A scalable, modular, multimodal training pipeline
π Latest UpdatesΒΆ
- 2025/09:
First version of the MultiMeditron pipeline!
β¨ OverviewΒΆ
MultiMeditron is a scalable and modular pipeline to train multimodal models.
Features:
Modular modality: Designed to be easily expanded to any types of modality
Scalable: Scalable to multinode training and efficient GPU memory usage using Deepspeed
Configurable: Trainings can be configured using a single YAML configuration file
Easy to install: We provide Docker images for easier reproducibility