CUDA 运行时错误：哪个 Cuda 版本与使用 BERT-NER 运行 NER 任务兼容答案

【问题标题】：CUDA Runtime Error: Which Cuda version is compatible to run NER task using BERT-NERCUDA 运行时错误：哪个 Cuda 版本与使用 BERT-NER 运行 NER 任务兼容
【发布时间】：2020-09-29 09:21:09
【问题描述】：

我已经设置了所有安装在我的虚拟机上的需求包，我发现没有安装 nvidia GPU 驱动程序，在需求中没有 nvidia GPU 驱动程序安装说明，我想知道哪个 cuda 版本和它兼容需要哪个 nvidia 驱动程序也解决以下错误。

Github 链接：github

错误日志：

  File "run_ner.py", line 594, in <module>
    main()
  File "run_ner.py", line 489, in main
    loss = model(input_ids, segment_ids, input_mask, label_ids,valid_ids,l_mask)
  File "/home/pt3_gcp/BERT-NER/ber_ner/lib/python3.7/site-packages/torch/nn/modules/module.py", line 547, in __call__
    result = self.forward(*input, **kwargs)
  File "run_ner.py", line 35, in forward
    valid_output = torch.zeros(batch_size,max_len,feat_dim,dtype=torch.float32,device='cuda')
  File "/home/pt3_gcp/BERT-NER/ber_ner/lib/python3.7/site-packages/torch/cuda/__init__.py", line 178, in _lazy_init
    _check_driver()
  File "/home/pt3_gcp/BERT-NER/ber_ner/lib/python3.7/site-packages/torch/cuda/__init__.py", line 99, in _check_driver
    http://www.nvidia.com/Download/index.aspx""")
AssertionError: 
**Found no NVIDIA driver on your system. Please check that you
have an NVIDIA GPU and installed a driver from
http://www.nvidia.com/Download/index.aspx
**

从以下链接安装最新的 cuda 版本后， cuda我收到以下错误，

06/04/2020 07:38:40 - INFO - __main__ -   ***** Running training *****
06/04/2020 07:38:40 - INFO - __main__ -     Num examples = 14041
06/04/2020 07:38:40 - INFO - __main__ -     Batch size = 32
06/04/2020 07:38:40 - INFO - __main__ -     Num steps = 2190
Epoch:   0%|                                                                                 | 0/5 [00:00<?, ?it/sTHCudaCheck FAIL file=/pytorch/aten/src/THC/THCGeneral.cpp line=50 error=38 : no CUDA-capable device is detectedt/s]
Traceback (most recent call last):
  File "run_ner.py", line 594, in <module>
    main()
  File "run_ner.py", line 489, in main
    loss = model(input_ids, segment_ids, input_mask, label_ids,valid_ids,l_mask)
  File "/home/pt3_gcp/.local/lib/python3.7/site-packages/torch/nn/modules/module.py", line 547, in __call__
    result = self.forward(*input, **kwargs)
  File "run_ner.py", line 35, in forward
    valid_output = torch.zeros(batch_size,max_len,feat_dim,dtype=torch.float32,device='cuda')
  File "/home/pt3_gcp/.local/lib/python3.7/site-packages/torch/cuda/__init__.py", line 179, in _lazy_init
    torch._C._cuda_init()
RuntimeError: cuda runtime error (38) : no CUDA-capable device is detected at /pytorch/aten/src/THC/THCGeneral.cpp:50

【问题讨论】：

安装Nvidia驱动后是否重启过虚拟机？
@MichaelJungo 是的
@MichaelJungo 从 sudo apt-get install cuda-driver-440 安装最新的 cuda 驱动程序时是否会出现此错误？或者哪个驱动版本更可取？
$ nvidia-smi NVIDIA-SMI 失败，因为它无法与 NVIDIA 驱动程序通信。确保已安装并运行最新的 NVIDIA 驱动程序。
是的，试过了，但还是一样，$ nvidia-smi NVIDIA-SMI 失败了，因为它无法与 NVIDIA 驱动程序通信。确保已安装并运行最新的 NVIDIA 驱动程序。

标签： pytorch named-entity-recognition huggingface-transformers bert-language-model

【解决方案1】：

前段时间我也遇到过同样的问题。为我修复了以下命令！

如果您有多个安装，并且您现在可能已经尝试了很多东西，这就是一个问题。基本上删除所有内容

sudo apt-get purge nvidia-*
sudo apt-get remove nvidia-cuda-toolkit
sudo apt autoremove --purge cuda-10-0 // you might have a different version, check it git cuda --version

同时删除用户群中的现有文件

rm -rf /usr/local/cuda* // anything related to cuda
rm -rf /usr/local/nvidia* // anything related to nvidia

现在，终于全新安装了

sudo apt-get update // update your packages

sudo apt search nvidia-driver  // to get the latest version of the driver. After finding out the latest version, install it with

sudo apt install nvidia-driver-450 (or any other number, depending on the latest version)

安装后必须重启！

sudo reboot

当你回来时，nvidia-smi 应该可以工作，你的 gpus 也应该工作

【讨论】：