triton_perf_analyzer_python_api_model_profiling_quickstart.py

1import triton_perf_analyzer.wrapper as pa
2
3def main():
4    # Model name to profile
5    model_name = "add_sub"
6    
7    # Run Perf Analyzer with basic arguments
8    # Equivalent to CLI: perf_analyzer -m add_sub --concurrency-range 1:4
9    cmd = f"-m {model_name} --concurrency-range 1:4"
10    
11    print(f"Running Perf Analyzer for model: {model_name}...")
12    
13    # Execute the performance profiling
14    # The wrapper returns a result object containing the metrics
15    result = pa.run(cmd)
16
17    # Check if the run was successful
18    if result.status() == 0:
19        print("Profiling complete!")
20        # Get the metrics from the last measurement
21        # This returns a dictionary of metrics like throughput and latency
22        metrics = result.get_metrics()
23        print(f"Final Throughput: {metrics['throughput_infer_per_sec']} infer/sec")
24        print(f"Avg Latency: {metrics['avg_latency_us']} us")
25    else:
26        print(f"Profiling failed with status: {result.status()}")
27        print(f"Error output: {result.stderr()}")
28
29if __name__ == "__main__":
30    main()