NVlabs/LongLive
An NVFP4 parallel infrastructure system for long video generation supporting both training and real-time inference.

LongLive 2.0 is an infrastructure project for AI video generation that leverages NVIDIA FP4 quantization and parallelism techniques. It provides optimized kernels for rotary position encoding and adaptive layer normalization via Triton, along with KV-cache synchronization and in-place quantization updates. The system supports teacher-forcing training and DMD distillation workflows for text-to-video models like Wan2.2-TI2V-5B, aiming to enable real-time generation of long-form video content.