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
perf(runtime): disable NNPACK by default for CPU convolutions
myhlolid05105e92525
diff --git a/mineru/model/runtime/cpu_threads.py b/mineru/model/runtime/cpu_threads.pyindex 555c0076..37d12912 100644--- a/mineru/model/runtime/cpu_threads.py+++ b/mineru/model/runtime/cpu_threads.py@@ -1,5 +1,5 @@# Copyright (c) Opendatalab. All rights reserved.-"""在首次 Torch 模型初始化和工作线程执行前限制 CPU 算子线程。"""+"""在首次 Torch 初始化关闭 CPU NNPACK,并限制后续工作线程的 CPU 算子线程。"""from __future__ import annotations@@ -43,12 +43,14 @@ def _positive_env(name: str) -> int | None:def initialize_torch_threads() -> int:- """仅在 Torch 模型边界加载 Torch,不调高调用方已有的更低线程数。"""+ """惰性初始化 Torch 的 CPU 默认后端与线程,不调高调用方已有的更低线程数。"""global _limitimport torchwith _lock:if _limit is None:+ # NNPACK 只参与 CPU 卷积;首次默认关闭,后续线程准备保留调用方显式设置。+ torch.backends.nnpack.set_flags(False)explicit = [_positive_env(name) for name in ("OMP_NUM_THREADS", "MKL_NUM_THREADS")]configured = (min(value for value in explicit if value is not None)diff --git a/tests/unittest/test_torch_cpu_threads.py b/tests/unittest/test_torch_cpu_threads.pyindex 3da413f7..cff4e433 100644--- a/tests/unittest/test_torch_cpu_threads.py+++ b/tests/unittest/test_torch_cpu_threads.py@@ -44,6 +44,7 @@ def test_torch_thread_limits(monkeypatch: pytest.MonkeyPatch, env: dict[str, strmonkeypatch.setattr(cpu_threads, "effective_cpu_count", lambda: 12)local = threading.local()calls: list[int] = []+ nnpack_calls: list[bool] = []def init() -> None:"""模拟新 OpenMP 工作线程需要显式初始化线程状态。"""@@ -54,12 +55,18 @@ def test_torch_thread_limits(monkeypatch: pytest.MonkeyPatch, env: dict[str, str"""只修改当前模拟工作线程,检测配置是否传播。"""local.count = count- fake = SimpleNamespace(init_num_threads=init, get_num_threads=lambda: local.count, set_num_threads=set_threads)+ fake = SimpleNamespace(+ init_num_threads=init,+ get_num_threads=lambda: local.count,+ set_num_threads=set_threads,+ backends=SimpleNamespace(nnpack=SimpleNamespace(set_flags=nnpack_calls.append)),+ )monkeypatch.setitem(sys.modules, "torch", fake)assert cpu_threads.initialize_torch_threads() == expectedwith ThreadPoolExecutor(max_workers=3) as pool:assert pool.submit(cpu_threads.initialize_torch_threads).result() == expectedassert len(set(calls)) >= 2+ assert nnpack_calls == [False]def test_thread_module_import_and_uninitialized_stage_are_light() -> None:@@ -85,12 +92,20 @@ os.environ.pop("MKL_NUM_THREADS",None)from mineru.model.runtime.cpu_threads import initialize_torch_threadsfrom concurrent.futures import ThreadPoolExecutorimport torch+torch.backends.nnpack.set_flags(True)initialize_torch_threads()+assert not torch._C._get_nnpack_enabled()def run(_):initialize_torch_threads()assert torch.get_num_threads() <= 2+ assert not torch._C._get_nnpack_enabled()return (torch.ones(64,64) @ torch.ones(64,64)).sum().item()with ThreadPoolExecutor(max_workers=3) as pool:assert list(pool.map(run,range(6))) == [262144.0]*6+with torch.backends.nnpack.flags(enabled=True):+ with ThreadPoolExecutor(max_workers=1) as pool:+ pool.submit(initialize_torch_threads).result()+ assert torch._C._get_nnpack_enabled()+assert not torch._C._get_nnpack_enabled()"""subprocess.run([sys.executable, "-c", code], check=True, timeout=60)