This repository adds vLLM support to the IndexTTS text-to-speech system. It re-implements the GPT model inference using vLLm, achieving roughly 3x speed improvement in decode throughput (90 to 280 tokens/s) and reduced real-time factor (0.3 to 0.1) on RTX 4090. The project supports IndexTTS v1, v1.5, and v2 variants, and provides API interfaces alongside a web UI for deployment.