Function profile_single_run

tests/quick_profile.py:18–104

Profile a single inference run.

()

def profile_single_run():
    """Profile a single inference run."""
    from questions.inference_server.model_cache import ModelCache
    from questions.models import GenerateParams
    from questions.text_generator_inference import fast_inference

    model_cache = ModelCache()
    params = GenerateParams(
        text="Once upon a time",
        max_length=20,  # Short generation for profiling
        temperature=0.7,
        top_p=0.9,
        top_k=40,
        number_of_results=1,
        model="any",
    )

    # Single warmup
    print("Warmup...")
    fast_inference(params, model_cache)
    if torch.cuda.is_available():
        torch.cuda.synchronize()

    gc.collect()
    if torch.cuda.is_available():
        torch.cuda.empty_cache()

    # Profile single run
    print("Profiling single inference run...")
    profiler = cProfile.Profile()
    profiler.enable()

    result = fast_inference(params, model_cache)

    if torch.cuda.is_available():
        torch.cuda.synchronize()

    profiler.disable()

    # Print stats
    stream = io.StringIO()
    stats = pstats.Stats(profiler, stream=stream)
    stats.strip_dirs()
    stats.sort_stats("cumulative")

    print("\n" + "="*60)
    print("TOP 25 FUNCTIONS BY CUMULATIVE TIME")
    print("="*60 + "\n")

    stats.print_stats(25)
    print(stream.getvalue())

    # Component analysis
    print("\n" + "="*60)
    print("TIME BY COMPONENT")
    print("="*60)

    component_times = {