鲲鹏920 Arm64 使用 OCR V5解析pdf或者图片报错：`Segmentation fault` is detected by the operating system

[sduqlsc]

🔎 Search before asking

• I have searched the PaddleOCR Docs and found no similar bug report.
• I have searched the PaddleOCR Issues and found no similar bug report.
• I have searched the PaddleOCR Discussions and found no similar bug report.

🐛 Bug (问题描述)

图片大小为1656*2339，无论enable_mkldnn是True还是False，都会报错。pdf识别同样的错误，145K以及3.8M都会报错，但是使用general_ocr_002.png正常。同样的文件，V3目前表现正常。并且这里也测试使用自己编译paddleinference以及ocr测试，同样效果不理想。

ocr = PaddleOCR(
            lang="ch",
            ocr_version="PP-OCRv3",
            enable_mkldnn=True,
            cpu_threads=8,
            use_doc_orientation_classify=False,
            use_doc_unwarping=False,
            use_textline_orientation=False,
            device='cpu'
        )

🏃‍♂️ Environment (运行环境)

OS : Ubuntu 22.04.4  ARM64 Docker
Python: 3.10.12
paddleocr                3.2.0
CPU: kunpeng-920
CUDA： None
NPU： None

🌰 Minimal Reproducible Example (最小可复现问题的Demo)

Python 3.10.12 (main, Aug 15 2025, 14:32:43) [GCC 11.4.0] on linux

>>> import os
>>> import sys
>>> import time
>>> import io
>>> import json
>>> from datetime import datetime
>>> from paddleocr import PaddleOCR
>>> img_file="../img_test/contract_00.jpg"
>>> ocr = PaddleOCR(
...             lang="ch",
...             device='cpu',
...             enable_mkldnn=False,
...             cpu_threads=8,
...             use_doc_orientation_classify=False,
...             use_doc_unwarping=False,
...             use_textline_orientation=True
...         )
Creating model: ('PP-LCNet_x1_0_textline_ori', None)
Model files already exist. Using cached files. To redownload, please delete the directory manually: `/root/.paddlex/official_models/PP-LCNet_x1_0_textline_ori`.
Creating model: ('PP-OCRv5_server_det', None)
Model files already exist. Using cached files. To redownload, please delete the directory manually: `/root/.paddlex/official_models/PP-OCRv5_server_det`.
Creating model: ('PP-OCRv5_server_rec', None)
Model files already exist. Using cached files. To redownload, please delete the directory manually: `/root/.paddlex/official_models/PP-OCRv5_server_rec`.
>>> result = ocr.predict(img_file)


--------------------------------------
C++ Traceback (most recent call last):
--------------------------------------
0   paddle::AnalysisPredictor::ZeroCopyRun(bool)
1   paddle::framework::NaiveExecutor::RunInterpreterCore(std::vector<std::string, std::allocator<std::string > > const&, bool, bool)
2   paddle::framework::InterpreterCore::Run(std::vector<std::string, std::allocator<std::string > > const&, bool, bool, bool, bool)
3   paddle::framework::PirInterpreter::Run(std::vector<std::string, std::allocator<std::string > > const&, bool, bool, bool, bool)
4   paddle::framework::PirInterpreter::TraceRunImpl()
5   paddle::framework::PirInterpreter::TraceRunInstructionList(std::vector<std::unique_ptr<paddle::framework::InstructionBase, std::default_delete<paddle::framework::InstructionBase> >, std::allocator<std::unique_ptr<paddle::framework::InstructionBase, std::default_delete<paddle::framework::InstructionBase> > > > const&)
6   paddle::framework::PirInterpreter::RunInstructionBase(paddle::framework::InstructionBase*)
7   paddle::framework::PhiKernelInstruction::Run()
8   phi::KernelImpl<void (*)(phi::CPUContext const&, phi::DenseTensor const&, phi::DenseTensor const&, std::vector<int, std::allocator<int> > const&, std::vector<int, std::allocator<int> > const&, std::string const&, std::vector<int, std::allocator<int> > const&, int, std::string const&, phi::DenseTensor*), &(void phi::ConvKernel<float, phi::CPUContext>(phi::CPUContext const&, phi::DenseTensor const&, phi::DenseTensor const&, std::vector<int, std::allocator<int> > const&, std::vector<int, std::allocator<int> > const&, std::string const&, std::vector<int, std::allocator<int> > const&, int, std::string const&, phi::DenseTensor*))>::Compute(phi::KernelContext*)
9   void phi::ConvKernelImpl<float, phi::CPUContext>(phi::CPUContext const&, phi::DenseTensor const&, phi::DenseTensor const&, std::vector<int, std::allocator<int> > const&, std::vector<int, std::allocator<int> > const&, std::string const&, int, std::vector<int, std::allocator<int> > const&, std::string const&, phi::DenseTensor*)
10  phi::funcs::Im2ColFunctor<(phi::funcs::ColFormat)0, phi::CPUContext, float>::operator()(phi::CPUContext const&, phi::DenseTensor const&, std::vector<int, std::allocator<int> > const&, std::vector<int, std::allocator<int> > const&, std::vector<int, std::allocator<int> > const&, phi::DenseTensor*, common::DataLayout)

----------------------
Error Message Summary:
----------------------
FatalError: `Segmentation fault` is detected by the operating system.
  [TimeInfo: *** Aborted at 1759193212 (unix time) try "date -d @1759193212" if you are using GNU date ***]
  [SignalInfo: *** SIGSEGV (@0xfff58b4b1480) received by PID 171857 (TID 0xfffe1c250420) from PID 18446744071751537792 ***]

Segmentation fault (core dumped)

[pulkitchowdry]

I have observed a similar issue on macOS. Following is the error message,

Creating model: ('PP-OCRv5_server_det', None)
Model files already exist. Using cached files. To redownload, please delete the directory manually: /Users/cc/.paddlex/official_models/PP-OCRv5_server_det.
Creating model: ('PP-OCRv5_server_rec', None)
Model files already exist. Using cached files. To redownload, please delete the directory manually: /Users/cc/.paddlex/official_models/PP-OCRv5_server_rec.
zsh: segmentation fault python3 preprocessing/image_proc.py

I observe this whenever my input is a png or pdf file while it never happens for jpg files. I am not sure if this is due to multi-processing. In v5, using cpu_thread=1 also did not solve it.

OCR Model
ocr = PaddleOCR(
text_detection_model_name="PP-OCRv5_server_det",
text_recognition_model_name="PP-OCRv5_server_rec",
use_doc_orientation_classify=False,
use_doc_unwarping=False,
use_textline_orientation=False,
)

Environment
PaddleOCR Version: 3.2.0
Python Version: 3.11.4
OS: macOS Sequoia Version 15.6.1

[cuicheng01]

Hello, this issue may be caused by a boundary condition bug in the Paddle framework, which has already been fixed in this PR. If you use the develop version of paddlepaddle, this problem should no longer occur.

[pulkitchowdry]

Thanks for sharing some insights. I tried using the develop version - paddlepaddle-3.3.0.dev20250930-cp311-cp311-macosx_11_0_arm64.whl

But the issue persists.

`
Error
zsh: segmentation fault python3 ./preprocessing/image_proc.py
Additional multi-processing warning after this,
UserWarning: resource_tracker: There appear to be 1 leaked semaphore objects to clean up at shutdown
warnings.warn('resource_tracker: There appear to be %d '

paddleocr version: 3.2.0
paddlepaddle version: 3.3.0.dev20250930
paddlex version: 3.2.1
Python version: 3.11.4
`