log_operator_inputs() — pytorch Function Reference

Architecture documentation for the log_operator_inputs() function in common.py from the pytorch codebase.

Function python DynamoBenchmarks Microbenchmarking called by 1

Entity Profile

DynamoBenchmarks→ Microbenchmarking→ log_operator_inputs() — pytorch Function Reference

Dependency Diagram

graph TD
  67009ec0_1e64_0f0a_7e5d_9f07d2bbbdde["log_operator_inputs()"]
  c9be2096_e6d7_2374_ad2e_a6e33f435ada["run()"]
  c9be2096_e6d7_2374_ad2e_a6e33f435ada -->|calls| 67009ec0_1e64_0f0a_7e5d_9f07d2bbbdde
  style 67009ec0_1e64_0f0a_7e5d_9f07d2bbbdde fill:#6366f1,stroke:#818cf8,color:#fff

Relationship Graph

Source Code

benchmarks/dynamo/common.py lines 4572–4607

def log_operator_inputs(model, example_inputs, model_iter_fn, name, args):
    mode = "training" if args.training else "eval"
    output = os.path.join(os.path.dirname(args.output), f"{name}_{mode}.txt")

    # TODO - add option for coalescing inputs over multiple runs
    if os.path.exists(output):
        print(f"Skipping {name}, {output} already exists")
        return

    print(f"Running {name}")
    try:
        from .microbenchmarks.operator_inp_utils import OperatorInputsMode
    except ImportError:
        from microbenchmarks.operator_inp_utils import OperatorInputsMode

    operator_mode = OperatorInputsMode()
    fake_tensor_mode = FakeTensorMode()

    with torch._subclasses.fake_tensor.FakeCopyMode(fake_tensor_mode):
        model_fake = copy.deepcopy(model)
        example_inputs_fake = copy.deepcopy(example_inputs)
    try:
        with fake_tensor_mode, operator_mode:
            model_iter_fn(model_fake, example_inputs_fake, collect_outputs=False)
    except Exception as e:
        print(f"{name} failed to run with fake tensors, trying real. Exception: {e}")
        operator_mode = OperatorInputsMode()
        try:
            with operator_mode:
                model_iter_fn(model, example_inputs, collect_outputs=False)
        except Exception as e2:
            print(f"{name} failed to run with real. Exception: {e2}")
            raise

    print(f"Writing output to {output}")
    operator_mode.log_to_file(output)