nv-dfm-lib-weather

Weather and climate data adapters for NVIDIA Data Federation Mesh (DFM).

What is DFM?

Data Federation Mesh (DFM) is a Python-based framework for creating and orchestrating complex workflows that process data from various distributed sources and stream results into applications. DFM determines where to run each operation of a data processing pipeline and handles data movement between sites automatically.

Experimental: This library is provided as a collection of examples and starting points for building your own adapters. Before any production use, review and extensively test all adapters in your target environment.

Overview

nv-dfm-lib-weather provides pre-built adapters for weather and climate data sources, AI models, and xarray processing:

Installation

pip install nv-dfm-lib-weather

Or with uv in the monorepo:

uv sync --package nv-dfm-lib-weather

Note: This package depends on earth2studio, which may require additional dependencies depending on your environment. See the Weather Data Library section of the installation guide for details.

Optional Extras

uv sync --package nv-dfm-lib-weather --extra sfno
uv sync --package nv-dfm-lib-weather --extra cbottle

GPU Prerequisites for AI Model Adapters

The SFNO and cBottle adapters perform GPU-accelerated AI inference. Before installing these extras, ensure you have:

• NVIDIA GPU with compute capability ≥ 8.9 and ≥ 40 GB GPU memory. See CUDA GPUs.
• NVIDIA GPU Drivers — see the Driver Installation Guide.
• CUDA Toolkit 12.8 (tested) — see CUDA Downloads.
• PyTorch with CUDA support matching your toolkit version — see pytorch.org. Example:

  pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128
  

• Earth2Studio 0.10.0 — installed automatically as a dependency. See the Earth2Studio docs for model-specific setup details.

Note: The cbottle and sfno extras depend on packages (cbottle, earth2grid, makani) that are not published on PyPI. When installing from source with uv, these resolve automatically via workspace git sources. When installing the published wheel via pip, users must install them manually — a clear error message with instructions is shown at runtime if they are missing.

Adapters

Data Loaders

Xarray Processing

AI Models (optional extras)

DFM Ecosystem

Dependencies

• nv-dfm-core – Core DFM framework
• nv-dfm-lib-common – Shared schemas (GeoJsonFile, TextureFile, etc.)
• earth2studio[data] – Data loaders (GFS, ECMWF, HRRR, CMIP6)
• xarray, netCDF4, rioxarray – Data handling

Documentation

• DFM Documentation
• Weather Tutorial – pipelines and adapters for weather data
• Tutorials

License

Apache License 2.0. See the LICENSE file for details.