We currently run our large scale experiments on the Stability AI HPC cluster. This subdirectory features a few helpful scripts that can help you get up and running on the cluster.
If you believe you need access to the cluster for your work please reach out to the core team on Discord.
-
Install Miniconda - installs miniconda for your cluster environment.
-
Create Environment - creates a basic conda environment for experiments.
- Creates a conda environment at the prefix
CONDA_ENV_PATH
path.Using the positional argument passed into the script
- Clones
chemnlp
into your personal clusterUSER
directory. - Installs the current revision of the
chemnlp
repository and dependencies that are in your personal directory into the conda environment.
# general case source experiments/scripts/stability-cluster/env_creation.sh where/to/store/conda where/to/build/conda/from/ # for creating a personal environment source experiments/scripts/stability-cluster/env_creation.sh jack/ jack/
- Creates a conda environment at the prefix
-
Running Experiment - runs a GPT-NeoX training pipeline
- creates a conda environment using the
env_creation.sh
script. - runs the GPT-NeoX
train.py
script using the user configurationas GPT-NeoX configurations can be combined, the PEFT configurations are held separately to the full model training and cluster configurations
# general case sbatch experiments/scripts/stability-cluster/sbatch_run.sh where/to/store/conda where/to/build/conda/from/ <cluster-config-name.yml> <training-config-names.yml> # for typical small model finetuning experiments sbatch experiments/scripts/stability-cluster/sbatch_run.sh experiments/my-experiment jack cluster_setup.yml 160M.yml # for typical small model soft-prompt experiments sbatch experiments/scripts/stability-cluster/sbatch_run.sh experiments/my-experiment jack cluster_setup.yml 160M.yml soft_prompt.yml
To interact with WandB services you need to authenticate yourself as per the Stability HPC guidelines to append a username + password to your .netrc file.
- creates a conda environment using the