Addition of a not-strictly-block-diffusion model

[settheworldonfireiii]

The dLLM framework currently prioritizes block-diffusion models (LLaDA 2.0, SDAR), and 'non-block diffusion LLMs' is unchecked on the roadmap. Do you consider adding Fast-dLLM v1's version of bidirectional LLaDA-8B / Dream-7B with approximate-KV-cache + confidence-based decoding thresholding with gptq_marlin quantization or is it out-of-scope? And if it's in scope, is anyone already working on it, and what's the approximate timeline/ETA?