AI智能摘要
你是否在为JetsonOrin平台编译PyTorch后，苦于找不到官方的TorchVision预编译包？这份详尽的指南，将带你一步步从源码成功构建出适配的whl安装文件。从环境变量设置、依赖库安装到最终测试验证，我们为你揭示了官方未提及的ARM架构编译关键步骤。跟随指南，你将彻底摆脱依赖困境，在边缘设备上自主部署完整的PyTorch视觉生态。
— AI 生成的文章内容摘要

基础信息

TorchVision 与 Torch 版本对照表

torchtorchvisionPython
main / nightlymain / nightly>=3.10, <=3.14
2.90.24>=3.10, <=3.14
2.80.23>=3.9, <=3.13
2.70.22>=3.9, <=3.13
2.60.21>=3.9, <=3.12

PyTorch 编译教程请参考

构建步骤

拉取项目代码

git clone --recursive --branch v0.22.0 https://github.com/pytorch/vision torchvision
cd torchvision
sudo apt-get update && sudo apt-get install -y libjpeg-dev libpng-dev libwebp-dev libavcodec-dev libavformat-dev libswscale-dev ffmpeg

使用编译 PyTorch 的虚拟环境

source ../pytorch/.venv/bin/activate
uv pip install numpy pillow

构建

export CPATH="/usr/include/aarch64-linux-gnu:/usr/local/cuda/include:$CPATH"
export LIBRARY_PATH="/usr/lib/aarch64-linux-gnu:/usr/local/cuda/lib64:$LIBRARY_PATH"
export LD_LIBRARY_PATH="/usr/lib/aarch64-linux-gnu:/usr/local/cuda/lib64:$LD_LIBRARY_PATH"
export FORCE_CUDA=1
export TORCH_CUDA_ARCH_LIST="8.7"

python3 setup.py bdist_wheel

whl 输出路径

./torchvision/dist/torchvision-0.22.0+9eb57cd-cp312-cp312-linux_aarch64.whl

安装

uv pip install ./dist/torchvision-0.22.0+9eb57cd-cp312-cp312-linux_aarch64.whl

测试验证

python -c "
import torch
import torchvision
print(f'Torchvision Version: {torchvision.__version__}')
input_tensor = torch.rand(5, 4).cuda()
scores = torch.rand(5).	cuda()
try:
    torchvision.ops.nms(input_tensor, scores, 0.5)
    print('✅ CUDA Operators: SUCCESS')
except Exception as e:
    print(f'❌ CUDA Operators: FAILED, error: {e}')
from torchvision.io import image
print('✅ Basic Image IO: Functional')
"