Multi-Block Diffusion LMs Introducing MBD-LMs: a post-training recipe that unlocks parallel block decoding in diffusion language models. MBD-LLaDA2-Mini jumps to 6.19 tokens/forward, a 78% speedup,…

3/ Add a new strategy? Modify the request state machine (or don't, if it's sampler-only). No kernel rewrites. No serving system rewrites.