Any chance of upgrading from CUDA 11.0 to 11.1? (installing PyG with pip takes >1 hour)

This is mostly related to Pytorch Geometric. From what i understand, issue #675 is unresolved due to issues with GPU tests. Fair enough, so I tried to use the following command on Kaggle Notebooks:

!pip install torch-scatter torch-sparse torch-cluster torch-spline-conv torch-geometric -f https://data.pyg.org/whl/torch-1.9.0+cu110.html

Unfortunately, this takes over 1 hour to install the packages, mostly building the wheel files for torch-sparse, torch-cluster, torch-spline-conv (some of which are required when importing torch_geometric). I believe that this is something related to the CUDA 11.0 version of these packages, given that I had a similar problem locally on my machine. On the other hand, if you run the following on Google Colab (currently with CUDA 11.1) or my own machine with CUDA 11.1:

!pip install torch-scatter torch-sparse torch-cluster torch-spline-conv torch-geometric -f https://data.pyg.org/whl/torch-1.9.0+cu111.html

, it takes only a few seconds.

Installing the CPU-only version of PyG does work in Kaggle Notebooks, but it’s inefficient to the point that waiting 1 hour at the start may still better in the long run.

[rosbo]

It actually entails a lot given the large number of packages installed on our image.

@djherbis is hard at work trying to complete this upgrade. You can follow his progress on this PR: https://github.com/Kaggle/docker-python/pull/1182

[jaimergp]

The release notes for Google Deep Learning containers even state:

February 28, 2022 M90 Release

• CUDA has been upgraded from 11.3.0 to 11.3.1 to address some NCCL issues.
• GPU TensorFlow containers are available and have a significantly smaller size.
• TensorFlow 2.7 containers are re-released.

So it has been available for a while, it looks like?