[BUG] Torch is not available in Docker image but used in CLX sample notebooks

[rapidsai]

[BUG] Torch is not available in Docker image but used in CLX sample notebooks

[drabastomek]

@efajardo-nv Thanks for the reply -- I will install from conda. It is, however, a bit misleading to put the notebooks for clx if the clx nor torch modules are not included.

[BartleyR]

This is a known issue. PyTorch 1.5 has a hard dependency on mkl, and the RAPIDS container uses the nomkl package to prevent installation of this. @efajardo-nv is working on deploying a separate CLX container that solves this short term. We’re also working on removing PyTorch as a hard dependency.

[drabastomek]

Thanks @BartleyR! Glad to hear you're seeing this too. I was quite surprised seeing this yesterday when I tried to run CLX from within the container.

[efajardo-nv]

@drabastomek The rapidsai and rapidsai-nightly images don't include CLX. You would have to either conda install or build from source on those containers. I just tested this on rapidsai/rapidsai:0.14-cuda10.1-runtime-ubuntu18.04-py3.7. You could then install CLX using:

conda install -c rapidsai -c nvidia -c pytorch -c conda-forge -c defaults clx

You would then have to install the dependencies for the notebook you want to run. For the cybert notebook, I had to install pytorch-transformers and seqeval:

pip install pytorch-transformers seqeval[gpu]

You will have PyTorch 1.3.1 with the above steps. As @BartleyR mentioned, we're working on CLX Docker images which will have PyTorch 1.5.0 and everything working out of the box.

[BartleyR]

@efajardo-nv Thanks for the reply -- I will install from conda. It is, however, a bit misleading to put the notebooks for clx if the clx nor torch modules are not included.

Yeah, you’re right. We will make that more clear/straightforward or look at removing and moving.

[BartleyR]

@drabastomek We're going to address this with a CLX image which will have all the dependencies. #201

[drabastomek]

Awesome! Looking forward to testing it.