End-to-end CUDA graph annotation and profiling demo.
()
| 539 | # annotations, profile it, and post-process the trace. |
| 540 | |
| 541 | def main(): |
| 542 | """End-to-end CUDA graph annotation and profiling demo.""" |
| 543 | if not torch.cuda.is_available(): |
| 544 | raise SystemExit("CUDA required for this tutorial") |
| 545 | |
| 546 | # Check if annotation support is available |
| 547 | # PyTorch will log a warning if cuda-bindings version is too old |
| 548 | supported = not _is_tools_id_unavailable() |
| 549 | print(f"Annotation support available: {supported}") |
| 550 | if not supported: |
| 551 | print("NOTE: Annotation API not available.") |
| 552 | print("This could be due to:") |
| 553 | print(" - Driver/CUDA-compat < 13.1") |
| 554 | print(" - Outdated cuda-bindings (check PyTorch warnings above)") |
| 555 | print("Annotations will not be recorded, but the demo will still run.") |
| 556 | print("Kernels will be reassigned to the default lane, not semantic lanes.\n") |
| 557 | |
| 558 | output_dir = Path("traces") |
| 559 | |
| 560 | # Build the model |
| 561 | print("\n1. Building transformer block model...") |
| 562 | model_fn = build_transformer_block() |
| 563 | |
| 564 | # Capture graph with annotations |
| 565 | print("\n2. Capturing CUDA graph with annotations...") |
| 566 | graph, output = capture_graph_with_annotations(model_fn) |
| 567 | |
| 568 | # Save annotations |
| 569 | print("\n3. Saving annotation metadata...") |
| 570 | annotations_path = save_annotations(output_dir) |
| 571 | |
| 572 | # Profile the graph |
| 573 | print("\n4. Profiling graph replays...") |
| 574 | raw_trace_path = profile_graph(graph, output_dir) |
| 575 | |
| 576 | # Post-process the trace |
| 577 | print("\n5. Post-processing: merging annotations into trace...") |
| 578 | annotated_path, raw_trace, annotated_trace = post_process_trace( |
| 579 | raw_trace_path, annotations_path, output_dir |
| 580 | ) |
| 581 | |
| 582 | # Compare before and after |
| 583 | print("\n6. Comparing traces...") |
| 584 | compare_traces(raw_trace, annotated_trace) |
| 585 | |
| 586 | # Summary |
| 587 | print("\n" + "="*60) |
| 588 | print("SUMMARY") |
| 589 | print("="*60) |
| 590 | print(f"Raw trace: {raw_trace_path}") |
| 591 | print(f"Annotated trace: {annotated_path}") |
| 592 | print(f"Annotations: {annotations_path}") |
| 593 | print("\nOpen the annotated trace in https://ui.perfetto.dev/ to visualize") |
| 594 | print("the semantic kernel lanes.") |
| 595 | print("="*60) |
| 596 | |
| 597 | # Example output: |
| 598 | # if __name__ == "__main__": |
nothing calls this directly
no test coverage detected