VLLM-Omni: A framework for efficient model inference with Omni-modality models

I feel like there's a story behind multimodal architectures not being integrated into the main package by default