Rex Framework Package Docs

Rex Framework is a package for inference with remotely stored model weights. It fetches chunks on demand, keeps residency bounded, and lets the full model stay remote.

This site documents the installed package only. It does not cover private repository internals, ADRs, or development workflow.

What the package gives you

  • Python APIs for loading manifests and running inference.
  • CLI tools for conversion, validation, inspection, serving, benchmarking, and demo runs.
  • Notebook-friendly usage patterns for Kaggle and Google Colab.
  • WebSocket streaming as an explicit backend mode for live weight delivery from rex-serve.
  • Configuration via RexConfig and REX_* environment variables.
  • Manifest hints such as storage_backend: streaming_ws and provenance.base_url for hosted chunks.

Start here

  1. Read the install guide.
  2. Follow the Kaggle notebook guide or Colab guide if you are running in a notebook.
  3. Review streaming if you want live weight delivery through rex-serve.
  4. Use configuration for runtime tuning.

Common package story

pip install "rex-framework[pytorch]"
rex-serve --dir ./rex_output/weights --port 8080

Then point your runtime at the manifest and base URL for the chunk host.