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
fix(hardware): preserve Windows DXGI memory in GPU inventory
Xiaomeng Zhao5c832ae37ff8
diff --git a/mineru/model/runtime/hardware.py b/mineru/model/runtime/hardware.pyindex a5b6ff65..f6479ea3 100644--- a/mineru/model/runtime/hardware.py+++ b/mineru/model/runtime/hardware.py@@ -17,7 +17,7 @@ from .hardware_probe import run_probefrom .hardware_system import collect_system_hardwarefrom .hardware_types import GpuInfo, HardwareInfo, HardwareInfoSection, PackageInfo, RuntimeProbefrom .onnx_calibration_cache import inspect_calibration_record-from .gpu import GPU_KIND_ORDER, GpuIdentity, gpu_identities_match, match_gpu_identities+from .gpu import GPU_KIND_ORDER, GpuIdentity, PciAddress, gpu_identities_match, match_gpu_identitiesif TYPE_CHECKING:from .webgpu import WebGPUDevice@@ -99,8 +99,26 @@ def config_summary() -> dict[str, Any]:}+def _merge_gpu_facts(target: GpuInfo, row: dict[str, Any], *, native_unique: bool = True) -> None:+ kind = row.get("kind", "unknown")+ target.kind = kind if kind in GPU_KIND_ORDER and native_unique else "unknown"+ target.kind_source = row.get("kind_source", "unknown")+ target.unified_memory = row.get("unified_memory")+ for attribute in ("dedicated_memory_bytes", "shared_memory_bytes"):+ if row.get(attribute) is not None:+ setattr(target, attribute, row[attribute])+++def _match_dxgi_pci(identities: list[GpuIdentity], targets: list[GpuIdentity]) -> dict[int, int]:+ """仅比较已知 PCI 地址,避免无地址的同型号条目阻止物理匹配。"""+ sources = [(index, identity) for index, identity in enumerate(identities) if identity.pci_address is not None]+ destinations = [(index, identity) for index, identity in enumerate(targets) if identity.bus_id]+ matches = match_gpu_identities([identity for _, identity in sources], [identity for _, identity in destinations])+ return {index: destinations[match][0] for (index, _), match in zip(sources, matches) if match is not None}++def merge_gpu_inventory(gpus: list[GpuInfo], probe: RuntimeProbe) -> None:- """按明确身份补充原生类型和显存;无法唯一匹配的多卡库存不猜测归属。"""+ """优先按身份补充显卡属性;DXGI 的同型号条目可补充一致的属性。"""if probe.status != "ok":returnvendors = {0x8086: "Intel", 0x10DE: "NVIDIA", 0x1002: "AMD"}@@ -108,18 +126,26 @@ def merge_gpu_inventory(gpus: list[GpuInfo], probe: RuntimeProbe) -> None:row for row in probe.data.get("devices", []) if row.get("vendor_id") is not None and row.get("device_id") is not None]identities = [- GpuIdentity(row["vendor_id"], row["device_id"], luid=row.get("luid"), bus_id=row.get("bus_id"), name=row.get("name"))+ GpuIdentity(+ row["vendor_id"],+ row["device_id"],+ luid=row.get("luid"),+ bus_id=row.get("bus_id"),+ name=row.get("name"),+ pci_address=PciAddress(**row["pci_address"]) if row.get("pci_address") is not None else None,+ )for row in rows]targets = [gpu.identity for gpu in gpus]matches = match_gpu_identities(identities, targets)+ if probe.name == "dxgi":+ pci_matches = _match_dxgi_pci(identities, targets)+ pci_targets = set(pci_matches.values())+ matches = tuple(pci_matches.get(index, None if match in pci_targets else match) for index, match in enumerate(matches))native_matches = match_gpu_identities(identities, identities)for index, (row, matched) in enumerate(zip(rows, matches)):vendor, chip = row["vendor_id"], row["device_id"]name = row.get("name")- kind = row.get("kind", "unknown")- if kind not in GPU_KIND_ORDER or native_matches[index] is None:- kind = "unknown"if matched is not None:target = gpus[matched]target.name = target.name or name@@ -133,12 +159,25 @@ def merge_gpu_inventory(gpus: list[GpuInfo], probe: RuntimeProbe) -> None:source=probe.name,)gpus.append(target)- target.kind = kind- target.kind_source = row.get("kind_source", "unknown")- target.unified_memory = row.get("unified_memory")- for attribute in ("dedicated_memory_bytes", "shared_memory_bytes"):- if row.get(attribute) is not None:- setattr(target, attribute, row[attribute])+ _merge_gpu_facts(target, row, native_unique=native_matches[index] is not None)++ if probe.name != "dxgi":+ return+ for index, identity in enumerate(targets):+ model = (identity.vendor_id, identity.device_id)+ if index in matches or sum((target.vendor_id, target.device_id) == model for target in targets) != 1:+ continue+ compatible = [i for i, native in enumerate(identities) if gpu_identities_match(identity, native)]+ if len(compatible) < 2:+ continue+ # 缺少可区分的身份时只报告一致的事实。选型可用的最小容量不能作为单卡的实际容量。+ first, *others = [rows[i] for i in compatible]+ facts = {+ key: first[key]+ for key in ("kind", "kind_source", "unified_memory", "dedicated_memory_bytes", "shared_memory_bytes")+ if key in first and all(row.get(key) == first[key] for row in others)+ }+ _merge_gpu_facts(gpus[index], facts)def _calibration_adapters(probes: list[RuntimeProbe]) -> tuple[GpuAdapter | WebGPUDevice, ...] | None:diff --git a/tests/unittest/test_hardware_info.py b/tests/unittest/test_hardware_info.pyindex e2b0a7bf..9a945bad 100644--- a/tests/unittest/test_hardware_info.py+++ b/tests/unittest/test_hardware_info.py@@ -237,6 +237,73 @@ def test_windows_cim_and_dxgi_memory() -> None:assert len(result.gpus) == 1 and result.gpus[0].dedicated_memory_bytes == 8 * 1024**3[email protected]("reverse", [False, True])+ ("bus_id", "alias_memory_gib", "expected_memory_gib"),+ [(None, 8, 8), (None, 16, None), (None, None, None), ("0000:01:00.0", 16, 8)],+)+def test_windows_dxgi_duplicate_model_preserves_known_memory(+ reverse: bool, bus_id: str | None, alias_memory_gib: int | None, expected_memory_gib: int | None+) -> None:+ """优先使用可唯一关联的 PCI 条目,否则只补充同型号条目一致的数据。"""+ gib = 1024**3+ result = SystemHardware()+ hardware_system.parse_windows_hardware(+ {"gpus": {"Name": "Radeon", "PNPDeviceID": "PCI\\VEN_1002&DEV_7550", "DriverVersion": "test"}}, result+ )+ result.gpus[0].bus_id = bus_id+ physical = {+ "vendor_id": 0x1002,+ "device_id": 0x7550,+ "luid": "physical",+ "pci_address": {"bus": 1, "device": 0, "domain": None, "function": 0},+ "kind": "discrete",+ "kind_source": "d3d12",+ "unified_memory": False,+ "dedicated_memory_bytes": 8 * gib,+ "shared_memory_bytes": 16 * gib,+ }+ alias = {+ **physical,+ "luid": "other",+ "pci_address": None,+ "dedicated_memory_bytes": alias_memory_gib * gib if alias_memory_gib is not None else None,+ }+ devices = [alias, physical] if reverse else [physical, alias]++ hardware.merge_gpu_inventory(result.gpus, RuntimeProbe("dxgi", "ok", {"devices": devices}))++ (gpu,) = result.gpus+ assert gpu.name == "Radeon" and gpu.driver == "test" and gpu.source == "cim"+ assert (gpu.kind, gpu.kind_source, gpu.unified_memory) == ("discrete", "d3d12", False)+ assert gpu.shared_memory_bytes == 16 * gib+ assert gpu.dedicated_memory_bytes == (expected_memory_gib * gib if expected_memory_gib is not None else None)++[email protected]("has_bus_id", [False, True])+def test_windows_dxgi_same_model_physical_cards_use_pci_addresses(has_bus_id: bool) -> None:+ """同型号多张卡按 PCI 地址保留各自的容量,缺少系统地址时不猜测归属。"""+ gpus = [+ GpuInfo(name="Radeon", device_id="0x1002:0x7550", bus_id=f"0000:0{bus}:00.0" if has_bus_id else None) for bus in (1, 2)+ ]+ devices = [+ {+ "vendor_id": 0x1002,+ "device_id": 0x7550,+ "luid": str(bus),+ "pci_address": {"bus": bus, "device": 0, "domain": None, "function": 0} if bus else None,+ "kind": "discrete",+ "dedicated_memory_bytes": capacity * 1024**3,+ }+ for bus, capacity in ((2, 16), (0, 4), (1, 8))+ ]++ hardware.merge_gpu_inventory(gpus, RuntimeProbe("dxgi", "ok", {"devices": devices}))++ expected = [8 * 1024**3, 16 * 1024**3] if has_bus_id else [None, None]+ assert [gpu.dedicated_memory_bytes for gpu in gpus] == expected++def test_windows_cim_partial_failure_keeps_memory_and_gpu(monkeypatch: pytest.MonkeyPatch) -> None:"""一个 CIM 类不可用时,其余硬件查询仍独立运行并进入报告。"""run = Mock(side_effect=[None, '{"TotalVisibleMemorySize": 1024}', '{"Name": "GPU"}'])