FastVideo is a unified framework designed to accelerate large video diffusion models, targeting researchers and developers working with state-of-the-art open video DiTs. It offers significant inference speedups and efficient fine-tuning capabilities, enabling faster iteration and deployment of video generation models.

How It Works

FastVideo leverages techniques like Sliding Tile Attention (STA) and distillation to achieve up to 8x inference speedup on models like FastHunyuan and FastMochi. It supports efficient training and fine-tuning using methods such as FSDP, sequence parallelism, selective activation checkpointing, LoRA, precomputed latents, and precomputed text embeddings, aiming for near-linear scaling up to 64 GPUs.

Quick Start & Requirements

• Installation: Instructions are available in the README.
• Prerequisites: Python, PyTorch. Specific CUDA versions or hardware requirements for optimal performance are not explicitly detailed but implied for advanced features.
• Resources: Single 4090 inference is supported for FastHunyuan.
• Links: Documentation, Inference Quick Start, V1 Inference API Guide (Coming soon!)

Highlighted Details

• Supports distillation/finetuning/inference for Mochi and Hunyuan video DiTs.
• Introduced Sliding Tile Attention for up to 3.4x speedup on StepVideo.
• Offers memory-efficient fine-tuning with LoRA.
• Scalable training with FSDP and sequence parallelism.

Maintenance & Community

• Supported by MBZUAI and Anyscale.
• Community support via Slack.
• Development is ongoing with a clear Development Plan.

Licensing & Compatibility

• The project's license is not explicitly stated in the README.
• Code is reused from various projects, some of which may have different licenses.

Limitations & Caveats

The project is described as "highly experimental" and "dev in progress." The V1 Inference API Guide is still pending. Support for additional models and optimization features is listed as under development.