conda not installing the correct nightly version

[anmolsjoshi]

OS: Ubuntu 18.04
Python Version: 3.7

Using the installation instructions found here, conda is installing the wrong version of torchvision despite specifying cpuonly.

conda install pytorch torchvision cpuonly -c pytorch-nightly

Output of the above command is -

(base) ➜  ~ conda activate testing
(testing) ➜  ~ conda install pytorch torchvision cpuonly -c pytorch-nightly
Collecting package metadata (current_repodata.json): done
Solving environment: done

## Package Plan ##

  environment location: /home/joshi/miniconda3/envs/testing

  added / updated specs:
    - cpuonly
    - pytorch
    - torchvision


The following NEW packages will be INSTALLED:

  _pytorch_select    pkgs/main/linux-64::_pytorch_select-0.2-gpu_0
  blas               pkgs/main/linux-64::blas-1.0-mkl
  cffi               conda-forge/linux-64::cffi-1.13.2-py37h8022711_0
  cpuonly            pytorch-nightly/noarch::cpuonly-1.0-0
  cudatoolkit        pkgs/main/linux-64::cudatoolkit-10.0.130-0
  cudnn              pkgs/main/linux-64::cudnn-7.6.5-cuda10.0_0
  freetype           conda-forge/linux-64::freetype-2.10.0-he983fc9_1
  intel-openmp       pkgs/main/linux-64::intel-openmp-2019.4-243
  jpeg               conda-forge/linux-64::jpeg-9c-h14c3975_1001
  libgfortran-ng     conda-forge/linux-64::libgfortran-ng-7.3.0-hdf63c60_5
  libpng             conda-forge/linux-64::libpng-1.6.37-hed695b0_0
  libtiff            conda-forge/linux-64::libtiff-4.1.0-hc3755c2_3
  lz4-c              conda-forge/linux-64::lz4-c-1.8.3-he1b5a44_1001
  mkl                pkgs/main/linux-64::mkl-2019.4-243
  mkl-service        conda-forge/linux-64::mkl-service-2.3.0-py37h516909a_0
  mkl_fft            conda-forge/linux-64::mkl_fft-1.1.0-py37hc1659b7_1
  mkl_random         conda-forge/linux-64::mkl_random-1.1.0-py37hb3f55d8_0
  ninja              conda-forge/linux-64::ninja-1.10.0-hc9558a2_0
  numpy              pkgs/main/linux-64::numpy-1.18.1-py37h4f9e942_0
  numpy-base         pkgs/main/linux-64::numpy-base-1.18.1-py37hde5b4d6_1
  olefile            conda-forge/noarch::olefile-0.46-py_0
  pillow             conda-forge/linux-64::pillow-7.0.0-py37hefe7db6_0
  pycparser          conda-forge/noarch::pycparser-2.19-py_2
  pytorch            pkgs/main/linux-64::pytorch-1.3.1-cuda100py37h53c1284_0
  six                conda-forge/linux-64::six-1.14.0-py37_0
  torchvision        pkgs/main/linux-64::torchvision-0.4.2-cuda100py37hecfc37a_0
  zstd               conda-forge/linux-64::zstd-1.4.4-h3b9ef0a_1


Proceed ([y]/n)?

If you see above cudatoolkit is being installed, along with the cuda versions of pytorch and torchvision.

How should I resolve this?

[anmolsjoshi]

Closing in favor of pytorch/pytorch#33103

[fmassa]

Thanks for the report!

I'm looking into fixing this

[fmassa]

This will hopefully be fixed with #1917

[fmassa]

Should be fixed now, let us know if you still have issues

[jamestwebber]

I'm having this issue again, when trying to create a Dockerfile with pytorch-nightly. In my case I'm trying to install the GPU version of pytorch but I am getting 1.3.1 and no GPUs detected :(

Is torchvision actually required for pytorch to work? I'm not using it for anything as far as I know. If I omit torchvision I get pytorch 1.5.0, which seems weird since it's behind the stable version? And I still don't have cuda available.

edit: It appears that this was fixed today with an update to the 1.7.0 rc.