.cuda.is_available() returns false after updating

1

Hey guys, so I am using Ubuntu 18.04 and CUDA 10.2. Installed pytorch and torchvision by conda. Initially.cuda.is_available() returns True. Now after updating it returns false… Any good solution besides reinstalling?

ps I am a big fan of conda, so please suggest a solution under conda environment, not something like compiling binaries. And idk when pytorch for cuda 10.2 is gonna be out, so won’t count on that…

2

Could you post the log from the conda installation?
Maybe the wrong version was picked (cpu instead of e.g. cudatoolkit=10.1).

3

I picked cudatoolkit=10.2 at initial installation, but updating didn’t ask me to pick this option. I’ll post log later.

4

The binaries are shipping with cudatoolkit=9.2 or 10.1.
If you specify 10.2, you might get the package from another source and PyTorch might fall back to an older release.
I just tested this command and got the following suggestion:

## Package Plan ##

  environment location: /home/pbialecki/anaconda3/envs/tmp

  added / updated specs:
    - cudatoolkit=10.2
    - pytorch
    - torchvision


The following packages will be downloaded:

    package                    |            build
    ---------------------------|-----------------
    cudatoolkit-10.2.89        |       hfd86e86_0       364.5 MB
    pytorch-1.0.0              |py3.7_cuda9.0.176_cudnn7.4.1_1       498.7 MB  pytorch
    torchvision-0.2.2          |             py_3          44 KB  pytorch
    ------------------------------------------------------------
                                           Total:       863.2 MB

The following NEW packages will be INSTALLED:

  ninja              pkgs/main/linux-64::ninja-1.9.0-py37hfd86e86_0
  pytorch            pytorch/linux-64::pytorch-1.0.0-py3.7_cuda9.0.176_cudnn7.4.1_1
  torchvision        pytorch/noarch::torchvision-0.2.2-py_3

As you can see, PyTorch would be downgraded to 1.0.0 with CUDA9.0.