Skip to content

vllm.compilation.counter

compilation_counter module-attribute

compilation_counter = CompilationCounter()

CompilationCounter dataclass

Source code in vllm/compilation/counter.py
@dataclasses.dataclass
class CompilationCounter:
    num_models_seen: int = 0
    num_graphs_seen: int = 0
    # including the splitting ops
    num_piecewise_graphs_seen: int = 0
    # not including the splitting ops
    num_piecewise_capturable_graphs_seen: int = 0
    num_backend_compilations: int = 0
    # Number of gpu_model_runner attempts to trigger CUDAGraphs capture
    num_gpu_runner_capture_triggers: int = 0
    # Number of CUDAGraphs captured
    num_cudagraph_captured: int = 0
    # InductorAdapter.compile calls
    num_inductor_compiles: int = 0
    # EagerAdapter.compile calls
    num_eager_compiles: int = 0

    def clone(self) -> "CompilationCounter":
        return copy.deepcopy(self)

    @contextmanager
    def expect(self, **kwargs):
        old = self.clone()
        yield
        for k, v in kwargs.items():
            assert getattr(self, k) - getattr(old, k) == v, (
                f"{k} not as expected, before it is {getattr(old, k)}"
                f", after it is {getattr(self, k)}, "
                f"expected diff is {v}")

num_backend_compilations class-attribute instance-attribute

num_backend_compilations: int = 0

num_cudagraph_captured class-attribute instance-attribute

num_cudagraph_captured: int = 0

num_eager_compiles class-attribute instance-attribute

num_eager_compiles: int = 0

num_gpu_runner_capture_triggers class-attribute instance-attribute

num_gpu_runner_capture_triggers: int = 0

num_graphs_seen class-attribute instance-attribute

num_graphs_seen: int = 0

num_inductor_compiles class-attribute instance-attribute

num_inductor_compiles: int = 0

num_models_seen class-attribute instance-attribute

num_models_seen: int = 0

num_piecewise_capturable_graphs_seen class-attribute instance-attribute

num_piecewise_capturable_graphs_seen: int = 0

num_piecewise_graphs_seen class-attribute instance-attribute

num_piecewise_graphs_seen: int = 0

__init__

__init__(
    num_models_seen: int = 0,
    num_graphs_seen: int = 0,
    num_piecewise_graphs_seen: int = 0,
    num_piecewise_capturable_graphs_seen: int = 0,
    num_backend_compilations: int = 0,
    num_gpu_runner_capture_triggers: int = 0,
    num_cudagraph_captured: int = 0,
    num_inductor_compiles: int = 0,
    num_eager_compiles: int = 0,
) -> None

clone

clone() -> CompilationCounter
Source code in vllm/compilation/counter.py
def clone(self) -> "CompilationCounter":
    return copy.deepcopy(self)

expect

expect(**kwargs)
Source code in vllm/compilation/counter.py
@contextmanager
def expect(self, **kwargs):
    old = self.clone()
    yield
    for k, v in kwargs.items():
        assert getattr(self, k) - getattr(old, k) == v, (
            f"{k} not as expected, before it is {getattr(old, k)}"
            f", after it is {getattr(self, k)}, "
            f"expected diff is {v}")