RTX3090踩坑实录-在配置pytorch3d遇到的問題
官方安裝:https://github.com/facebookresearch/pytorch3d/blob/v0.3.0/INSTALL.md
archive:https://pytorch.org/get-started/previous-versions/
详细系统环境见后文:
总结:只能pip pytorch,使用cub版本10.0, python3.8+
GeForce RTX 3090 with CUDA capability sm_86 is not compatible with the current PyTorch installation
- 原来的步骤
- 按照原來的文檔指示py=3.7
- pip install torch1.6.0 torchvision0.7.0、
- conda install -c conda-forge -c fvcore fvcore
- conda install -c bottler nvidiacub
- curl -LO https://github.com/NVIDIA/cub/archive/1.10.0.tar.gz
- tar xzf 1.10.0.tar.gz
- export CUB_HOME=$PWD/cub-1.10.0
- conda install pytorch3d -c pytorch3d
- 报错:与3090不兼容
- 参考文档:https://www.cnblogs.com/pikachuuuuu/p/15126905.html
OSError: undefined symbol: free_gemm_select, version libcublasLt.so.11
-
改成py=3.8
-
然后还是按原来的步骤
-
后面改用pip
- pip install torch1.8.1+cu111 torchvision0.9.1+cu111 torchaudio==0.8.1 -f https://download.pytorch.org/whl/torch_stable.html
后面参考的文档https://blog.csdn.net/weixin_40022595/article/details/115359656 - 修改过程
-
pip上面的版本
-
pip install -r requirements.txt
-
pip install fvcore
CUB的release版本用最新的链接:https://github.com/NVIDIA/cub/releases
https://github.com/NVIDIA/cub/archive/refs/tags/1.16.0.tar.gz -
替换下面对应的内容
curl -LO https://github.com/NVIDIA/cub/archive/1.10.0.tar.gz tar xzf 1.10.0.tar.gz export CUB_HOME=$PWD/cub-1.10.0
-
pip install ‘git+https://github.com/facebookresearch/pytorch3d.git’
-
pip install ‘git+https://github.com/facebookresearch/pytorch3d.git@stable’
-
- pip install torch1.8.1+cu111 torchvision0.9.1+cu111 torchaudio==0.8.1 -f https://download.pytorch.org/whl/torch_stable.html
fatal error: cuda_runtime_api.h: 没有那个文件或目录
- 参考文档:https://blog.csdn.net/walker0001/article/details/115460741
- sudo apt-get install nvidia-cuda-toolkit
- 报下面的错
:/usr/local/cuda-11.4/bin/nvcc’: No such file or directory
-
查看自己的环境
这里显示是cuda11.4
nvcc -V与smi信息不一致,但似乎可以,参考这篇:https://blog.csdn.net/sophicchen/article/details/120782209
-
查看环境变量:
export CUDA_HOME=/usr/local/cuda-11.4
gedit ~/.bashrc
查看路径是不是加进去了
-
一堆问题
居然还有这种错误
降cub版本回到10.0