【已解决】RuntimeError: PyTorch has CUDA version 12.1 and torch_sparse has CUDA version 11.8

问题描述

        今天遇到这么一个问题：RuntimeError: Detected that PyTorch and torch_sparse were compiled with different CUDA versions. PyTorch has CUDA version 12.1 and torch_sparse has CUDA version 11.8. Please reinstall the torch_sparse that matches your PyTorch install.
        详细内容如下：

/home/visionx/anaconda3/envs/gdem/lib/python3.8/site-packages/torch_geometric/typing.py:101: UserWarning: An issue occurred while importing 'torch-sparse'. Disabling its usage. Stacktrace: Detected that PyTorch and torch_sparse were compiled with different CUDA versions. PyTorch has CUDA version 12.1 and torch_sparse has CUDA version 11.8. Please reinstall the torch_sparse that matches your PyTorch install.
  warnings.warn(f"An issue occurred while importing 'torch-sparse'. "
Traceback (most recent call last):
  File "preprocess.py", line 7, in <module>
    from dataset import get_dataset, get_largest_cc, get_eigh, Transd2Ind, DataGraphSAINT
  File "/media/visionx/monica/project24/GDEM/dataset.py", line 23, in <module>
    from utils import normalize_adj
  File "/media/visionx/monica/project24/GDEM/utils.py", line 14, in <module>
    from torch_sparse import SparseTensor
  File "/home/visionx/anaconda3/envs/gdem/lib/python3.8/site-packages/torch_sparse/__init__.py", line 32, in <module>
    raise RuntimeError(
RuntimeError: Detected that PyTorch and torch_sparse were compiled with different CUDA versions. PyTorch has CUDA version 12.1 and torch_sparse has CUDA version 11.8. Please reinstall the torch_sparse that matches your PyTorch install.
 

原因分析与解决

        一开始看到这个的时候我是懵逼的，为什么呢？我明明按照requirements.txt安装了

pip install -r requirements.txt

         requirements.txt内容如下：

deeprobust==0.2.9
gdown==4.7.3
networkx==3.2.1
numpy==1.26.3
ogb==1.3.6
pandas==2.1.4
scikit-learn==1.3.2
scipy==1.11.4
torch==2.1.2
torch_geometric==2.4.0
torch-sparse==0.6.18

        但首先是报错如下图（解决办法可以看我的另一篇博客：点击这里​​​​​​​）：

        解决完上面这个问题之后呢，又会出现新的问题：RuntimeError: PyTorch has CUDA version 12.1 and torch_sparse has CUDA version 11.8，显然就是版本version不对,但是为什么会这样呢？我们从pytorch故那王找到了答案： 

        我们会发现同样是torch2.1.2，但是却对应了不同的cuda版本，也就是我们在pip install -r requirements.txt的时候torch里的2.1.2并不是cuda11.8的而是cuda12.1的，这个大家也熟，pip install一般就是说安装最新版本的，所以出了问题。那，怎么解决呢？

        很多人的想法就是直接：

pip install torch==2.1.2 torchvision==0.16.2 torchaudio==2.1.2 --index-url https://download.pytorch.org/whl/cu118

        但别忘了，pip install是安装最新的，这个版本比我们现在已有的旧，所以安装不上！！！

        所以需要先卸载：

pip uninstall torch

        然后再安装：

pip install torch==2.1.2 torchvision==0.16.2 torchaudio==2.1.2 --index-url https://download.pytorch.org/whl/cu118

        这样就ok了！ 

相关链接

Previous PyTorch Versions | PyTorchInstalling previous versions of PyTorchhttps://pytorch.org/get-started/previous-versions/conda install 和pip install有什么区别？-CSDN博客文章浏览阅读1k次，点赞3次，收藏12次。本篇为分享贴，截图部分选自，部分选自，文字内容是结合自己实践进行总结。_conda installhttps://blog.csdn.net/BetrayFree/article/details/132115996?ops_request_misc=%257B%2522request%255Fid%2522%253A%2522171652690116800222883586%2522%252C%2522scm%2522%253A%252220140713.130102334.pc%255Fblog.%2522%257D&request_id=171652690116800222883586&biz_id=0&utm_medium=distribute.pc_search_result.none-task-blog-2~blog~first_rank_ecpm_v1~rank_v31_ecpm-1-132115996-null-null.nonecase&utm_term=pip%20install&spm=1018.2226.3001.4450

完结撒花

        就让一切随风～