| """Pytest plugin: per-test wall time (setup/call/teardown) and peak CUDA memory, written as JSONL. |
| |
| Load with `-p profile_plugin` (directory on PYTHONPATH). Output path: $PROFILE_OUT. Extra metadata copied into every |
| row: $PROFILE_META (JSON object). Peak memory covers the PyTorch allocator of the pytest process only (not vLLM servers |
| or `accelerate launch` subprocesses). |
| """ |
|
|
| import json |
| import os |
|
|
| import pytest |
| import torch |
|
|
|
|
| _rows = [] |
| _meta = json.loads(os.environ.get("PROFILE_META", "{}")) |
|
|
|
|
| @pytest.hookimpl(hookwrapper=True) |
| def pytest_runtest_protocol(item, nextitem): |
| row = {"nodeid": item.nodeid, "file": item.nodeid.split("::")[0], "outcome": "passed"} |
| item._profile_row = row |
| torch.cuda.reset_peak_memory_stats() |
| yield |
| row["peak_alloc_mib"] = torch.cuda.max_memory_allocated() / 2**20 |
| row["peak_reserved_mib"] = torch.cuda.max_memory_reserved() / 2**20 |
| row["total_s"] = sum(row.get(f"{when}_s", 0.0) for when in ("setup", "call", "teardown")) |
| _rows.append({**_meta, **row}) |
|
|
|
|
| @pytest.hookimpl(hookwrapper=True) |
| def pytest_runtest_makereport(item, call): |
| outcome = yield |
| report = outcome.get_result() |
| row = item._profile_row |
| row[f"{report.when}_s"] = report.duration |
| if report.outcome != "passed" and row["outcome"] == "passed": |
| row["outcome"] = report.outcome if report.when == "call" or report.skipped else "error" |
|
|
|
|
| def pytest_sessionfinish(session, exitstatus): |
| with open(os.environ["PROFILE_OUT"], "w") as f: |
| for row in _rows: |
| f.write(json.dumps(row) + "\n") |
|
|
