Transforms complex documents like PDFs and Office docs into LLM-ready markdown/JSON for your Agentic workflows.
Python ai4sciencedocument-analysisdocxextract-datalayout-analysisocrparserpdfpdf-converterpdf-extractor-llmpdf-extractor-pretrainpdf-extractor-ragpdf-parserpptxpythonxlsx
"""验证 GPU 验收入口的报告与退出码一致,不加载模型或替代真机验收。"""
from __future__ import annotations
import argparse
import json
import sys
from pathlib import Path
from types import SimpleNamespace
from typing import Any
import pytest
from scripts.acceptance import lmdeploy_lifecycle
@pytest.mark.parametrize("engine,platform", [("lmdeploy", "win32"), ("vllm", "linux")])
@pytest.mark.parametrize("scenario", ["success", "execution_failure", "cleanup_failure", "incomplete"])
def test_acceptance_report_matches_exit_status(
engine: str,
platform: str,
scenario: str,
tmp_path: Path,
monkeypatch: pytest.MonkeyPatch,
capsys: pytest.CaptureFixture[str],
) -> None:
"""清理失败必须撤销先前的成功标记,未完成全部验收也不能因正常返回而变为通过。"""
sample = tmp_path / "sample.pdf"
sample.write_bytes(b"%PDF-1.7\n")
evidence = tmp_path / "evidence"
monkeypatch.setattr(sys, "argv", ["acceptance", "--sample", str(sample), "--evidence", str(evidence)])
# 只替换脚本自身的平台视图,避免影响 macOS 上 asyncio 的事件循环选择。
monkeypatch.setattr(lmdeploy_lifecycle, "sys", SimpleNamespace(platform=platform, stderr=sys.stderr))
monkeypatch.setenv("PYTHONUTF8", "1")
monkeypatch.setenv("MINERU_LMDEPLOY_BACKEND", "turbomind")
monkeypatch.setattr(lmdeploy_lifecycle, "_versions", lambda engine: {"synthetic": True})
cleanup_calls = []
class Pool:
"""提供可控关闭故障,保持验收执行完成后才清理的时序。"""
async def close(self) -> None:
"""分别模拟正常关闭与 Windows 上传目录清理失败。"""
cleanup_calls.append("close")
if scenario == "cleanup_failure":
raise PermissionError("synthetic upload directory cleanup failure")
async def accept(args: argparse.Namespace, report: dict[str, Any]) -> None:
"""隔离 GPU 操作,保留成功记录、执行异常与 finally 关闭的入口契约。"""
assert args.engine == engine
pool = Pool()
try:
if scenario == "execution_failure":
raise RuntimeError("synthetic inference failure")
report["inflight_task_terminal_states_observed"] = scenario != "incomplete"
report["passed"] = report["inflight_task_terminal_states_observed"]
finally:
await pool.close()
monkeypatch.setattr(lmdeploy_lifecycle, "_accept", accept)
exit_code = lmdeploy_lifecycle.main(engine)
report = json.loads((evidence / "report.json").read_text(encoding="utf-8"))
captured = capsys.readouterr()
expected_pass = scenario == "success"
assert report["passed"] is expected_pass
assert exit_code == (0 if expected_pass else 1)
assert report["engine"] == engine
assert cleanup_calls == ["close"]
if scenario in {"execution_failure", "cleanup_failure"}:
error = "PermissionError" if scenario == "cleanup_failure" else "RuntimeError"
assert error in report["error"]
assert "Traceback" in report["error"]
assert report["error"] in captured.err
else:
assert "error" not in report
assert not captured.err