FastVideo: Open-Source Framework Accelerates Large Video Diffusion Models
2024-12-17
FastVideo is an open-source framework designed to accelerate inference and training of large video diffusion models. Supporting models like FastMochi and FastHunyuan, it achieves up to 8x faster inference speeds. The framework leverages knowledge distillation techniques and supports distillation, fine-tuning, and inference for video DiT models based on PCM. Furthermore, FastVideo features scalable training capabilities with FSDP, sequence parallelism, and selective activation checkpointing, while enabling memory-efficient fine-tuning.