导入Pointnet2_ops_lib后GPU加载触发Segmentation Fault求助

问题情况 #

使用Pointnet2模型的加速扩展库pointnet2_ops_lib，导入该库后，任何将模型或张量加载到GPU的操作都会触发Segmentation fault(core dumped)错误。

测试代码 #

import os
import torch
from pointnet_ops import pointnet2_utils 
os.environ["CUDA_VISIBLE_DEVICES"] ="0"
a = torch.tensor([1,2])
a = a.cuda()
print(a.device)

该库通过C++/CUDA扩展实现加速，安装方式为执行python setup.py install。

运行错误 #

segmentation fault(core dumped)

cuda-gdb调试信息 #

Starting program: /home/xxx/anaconda3/envs/py37/bin/python test.py
[Thread debugging using libthread_db enabled]
Using host libthread_db library "/lib/x86_64-linux-gnu/libthread_db.so.1".
[Detaching after fork from child process 23757]
[New Thread 0x7fff8a725700 (LWP 23796)]
[New Thread 0x7fff89f24700 (LWP 23797)]
[New Thread 0x7fff85723700 (LWP 23798)]
[New Thread 0x7fff82f22700 (LWP 23799)]
[New Thread 0x7fff80721700 (LWP 23800)]
[New Thread 0x7fff7df20700 (LWP 23801)]
[New Thread 0x7fff7b71f700 (LWP 23802)]
[New Thread 0x7fff78f1e700 (LWP 23803)]
[New Thread 0x7fff7671d700 (LWP 23804)]
[New Thread 0x7fff73f1c700 (LWP 23805)]
[New Thread 0x7fff7171b700 (LWP 23806)]
[New Thread 0x7fff6ef1a700 (LWP 23807)]
[New Thread 0x7fff6c719700 (LWP 23808)]
[New Thread 0x7fff69f18700 (LWP 23809)]
[New Thread 0x7fff67717700 (LWP 23810)]
[New Thread 0x7fff64f16700 (LWP 23811)]
[New Thread 0x7fff62715700 (LWP 23812)]
[New Thread 0x7fff5ff14700 (LWP 23813)]
[New Thread 0x7fff5d713700 (LWP 23814)]
[New Thread 0x7fff5af12700 (LWP 23815)]
[New Thread 0x7fff58711700 (LWP 23816)]
[New Thread 0x7fff55f10700 (LWP 23817)]
[New Thread 0x7fff5370f700 (LWP 23818)]
[New Thread 0x7fff50f0e700 (LWP 23819)]
[New Thread 0x7fff4e70d700 (LWP 23820)]
[New Thread 0x7fff4bf0c700 (LWP 23821)]
[New Thread 0x7fff4970b700 (LWP 23822)]
[New Thread 0x7fff46f0a700 (LWP 23823)]
[New Thread 0x7fff44709700 (LWP 23824)]
[New Thread 0x7fff41f08700 (LWP 23825)]
[New Thread 0x7fff3f707700 (LWP 23826)]
[New Thread 0x7fff3cf06700 (LWP 23827)]
[New Thread 0x7fff3a705700 (LWP 23828)]
[New Thread 0x7fff37f04700 (LWP 23829)]
[New Thread 0x7fff35703700 (LWP 23830)]
[New Thread 0x7fff32f02700 (LWP 23831)]
[New Thread 0x7fff30701700 (LWP 23832)]
[New Thread 0x7fff2df00700 (LWP 23833)]
[New Thread 0x7fff2d6ff700 (LWP 23834)]
[Detaching after fork from child process 23836]
[New Thread 0x7fff1da72700 (LWP 23840)]
[New Thread 0x7fff1d271700 (LWP 23841)]
[New Thread 0x7fff1ca70700 (LWP 23842)]

Thread 1 "python" received signal SIGSEGV, Segmentation fault.
0x00007fff8d85b220 in ?? ()
   from /home/xxx/anaconda3/envs/py37/lib/python3.7/site-packages/torch/lib/libcudart-a7b20f20.so.11.0

环境配置 #

cuda:11.1
torch:1.12.1+cu113（也曾尝试cu111版本，故障依旧）

解决建议 #

• 严格匹配CUDA与PyTorch版本：将PyTorch切换为与系统CUDA11.1完全匹配的版本（如torch1.12.1+cu111），然后重新从源码编译pointnet2_ops_lib，编译前确保环境变量CUDA_PATH指向系统CUDA11.1路径，避免版本不兼容导致的CUDA上下文错误。
• 调整环境变量设置顺序：将os.environ["CUDA_VISIBLE_DEVICES"] ="0"放在所有导入语句之前，避免CUDA上下文初始化冲突，修改后的测试代码如下：

import os
os.environ["CUDA_VISIBLE_DEVICES"] ="0"
import torch
from pointnet_ops import pointnet2_utils 
a = torch.tensor([1,2])
a = a.cuda()
print(a.device)

• 适配GPU架构编译：修改pointnet2_ops_lib的setup.py，添加与你的GPU型号匹配的CUDA架构编译参数。例如Ampere架构（RTX30/40系列）需添加-gencode=arch=compute_80,code=sm_80，避免因架构不兼容导致的内存访问错误。
• 降级PyTorch版本：尝试使用兼容性更稳定的旧版本PyTorch，如torch1.9.0+cu111，该版本对老CUDA扩展的支持更好。
• 检查NVIDIA驱动版本：确保驱动版本不低于CUDA11.1要求的最低版本（455.23.05），驱动版本不足会导致CUDA上下文初始化失败，触发段错误。

内容的提问来源于stack exchange，提问作者name555difficult