compute_speedups() — pytorch Function Reference

Architecture documentation for the compute_speedups() function in microbench.py from the pytorch codebase.

Function python DynamoBenchmarks Configuration called by 1

Entity Profile

DynamoBenchmarks→ Configuration→ compute_speedups() — pytorch Function Reference

Dependency Diagram

graph TD
  0c913e59_9191_0190_ce44_ba814a18cb93["compute_speedups()"]
  f3575d6f_d852_9193_8d21_437bde29bfae["microbenchmark()"]
  f3575d6f_d852_9193_8d21_437bde29bfae -->|calls| 0c913e59_9191_0190_ce44_ba814a18cb93
  style 0c913e59_9191_0190_ce44_ba814a18cb93 fill:#6366f1,stroke:#818cf8,color:#fff

Relationship Graph

Source Code

benchmarks/dynamo/microbenchmarks/microbench.py lines 26–39

def compute_speedups(args, models, example_inputs):
    expected = models[0](*example_inputs)
    for model in models[1:]:
        actual = model(*example_inputs)
        if not same(actual, expected):
            raise AssertionError(f"Output mismatch: diff={expected[0] - actual[0]}")

    timings = np.zeros((args.repeat, len(models)), np.float64)
    for rep in range(args.repeat):
        # interleave the runs to handle frequency scaling and load changes
        for m, model in enumerate(models):
            timings[rep, m] = timed(model, example_inputs)
    median = np.median(timings, axis=0)
    return (median[0] / median[1:]).tolist()