[Feature Request] Support Stable-DiffCoder Model

[Seas0]

Hello, I am a contributor to the Stable-DiffCoder model, a new open-source code diffusion Large Language Model trained from standard AR model's weight that recently achieved state-of-the-art performance among 8B-scale models.

Given dInfer's position as a leading efficient framework for Diffusion LLMs, I believe integrating Stable-DiffCoder would be a highly valuable addition to the framework's supported model library and our model. I am interested in contributing to this integration and would like to discuss the roadmap.

I am happy to assist with the implementation details like model architecture and sampling tricks.

[DavidBao03]

Hi @Seas0 @zheng-da , this sounds like a great addition! I’ve been following the DLLM space for quite a while and would love to help with this integration. Has anyone started working on a PR yet? If not, I’d love to do it.