PyTorch + CUDA 7.5 错误

Question

I have non-sudo access to a machine with NVIDIA GPUs and CUDA 7.5 installed. I installed PyTorch with CUDA 7.5 support, which seems to have worked:

>>> import torch
>>> torch.cuda.is_available()
True

To get some practice, I followed tutorial for machine translation using RNNs . When I set USE_CUDA = False and the CPUs are used, everything works quite alright. However, when want to utilize the GPUs with USE_CUDA = True I get the following error:

Traceback (most recent call last):
  ...
  File "seq2seq.py", line 229, in train
    encoder_output, encoder_hidden = encoder(input_variable[ei], encoder_hidden)
  File "/.../python2.7/site-packages/torch/nn/modules/module.py", line 206, in __call__
    result = self.forward(*input, **kwargs)
  File "seq2seq.py", line 144, in forward
    output, hidden = self.gru(embedded, hidden)
  File "/.../python2.7/site-packages/torch/nn/modules/module.py", line 206, in __call__
    result = self.forward(*input, **kwargs)
  File "/.../python2.7/site-packages/torch/nn/modules/rnn.py", line 91, in forward
    output, hidden = func(input, self.all_weights, hx)
  ...  
  File "/.../python2.7/site-packages/torch/backends/cudnn/rnn.py", line 42, in init_rnn_descriptor
    cudnn.DropoutDescriptor(handle, dropout_p, fn.dropout_seed)
  File "/usr/lib/python2.7/ctypes/__init__.py", line 383, in __getitem__
    func = self._FuncPtr((name_or_ordinal, self))
AttributeError: python: undefined symbol: cudnnCreateDropoutDescriptor
Exception AttributeError: 'python: undefined symbol: cudnnDestroyDropoutDescriptor' in <bound method DropoutDescriptor.__del__ of <torch.backends.cudnn.DropoutDescriptor object at 0x7fe540efec10>> ignored

I've tried to use Google to search for that error but got no meaningful results. Since I'm rather a newbie with PyTorch and CUDA, I have no idea how to go on from here. The full setup is Ubuntu 14.04, Python 2.7, CUDA 7.5.

Answer 1

As stated in the comments: your error is with outdated CUDNN, and can be resolved by upgrading.

Install current versions of CUDA, CUDNN, and PyTorch, then you'll be fine.