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
Select table ONNX CUDA automatically when available
myhlolibf8a518d235e
diff --git a/docs/en/usage/tiers.md b/docs/en/usage/tiers.mdindex 709be4f0..a7777a9e 100644--- a/docs/en/usage/tiers.md+++ b/docs/en/usage/tiers.md@@ -28,7 +28,9 @@ PDF and images support all four tiers. Office, OpenDocument, RTF, EPUB, OFD, HTM| Windows, accelerator and `full` | Torch | LMDeploy || CPU environment | ONNX / CPU | llama.cpp |-Automatic selection uses installed dependencies and available devices. An explicit choice with missing dependencies fails instead of silently selecting another backend. ONNX small models run on CPU; VLM device selection is independent.+Automatic selection uses installed dependencies and available devices. An explicit choice with missing dependencies fails instead of silently selecting another backend. ONNX Layout, OCR, and formula models run on CPU; VLM device selection is independent.++Table classification, SLANet, and UNet use `MINERU_TABLE_DEVICE=auto` by default: the CPU `onnxruntime` package uses CPU; replacing it with a compatible `onnxruntime-gpu` package enables CUDA when its provider is available. CUDA initialization failures are logged and fall back to CPU. Do not install both ORT distributions together. Set `MINERU_TABLE_DEVICE=cpu` to force CPU or `cuda` to explicitly attempt CUDA. This option affects only table models, and logs report the actual providers.Intel XPU is excluded from automatic LMDeploy selection. Linux prefers an installed XPU-compatible vLLM and otherwise falls back to llama.cpp; Windows uses llama.cpp.diff --git a/docs/next/model-assets.md b/docs/next/model-assets.mdindex 1a7a4b98..be5150a2 100644--- a/docs/next/model-assets.md+++ b/docs/next/model-assets.md@@ -30,7 +30,11 @@ mineru-kit parse input.pdf --tier basic -o output.md````MINERU_MODEL_BASE_DIR` 可指定独立模型根目录。`--small-backend`、`--vlm-engine` 只覆盖当前模型管理命令;解析和服务运行时读取环境变量或配置文件。-所有 ONNX 会话固定使用 CPU,最低版本为 ONNX Runtime 1.20.1,支持资源包所需的 ONNX IR 10。+Layout、OCR 和公式的 ONNX 会话固定使用 CPU,最低版本为 ONNX Runtime 1.20.1,支持资源包所需的 ONNX IR 10。++表格分类、SLANet 和 UNet 共用独立的设备策略:`MINERU_TABLE_DEVICE=auto`(默认)根据 ONNX Runtime 的可用 provider 选择设备。安装 CPU 版 `onnxruntime` 时使用 CPU;替换为与 CUDA/cuDNN 环境兼容的 `onnxruntime-gpu` 后优先使用 CUDA,CUDA 会话初始化失败则记录原因并回退 CPU。不要同时安装两个发行包,它们共享同一个 Python 模块。日志中的实际 providers 表示最终加载结果。++设置 `MINERU_TABLE_DEVICE=cpu` 可固定表格在 CPU 执行(ZeroGPU Space 默认采用此配置);设置 `cuda` 可显式尝试 CUDA,仍保留故障回退。该选项不影响 Layout、OCR 和公式的 ONNX 后端,也不影响 Torch 设备选择。所有表格会话均读取 `MINERU_INTRA_OP_NUM_THREADS` 和 `MINERU_INTER_OP_NUM_THREADS`,限制 CPU 算子及回退路径的线程数。llama.cpp 的设备参数独立于 ONNX provider;Torch 和 VLM 仍沿用各自设备策略。## 资源结构diff --git a/docs/zh/usage/tiers.md b/docs/zh/usage/tiers.mdindex d8fd5d95..3d200749 100644--- a/docs/zh/usage/tiers.md+++ b/docs/zh/usage/tiers.md@@ -28,7 +28,9 @@ PDF 和图片可选四档。Office、OpenDocument、RTF、EPUB、OFD、HTML、CS| Windows,有加速器并安装 `full` | Torch | LMDeploy || CPU 环境 | ONNX / CPU | llama.cpp |-自动选择依据已安装依赖和可用设备;显式指定的后端缺依赖时会报错,不会静默改用另一种。ONNX 小模型在 CPU 上执行,VLM 的设备选择独立。+自动选择依据已安装依赖和可用设备;显式指定的后端缺依赖时会报错,不会静默改用另一种。Layout、OCR 和公式的 ONNX 模型在 CPU 上执行,VLM 的设备选择独立。++表格分类、SLANet 和 UNet 默认使用 `MINERU_TABLE_DEVICE=auto`:CPU 版 `onnxruntime` 使用 CPU;替换为兼容的 `onnxruntime-gpu` 后,有可用 CUDA provider 时优先使用 CUDA,初始化失败则记录原因并回退 CPU。两个 ORT 发行包不要同时安装。设置 `MINERU_TABLE_DEVICE=cpu` 可固定 CPU,设置 `cuda` 可显式尝试 CUDA;该选项仅影响表格模型。日志会记录实际使用的 providers。Intel XPU 不会自动选择 LMDeploy。Linux 上安装了支持 XPU 的 vLLM 时优先使用 vLLM,否则回退 llama.cpp;Windows 上使用 llama.cpp。diff --git a/mineru/model/runtime/onnx.py b/mineru/model/runtime/onnx.pyindex 419d5d23..34d5c12e 100644--- a/mineru/model/runtime/onnx.py+++ b/mineru/model/runtime/onnx.py@@ -1,5 +1,5 @@# Copyright (c) Opendatalab. All rights reserved.-"""ONNX Runtime CPU 会话与可选的表格 CUDA 会话。"""+"""ONNX Runtime CPU 会话与自动选择设备的表格会话。"""import osimport time@@ -22,7 +22,7 @@ def get_op_num_threads(env_name: str) -> int:def ort_providers(device: str | None = None) -> list[tuple[str, dict[str, object]]]:- """所有 ONNX 模型固定使用 CPU,宿主 Torch 设备不参与 provider 选择。"""+ """通用 ONNX 模型固定使用 CPU;表格工厂独立选择 CUDA provider。"""return [("CPUExecutionProvider", {"arena_extend_strategy": "kSameAsRequested"})]@@ -41,10 +41,10 @@ def ort_session(model_path: str, device: str | None = None, intra_op_num_threadsdef table_ort_session(model_path: str, *, sess_options: ort.SessionOptions | None = None) -> ort.InferenceSession:- """表格独立选择 CPU/CUDA;创建失败或 provider 静默降级时明确记录并保留 CPU 回退。"""- device = os.getenv("MINERU_TABLE_DEVICE", "cpu").strip().lower()- if device not in {"cpu", "cuda"}:- raise ValueError("MINERU_TABLE_DEVICE must be 'cpu' or 'cuda'")+ """表格默认按可用 provider 选择设备;支持固定 CPU/CUDA,并明确记录 CUDA 失败回退。"""+ device = os.getenv("MINERU_TABLE_DEVICE", "auto").strip().lower()+ if device not in {"auto", "cpu", "cuda"}:+ raise ValueError("MINERU_TABLE_DEVICE must be 'auto', 'cpu' or 'cuda'")opts = sess_options if sess_options is not None else ort.SessionOptions()opts.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALLopts.log_severity_level = 3@@ -61,8 +61,10 @@ def table_ort_session(model_path: str, *, sess_options: ort.SessionOptions | Nonopts.enable_profiling = Trueopts.profile_file_prefix = str(Path(profile_dir) / f"{Path(model_path).stem}-{os.getpid()}")providers = ort_providers()- if device == "cuda":- if "CUDAExecutionProvider" in ort.get_available_providers():+ use_cuda = False+ if device != "cpu":+ use_cuda = "CUDAExecutionProvider" in ort.get_available_providers()+ if use_cuda:providers = [("CUDAExecutionProvider",@@ -75,18 +77,18 @@ def table_ort_session(model_path: str, *, sess_options: ort.SessionOptions | Non),*providers,]- else:+ elif device == "cuda":logger.warning("Table CUDA unavailable; falling back to CPU: {}", model_path)started = time.perf_counter()try:session = ort.InferenceSession(model_path, sess_options=opts, providers=providers)except Exception as exc:- if device != "cuda" or providers == ort_providers():+ if not use_cuda:raiselogger.warning("Table CUDA initialization failed; falling back to CPU: {}: {}", model_path, exc)session = ort.InferenceSession(model_path, sess_options=opts, providers=ort_providers())actual = session.get_providers()- if device == "cuda" and "CUDAExecutionProvider" not in actual:+ if use_cuda and "CUDAExecutionProvider" not in actual:logger.warning("Table requested CUDA but actual providers are {}: {}", actual, model_path)logger.info("Table ONNX ready: model={}, requested={}, providers={}, init_s={:.3f}, pid={}",diff --git a/tests/unittest/test_table_onnx_runtime.py b/tests/unittest/test_table_onnx_runtime.pyindex d7c8e30e..bd811e03 100644--- a/tests/unittest/test_table_onnx_runtime.py+++ b/tests/unittest/test_table_onnx_runtime.py@@ -1,4 +1,4 @@-"""表格 CUDA provider 的显式选择、回退与 CPU 模型边界。"""+"""表格 CUDA provider 的自动选择、显式覆盖、回退与 CPU 模型边界。"""from types import SimpleNamespacefrom unittest.mock import Mock@@ -9,10 +9,11 @@ import pytestfrom mineru.model.runtime import onnx-def test_table_default_is_cpu(monkeypatch: pytest.MonkeyPatch) -> None:- """Torch 使用 CUDA 不应隐式改变表格模型设备。"""+def test_table_default_uses_cpu_without_cuda_provider(monkeypatch: pytest.MonkeyPatch) -> None:+ """CPU ORT 包保持 CPU 执行,Torch 设备不参与表格 provider 决策。"""monkeypatch.delenv("MINERU_TABLE_DEVICE", raising=False)monkeypatch.setenv("MINERU_DEVICE_MODE", "cuda")+ monkeypatch.setattr(onnx.ort, "get_available_providers", lambda: ["CPUExecutionProvider"])session = SimpleNamespace(get_providers=lambda: ["CPUExecutionProvider"])factory = Mock(return_value=session)monkeypatch.setattr(onnx.ort, "InferenceSession", factory)@@ -20,9 +21,13 @@ def test_table_default_is_cpu(monkeypatch: pytest.MonkeyPatch) -> None:assert factory.call_args.kwargs["providers"] == onnx.ort_providers()-def test_cuda_is_explicit_and_other_onnx_models_stay_cpu(monkeypatch: pytest.MonkeyPatch) -> None:- """表格 CUDA 使用低冷启动开销选项,OCR/Layout ONNX 仍保留 CPU 策略。"""- monkeypatch.setenv("MINERU_TABLE_DEVICE", "cuda")[email protected]("device", [None, "auto", "cuda"])+def test_cuda_selection_and_other_onnx_models_stay_cpu(monkeypatch: pytest.MonkeyPatch, device: str | None) -> None:+ """默认自动和显式 CUDA 都选择 GPU,但 OCR/Layout ONNX 仍保留 CPU 策略。"""+ if device is None:+ monkeypatch.delenv("MINERU_TABLE_DEVICE", raising=False)+ else:+ monkeypatch.setenv("MINERU_TABLE_DEVICE", device)monkeypatch.setenv("MINERU_INTRA_OP_NUM_THREADS", "4")monkeypatch.setenv("MINERU_INTER_OP_NUM_THREADS", "1")monkeypatch.setattr(onnx.ort, "get_available_providers", lambda: ["CUDAExecutionProvider", "CPUExecutionProvider"])@@ -38,10 +43,21 @@ def test_cuda_is_explicit_and_other_onnx_models_stay_cpu(monkeypatch: pytest.Monassert factory.call_args.kwargs["providers"] == onnx.ort_providers()+def test_explicit_cpu_overrides_available_cuda_provider(monkeypatch: pytest.MonkeyPatch) -> None:+ """Space 的显式 CPU 配置不能因安装 GPU ORT 包而申请 CUDA 会话。"""+ monkeypatch.setenv("MINERU_TABLE_DEVICE", "cpu")+ monkeypatch.setattr(onnx.ort, "get_available_providers", lambda: ["CUDAExecutionProvider", "CPUExecutionProvider"])+ factory = Mock(return_value=SimpleNamespace(get_providers=lambda: ["CPUExecutionProvider"]))+ monkeypatch.setattr(onnx.ort, "InferenceSession", factory)+ onnx.table_ort_session("table.onnx")+ assert factory.call_args.kwargs["providers"] == onnx.ort_providers()++[email protected]("device", ["auto", "cuda"])@pytest.mark.parametrize("mode", ["unavailable", "exception", "silent"])-def test_cuda_failure_is_reported_and_cpu_remains_available(monkeypatch: pytest.MonkeyPatch, mode: str) -> None:+def test_cuda_failure_is_reported_and_cpu_remains_available(monkeypatch: pytest.MonkeyPatch, mode: str, device: str) -> None:"""覆盖未安装 CUDA、动态库失败以及 ORT 自动降级,不能把回退误报为 GPU。"""- monkeypatch.setenv("MINERU_TABLE_DEVICE", "cuda")+ monkeypatch.setenv("MINERU_TABLE_DEVICE", device)monkeypatch.setattr(onnx.ort,"get_available_providers",@@ -53,7 +69,7 @@ def test_cuda_failure_is_reported_and_cpu_remains_available(monkeypatch: pytest.monkeypatch.setattr(onnx.ort, "InferenceSession", factory)monkeypatch.setattr(onnx.logger, "warning", warning)assert onnx.table_ort_session("table.onnx") is cpu- assert warning.called+ assert warning.called == (mode != "unavailable" or device == "cuda")if mode in {"unavailable", "exception"}:assert factory.call_args.kwargs["providers"] == onnx.ort_providers()