Getting Started

Getting Started#

Installation for Single Server/Node Deployment#

If you plan to use CREDIT only for running pretrained models or training on a single server/node, then the standard Python install process will install both CREDIT and all necessary dependencies, including the right versions of PyTorch and CUDA, for you. If you are running CREDIT on the Casper system, then the following instructions should work for you.

Create a minimal conda or virtual environment.

conda create -n credit python=3.12
conda activate credit

Important

When installing PyTorch on Linux, the default option (v2.11) uses CUDA 13, which is currently not compatible with the CUDA driver on Casper. Add --extra-index-url https://download.pytorch.org/whl/cu126 to your pip install commands to install a PyTorch version compiled with CUDA 12.6, which will work on Casper NVIDIA GPUs from the V100s to the H100s.

If you want to install the CPU-only version of PyTorch, use --extra-index-url https://download.pytorch.org/whl/cpu.

If you want to install the latest stable release from PyPI:

pip install miles-credit

If you want to install the main development branch:

git clone https://github.com/NCAR/miles-credit.git
cd miles-credit
# Linux with NVIDIA GPU support
pip install -e . --extra-index-url https://download.pytorch.org/whl/cu126
# Mac
pip install -e .

If you plan to perform development work on CREDIT, install additional dependencies for development.

pip install -e .[develop]

Important

macOS users will need to ensure that the required compilers are present and properly configured before installing miles-credit for versions requiring pySTEPS (miles-credit > 2025.2.0).
See this note in the pySTEPS documentation for details.

Installation on Derecho#

If you want to build a conda environment and install a Derecho-compatible version of PyTorch, run the create_derecho_env.sh script.

git clone https://github.com/NCAR/miles-credit.git
cd miles-credit
./create_derecho_env.sh

This script will create a Derecho-compatible conda environment with a user-specified name taken from the CREDIT_ENV_NAME environment variable (defaults to credit-derecho).

Important

The credit conda environment requires multiple gigabytes of space. Use the gladequota command to verify that you have sufficient space in your home or work directories before installing. You can specify where to install your conda environments in a .condarc file with the section envs_dirs.

If you are finding your home directory to be surprisingly full with minimal installation of files, check your ~/.cache directory and delete everything inside it. This directory is where python installers tend to download their files. You can shift the .cache directory to a location with more space by adding export XDG_CACHE_HOME="/glade/work/$USER/.cache" to your ~/.bashrc file.

Installation from source#

See Installing CREDIT from source for detailed instructions on building CREDIT and its dependencies from source or for building CREDIT on the Derecho supercomputer.

Making sure training and inference work#

To train a basic WXFormer model on Casper or Derecho, run the following command from the miles-credit directory:

credit_train -c config/example-v2026.1.0.yml

This script will train a simple WXFormer model on 1 degree ERA5 data for 10 batches. The model will not be good, but if it completes successfully, you can have some confidence that CREDIT has been set up correctly in your environment.

Next, make sure inference is working by running:

credit_rollout_to_netcdf -c config/example-v2026.1.0.yml

This will perform a short forecast based on ERA5 and will save the outputs to a specified directory.

If you want to run a realtime forecast, first download and regrid GFS data to your domain.

credit_gfs_init -c config/example-v2026.1.0.yml -p 1

The script will create a GFS initialization file in your scratch directory. It will also create a realtime config file config/example-v2026.1.0_realtime.yml. Edit the realtime config file so that data:save_loc_dynamic_forcing is /glade/campaign/cisl/aiml/credit/credit_solar_6h_1deg_era5_mlevel/*.nc. Then you can run a realtime forecast with

credit_rollout_realtime -c config/example-v2026.1.0_realtime.yml -p 1

This script will output 6-hourly netcdf files to your scratch directory and performs interpolation to pressure levels. From there you can view the data with your favorite visualization tool.

Running a pretrained model#

See Prediction Rollouts for more details on how to run one of the pretrained CREDIT models.