Skip to content

Latest commit

 

History

History
29 lines (19 loc) · 833 Bytes

File metadata and controls

29 lines (19 loc) · 833 Bytes

Style Guide

Follow these standards so that each example can be run consistently.

Secrets

Secrets should be consistently named and used.

  • huggingface-secret should be used for the huggingface token. This secret should always have the HF_TOKEN environment variable set.
  • wandb-secret should be used for Weights & Biases. This secret should always be optional i.e. the script should work without it.

File Structure

Most of the examples should be structured as follows:

multinode-training-guide/
    foo/
        modal_train.py
        train.py
        README.md
        ...

modal_train.py should be the entrypoint for training. In most cases, this will use torchrun to launch copies of train.py.

Dependencies

Dependencies in images should always be pinned, and python_version should be set.