Cornstarch
Build, Train, and Run Your Own Multimodal Model
Cornstarch is a multimodal model training framework that allows you to create your own multimodal model from a set of HuggingFace unimodal models, train, or use it.
Documentation Organization
- Getting Started: Instructions on installation and setup.
- Creating a multimodal LLM: How to create a multimodal LLM from unimodal models.
- Preprocessing multimodal inputs: How to preprocess muiltimodal inputs to run a multimodal LLM.
- Training a multimodal LLM: How to train a multimodal LLM using Cornstarch.
- Multimodal LLM Parallelization: How to parallelize a multimodal LLM training.
Research Works
Below is a list of works powered by Cornstarch.
- Oobleck: resilient distributed training using pipeline template
Contact
Insu Jang (insujang@umich.edu)