Motivation

Contritbutions are welcome!

Focus

• Diverse model support with training and inference
• Launch RL training infra and an effective training recipe
• New Distillation Recipes
• Quantization Support
• Improve training performance and infra
• CI/CD

Diverse model support with training and inference

• Longcat
  • Add LongCat T2V (Base, Distillation and Refinement) Support to FastVideo #883
• Cosmos2.5
  • [feat]: add COSMOS 2.5 DiT implementation #897
• Hunyuanvideo Training Pipeline
  • [Preprocess] [feat] Support HunyuanVideo Model #754
• Hunyuanvideo-v1.5
• Hunyuanvideo-v1.5 training
• Lucy-Edit-Dev
• LTX
  • Initial DiT and VAE changes for LTX inference #783
  • LTX specific pipeline stages #876

New Distillation Recipes

• Self-Forcing for Wan2.2 MoE
  • [feat] Add inference for MoE SF #880
• Rolling-Forcing support

Launch RL training infra and an effective training recipe

• FlowGRPO
  • in progress

Preprocessing & Dataset

• Enable training with no precomputed latents

Training Infra

• Activation offloading
• Multi-resolution and variable length bucketing for long video input compatibility
  • [feat] Support sequence packing and shard after pachification for USP #894
• LoRA extraction, verification, and comparison scripts
  • Add LoRA extraction, verification, and comparison scripts #865
• USP for Self-Forcing
  • in progress
• Ring Attention

CI/CD

• Add MFU into CI
  • [feat] training mfu calculation scripts #871
• Add FVD into CI
  • [feat] Add fvd implementation #923

Related resources

No response