Any suggestions on implementing this in the Wan repo? I've been looking at it the last couple weeks to try to implement it into a fork of my own, but I'm getting stuck on where to work the code change in since the inference/pipeline structure is a bit different than Hunyuan.