RedCoast

Red Coast (redco) is a lightweight and user-friendly tool designed to automate distributed training and inference for large models while simplifying the ML pipeline development process without necessitating MLSys expertise from users.

RedCoast supports Large Models + Complex Algorithms, in a lightweight and user-friendly way:

• Large Models beyond Transformers, e.g, Stable Diffusion, etc.
• Complex algorithms beyond cross entropy, e.g., Meta Learning, DP Training, etc.

With RedCoast, to define a ML pipeline, only three functions are needed:

• Collate function: convert raw data into model inputs (e.g., text tokenization);
• Loss function: execute the model and compute loss (e.g., cross-entropy);
• Predict function: run the model and deliver outcomes (e.g., beam search).

Redco automates all the remaining of pipeline execution such as data and model parallelism, multi-host related processing, distributed checkpointing, randomness controlling, logging, etc.