LLaMA2-Accessory is an open-source toolkit that supports the complete development cycle of LLMs and multimodal LLMs, including pretraining, finetuning, and deployment. The toolkit includes SPHINX, a versatile multimodal large language model that combines diverse training tasks and visual embeddings, along with tools like Large-DiT for image generation tasks. It extends from LLaMA-Adapter with additional features for training and serving models.