPUExactGraphRunner

Capture and replay tensor-only functions for exact NPU signatures.

component_name instance-attribute 

component_name = component_name

disable_config_hint instance-attribute 

disable_config_hint = disable_config_hint

max_graphs instance-attribute 

max_graphs = max(0, int(max_graphs))

stats property 

stats: dict[str, int]

capture 

capture(
    inputs: tuple[Tensor, ...],
    compute: Callable[..., tuple[Tensor, ...]],
) -> CapturedDeviceGraph

is_supported staticmethod 

is_supported() -> bool

run 

run(
    operation: str,
    inputs: tuple[Tensor, ...],
    constants: tuple[object, ...],
    compute: Callable[..., tuple[Tensor, ...]],
) -> tuple[Tensor, ...]