因为pytorch1.11.0与cuda版本的关系 需要用到cuda11.5

否则报错

"addmm_sparse_cuda" not implemented for Half

cuda11.5.0及以前的版本不会出现这个问题

因此重新安装，步骤如下：

安装CUDA-11.5.0

wget https://developer.download.nvidia.com/compute/cuda/11.5.0/local_installers/cuda_11.5.0_495.29.05_linux.run
sudo sh cuda_11.5.0_495.29.05_linux.run

修改路径~/.bashrc

export PATH=$PATH:/usr/local/cuda/bin
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda/lib64
export CUDA_HOME=/usr/local/cuda

安装cuDNN-8.8.0

下载路径：

cuDNN Archive | NVIDIA Developer

sudo dpkg -i cudnn-local-repo-ubuntu2204-8.8.0.121_1.0-1_amd64.debsudo apt-get updatesudo apt-get install libcudnn8=8.8.0.121-1+cuda11.8sudo apt-get install libcudnn8-dev=8.8.0.121-1+cuda11.8sudo apt-get install libcudnn8-samples=8.8.0.121-1+cuda11.8

安装完成后可查看
 

cat /usr/include/cudnn_version.h | grep CUDNN_MAJOR -A 2#define CUDNN_MAJOR 8
#define CUDNN_MINOR 8
#define CUDNN_PATCHLEVEL 0
--
#define CUDNN_VERSION (CUDNN_MAJOR * 1000 + CUDNN_MINOR * 100 + CUDNN_PATCHLEVEL)/* cannot use constexpr here since this is a C-only file */

表示输出正常！

正确的计算代码应该是

import torch
a = torch.randn(3,2).half().cuda()
i = torch.LongTensor([[0, 1, 1],  [2, 0, 2]]) 
v = torch.FloatTensor([3, 4, 5]) 
b = torch.sparse.FloatTensor(i, v, torch.Size([2,3])).half().cuda()
b = b.to_sparse_csr()
c = torch.spmm(b, a)