I have been playing around a lot with analysing single cell RNA datasets on GPUs. Setting up the environments is kind of a pain so created I a guide to help with the conda envirmoment setup.
You can check it out at: GitHub - Intron7/GPU_SingleCell_Setup: Guides to create Conda environments for single cell analysis that leverage GPU acceleration for faster computing
If want more tools included, I can try to also fit them into the enviroment. So tell me whats missing.
Thanks for this! I think it should work to make a conda env yaml file that works with all of these packages rather than just a guide. I would also consider using mamba in your guide as it’s much faster than conda. Also:
- All scverse packages are on conda, so I would try to use conda-forge distributed packages as much as possible.
- In the case of jax, it can also be installed via conda-forge. I noticed to get jax gpu working via conda it was necessary to install Cudatoolkit Dev :: Anaconda.org
there might be some caveats with the computer’s nvidia driver versions though
I personally have not worked with mamba yet, but if its faster i’ll give it a try. About the conda-forge packages. The Rapids packages are really picky what you can and cannot install with them in a conda env. In Version 22.02 it was still possible to install scanpy via conda-forge together with rapids. In version 22.04 and 22.06 this doesn’t work anymore. I haven’t tested this with release 22.08 yet. The next issue it that rapids and pytorch work with different cuda versions (rapids 11.5, pytorch 11.6) This is why install it in such a way and order. If there is a better way please let me know. Another issue with conda environment is that tensorboard (needed for scvi-tools) requires protobuf 3.19.X and rapids need at least 3.20.1. Everything works to my knowledge if you install protobuf manually after installing scvi-tools. I don’t think conda or mamba would resolve this conflict. I’ll never the less try to create a yaml file to make the setup even easier.