DeepSpeed-VisualChat is a scalable, efficient, and user-friendly multi-modal training pipeline that leverages a novel multi-modal causal attention mechanism for better alignment of visual and text features. It uses data blending techniques to address the scarcity of interleaved text-and-image inputs in datasets.
The framework trains using a 2B visual encoder from QWen-VL and a 70B language decoder from LLaMA-2, showcasing its extraordinary scalability. DeepSpeed-VisualChat is now open-sourced and encourages community contributions and collaborations. Visit the GitHub page to get started.