Hi HN, this is my research project that allow people to locally deploy MoE Diffusion LLMs more efficiently. With this method, you can fit a 100B LLaDA2.0-flash model into a PC with a RTX5090 and run it faster than other methods.