Support for OPSD-V

Hi FastGen team,

Thank you for releasing this great framework!

We recently released [OPSD-V: On-Policy Self-Distillation for Post-Training Few-Step Autoregressive Video Generators](https://arxiv.org/abs/2607.08766). OPSD-V performs self-distillation under the student's actual inference-time rollout and KV cache, helping reduce error accumulation and motion degradation in long autoregressive video generation while preserving fast few-step inference.

Since FastGen already supports methods such as Self-Forcing and CausVid, we believe OPSD-V could also be relevant to this framework. Would you be interested in supporting OPSD-V or discussing a possible integration? We would be happy to contribute and provide implementation details.

• [Paper](https://arxiv.org/abs/2607.08766)
• [Project Page](https://meigen-ai.github.io/OPSD-V/)
• [Code](https://github.com/MeiGen-AI/OPSD-V)

Thank you!

Start by reading the linked OPSD-V paper, project page, and implementation, then compare its inference-time rollout and KV-cache approach with FastGen's existing Self-Forcing and CausVid support. Done would require an agreed integration scope and working OPSD-V support, but this issue does not identify files, tests, or a concrete acceptance criterion.