FMod

Foundation Models and Super Resolution with MERRA2 training / fine tuning.

Installation

OSX Local Development Installation

* conda create -n fmod python=3.10 mamba -c conda-forge 
* conda activate fmod
* mamba install pytorch torchvision -c pytorch
* mamba install -c dglteam dgl 
* mamba install -c conda-forge s3fs tqdm pydantic ipython h5py h5netcdf matplotlib scipy netCDF4 ipympl jupyterlab ipykernel ipywidgets numpy xarray dask  pandas typing_extensions
* pip install hydra-core --upgrade
* pip install --no-deps nvidia-modulus nvidia-modulus-sym

Production Server Installation

Prerequesites:
- Must have mpicc in PATH (e.g. /app/openmpi/platform/x86_64/rhel/8.9/4.1.2_gcc-12.1.0/bin/mpicc )
Install Packages:
- conda create -n fmod python=3.10 cuda-python -c nvidia
- conda activate fmod
- pip install ninja
- pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu121
- pip install lightning-bolts
- pip install -v --disable-pip-version-check --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" pytorch-extension
- pip install nvidia-modulus[all] nvidia-modulus-sym
- pip install tensorly tensorly-torch netCDF4 h5py h5netcdf parameterized cartopy ipympl opt_einsum hydroeval
- cd torch-harmonics; pip install .
Alternate pytorch-extension installation
- cd /explore/nobackup/projects/ilab/software/pytorch-extension-0.2
- python setup.py egg_info
- python setup.py bdist_wheel
- python setup.py install
- python setup.py clean
For graphcast:
- pip install dgl -f https://data.dgl.ai/wheels/cu121/repo.html
- pip install dglgo -f https://data.dgl.ai/wheels-test/repo.html
- pip install wandb pydantic quadpy orthopy ndim gdown netCDF4 h5py h5netcdf mlflow torch-harmonics
[GraphCast]: Install pymesh from source: https://pymesh.readthedocs.io/en/latest/installation.html
[SFNO]: Install apex from source: https://github.com/NVIDIA/apex

Recommend Projects