Run Cheetah simulations for RL training on GPU#72
Conversation
| - pip | ||
| - python | ||
| - python-build | ||
| - pytorch-cpu |
There was a problem hiding this comment.
As discussed: looks splendid, feel free to rename this file (remove the -cpu) and also update in the REAMDE.md now.
There was a problem hiding this comment.
Background: I had this initially because I wanted to avoid downloading huge cuda packages for CPU usage, but in the end that package did not work properly for compilation of torch code.
RemiLehe
changed the title
RemiLehe
changed the title
RemiLehe
changed the title
|
|
||
| channels: | ||
| - conda-forge | ||
| - nodefaults |
There was a problem hiding this comment.
I had to remove unused packages here, to fit within the memory quotas of the $HOME folder at NERSC
|
|
||
| 2. Create and activate the environment: | ||
| ```bash | ||
| conda clean --all -y ; pip cache purge |
There was a problem hiding this comment.
I also modified my conda pkg_dir now, because it clutters up $HOME on PM while downloading all the CUDA stuff:
conda config --add pkgs_dirs $PSCRATCH/conda/pkgs
| ```bash | ||
| conda clean --all -y ; pip cache purge | ||
| mkdir -p /global/cfs/cdirs/m558/$(whoami)/sw/perlmutter/rl_training_conda_env | ||
| conda env create --prefix /global/cfs/cdirs/m558/$(whoami)/sw/perlmutter/rl_training_conda_env -f rl-training-conda.yaml |
There was a problem hiding this comment.
I think you also want to set the envs_dirs in the conda config.
What I did not instead of prefix is set a envs_dirs to CFS or PSCRATCH upfront and then just conda env create away - it will then pick that dir by default and conda activate and conda env ... will know where to find it.
2 participants
This adds installation instructions and code changes to run RL training on NERSC