Profile the model and return a list of KernelRecords sorted by GPU time desc. Returns: Tuple of (records, extras) where extras contains paths to exported artifacts and optional HTA analysis results.
(
model: nn.Module,
inputs: Dict[str, Any],
warmup_iters: int = WARMUP_ITERS,
profile_iters: int = PROFILE_ITERS,
export_trace: bool = False,
memory_snapshot: bool = False,
)
| 527 | |
| 528 | |
| 529 | def profile_model( |
| 530 | model: nn.Module, |
| 531 | inputs: Dict[str, Any], |
| 532 | warmup_iters: int = WARMUP_ITERS, |
| 533 | profile_iters: int = PROFILE_ITERS, |
| 534 | export_trace: bool = False, |
| 535 | memory_snapshot: bool = False, |
| 536 | ) -> Tuple[List[KernelRecord], Dict[str, Any]]: |
| 537 | """Profile the model and return a list of KernelRecords sorted by GPU time desc. |
| 538 | |
| 539 | Returns: |
| 540 | Tuple of (records, extras) where extras contains paths to exported |
| 541 | artifacts and optional HTA analysis results. |
| 542 | """ |
| 543 | extras: Dict[str, Any] = {} |
| 544 | |
| 545 | os.makedirs(WORKSPACE_DIR, exist_ok=True) |
| 546 | trace_path = os.path.join(WORKSPACE_DIR, "trace.json") |
| 547 | snapshot_path = os.path.join(WORKSPACE_DIR, "memory_snapshot.pickle") |
| 548 | |
| 549 | # --- Warmup --- |
| 550 | with torch.no_grad(): |
| 551 | for _ in range(warmup_iters): |
| 552 | _run_forward(model, inputs) |
| 553 | |
| 554 | torch.cuda.synchronize() |
| 555 | |
| 556 | # --- Start memory recording if requested --- |
| 557 | if memory_snapshot: |
| 558 | try: |
| 559 | torch.cuda.memory._record_memory_history(max_entries=100000) |
| 560 | except Exception as e: |
| 561 | print(f" WARNING: Could not start memory history recording: {e}") |
| 562 | memory_snapshot = False |
| 563 | |
| 564 | # --- Profile --- |
| 565 | with torch.no_grad(): |
| 566 | with torch.profiler.profile( |
| 567 | activities=[ |
| 568 | torch.profiler.ProfilerActivity.CPU, |
| 569 | torch.profiler.ProfilerActivity.CUDA, |
| 570 | ], |
| 571 | record_shapes=True, |
| 572 | with_stack=False, |
| 573 | ) as prof: |
| 574 | for _ in range(profile_iters): |
| 575 | _run_forward(model, inputs) |
| 576 | torch.cuda.synchronize() |
| 577 | |
| 578 | # --- Export Chrome trace --- |
| 579 | if export_trace: |
| 580 | try: |
| 581 | prof.export_chrome_trace(trace_path) |
| 582 | extras["trace_path"] = trace_path |
| 583 | except Exception as e: |
| 584 | print(f" WARNING: Could not export Chrome trace: {e}") |
| 585 | |
| 586 | # --- Capture memory snapshot --- |
no test coverage detected