vllm-project /vllm-omniPython7.6K2.1KA framework for efficient model inference with omni-modality modelsdiffusioninferencemodel-servingpytorch