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
"""原生独显容量与保留的硬件诊断 ABI 回归。"""
from __future__ import annotations
import ctypes as ct
from types import SimpleNamespace
from typing import Any
from unittest.mock import Mock
import pytest
from mineru.model.runtime import vulkan, windows_gpu
from mineru.model.runtime.gpu import PciAddress
@pytest.mark.parametrize("stage", ["success", "open", "query", "invalid", "close"])
def test_windows_pci_query_abi_and_handle_lifetime(monkeypatch: pytest.MonkeyPatch, stage: str) -> None:
"""验证系统接口结构布局,失败/无效地址均不伪造坐标,已打开句柄始终关闭。"""
def open_adapter(output: Any) -> int:
assert output._obj.luid.low == 0x1709C and output._obj.luid.high == 1
output._obj.handle = 123
return -1 if stage == "open" else 0
def query_adapter(output: Any) -> int:
query = output._obj
assert query.handle == 123 and query.type == 6 and query.size == 12
address = ct.cast(query.data, ct.POINTER(windows_gpu._AdapterAddress)).contents
address.bus, address.device, address.function = (0xFFFFFFFF, 0xFFFF, 0xFFFF) if stage == "invalid" else (195, 0, 1)
return -1 if stage == "query" else 0
def close_adapter(output: Any) -> int:
assert output._obj.handle == 123
return -1 if stage == "close" else 0
library = SimpleNamespace(
D3DKMTOpenAdapterFromLuid=Mock(side_effect=open_adapter),
D3DKMTQueryAdapterInfo=Mock(side_effect=query_adapter),
D3DKMTCloseAdapter=Mock(side_effect=close_adapter),
)
monkeypatch.setattr(windows_gpu.C, "WinDLL", lambda *args, **kwargs: library, raising=False)
if stage == "success":
assert windows_gpu._pci_address(windows_gpu._Luid(0x1709C, 1)) == PciAddress(195, 0, function=1)
else:
with pytest.raises(ValueError if stage == "invalid" else OSError):
windows_gpu._pci_address(windows_gpu._Luid(0x1709C, 1))
assert library.D3DKMTCloseAdapter.call_count == int(stage != "open")
assert ct.sizeof(windows_gpu._OpenAdapterFromLuid) == 12
assert windows_gpu._QueryAdapterInfo.data.offset == 8
@pytest.mark.parametrize("vendor", [0x10DE, 0x1002, 0x8086])
def test_windows_discovery_carries_native_64bit_memory(monkeypatch: pytest.MonkeyPatch, vendor: int) -> None:
"""模拟系统 DXGI 调用,验证三家独显均保留超过 32 位的总容量。"""
monkeypatch.setattr(windows_gpu, "sys", SimpleNamespace(platform="win32"))
def create(iid: Any, output: Any) -> int:
"""使用旧版工厂路径,避免测试依赖宿主 Windows 接口。"""
output._obj.value = 1
return -1 if factory.call_count == 1 else 0
def enumerate_adapter(pointer: Any, index: int, output: Any) -> int:
"""枚举一张独显后正常结束。"""
output._obj.value = 2
return 0 if index == 0 else 0x887A0002
factory = Mock(side_effect=create)
library = SimpleNamespace(CreateDXGIFactory1=factory, D3D12CreateDevice=Mock())
monkeypatch.setattr(windows_gpu.C, "WinDLL", lambda *args, **kwargs: library, raising=False)
monkeypatch.setattr(windows_gpu, "_method", lambda *args: enumerate_adapter)
monkeypatch.setattr(windows_gpu, "_architecture", Mock(return_value=(True, False)))
description = windows_gpu._Description()
description.name, description.vendor, description.device = "GPU", vendor, 1
description.video_memory = 8 * 1024**3
monkeypatch.setattr(windows_gpu, "_description", Mock(return_value=description))
release = Mock()
monkeypatch.setattr(windows_gpu, "_release", release)
selected = windows_gpu.preferred_gpu(windows_gpu.discover_gpus())
assert selected.kind == "discrete" and selected.vendor_id == vendor
assert selected.dedicated_memory_bytes == 8 * 1024**3
assert release.call_count == 2
@pytest.mark.parametrize("include_memory", [False, True])
@pytest.mark.parametrize("failed", [False, True])
def test_vulkan_discovery_only_reads_discrete_local_heaps(
monkeypatch: pytest.MonkeyPatch, include_memory: bool, failed: bool
) -> None:
"""完整枚举夹具验证仅查询独显、本地堆不重复计数、失败保留原因且释放实例。"""
monkeypatch.setattr(vulkan, "sys", SimpleNamespace(platform="linux"))
def create(info: Any, allocator: Any, output: Any) -> int:
"""创建测试实例,不创建逻辑设备或分配显存。"""
output._obj.value = 1
return 0
def enumerate_devices(instance: Any, count: Any, handles: Any) -> int:
"""返回一张独显和一张集显。"""
count._obj.value = 2
if handles is not None:
handles[0], handles[1] = 2, 3
return 0
def properties(handle: int, output: Any) -> None:
"""独显和集显使用不同的原生类型。"""
output._obj.prefix.vendorID = 0x1002
output._obj.prefix.deviceID = handle
output._obj.prefix.deviceType = 2 if handle == 2 else 1
output._obj.prefix.deviceName = b"GPU"
def memory(handle: int, output: Any) -> None:
"""两个内存类型指向同一个堆,另一个大堆属于主机内存。"""
assert handle == 2
if failed:
raise OSError("memory properties unavailable")
value = output._obj
value.type_count, value.heap_count = 2, 3
value.types[0].heap_index = value.types[1].heap_index = 0
value.heaps[0].flags, value.heaps[0].size = 1, 8 * 1024**3
value.heaps[1].flags, value.heaps[1].size = 1, 2 * 1024**3
value.heaps[2].flags, value.heaps[2].size = 0, 128 * 1024**3
library = SimpleNamespace(
vkCreateInstance=Mock(side_effect=create),
vkDestroyInstance=Mock(),
vkEnumeratePhysicalDevices=Mock(side_effect=enumerate_devices),
vkGetPhysicalDeviceProperties=Mock(side_effect=properties),
vkGetPhysicalDeviceMemoryProperties=Mock(side_effect=memory),
)
monkeypatch.setattr(vulkan.ct, "CDLL", lambda *args: library)
discrete, integrated = vulkan.enumerate_vulkan_devices(include_memory=include_memory)
assert discrete.kind == "discrete" and integrated.kind == "integrated"
assert discrete.dedicated_memory_bytes == (10 * 1024**3 if include_memory and not failed else None)
assert bool(discrete.memory_error) is (include_memory and failed)
assert integrated.dedicated_memory_bytes is None and integrated.memory_error is None
assert library.vkGetPhysicalDeviceMemoryProperties.call_count == int(include_memory)
library.vkDestroyInstance.assert_called_once()