dLLM Notes

Paper notes · arXiv 2025

Diffusion LLMs Can Do Faster-Than-AR Inference via Discrete Diffusion Forcing (D2F)

Xu Wang, Chenkai Xu, Yijie Jin, Jiachun Jin, Hao Zhang, Zhijie Deng

Block diffusion (M08) gives an exact KV cache, but blocks run one after another. D2F keeps the block-causal mask and adds inter-block parallelism: the next block starts while the previous one is still being decoded. Two parts:

What “forcing” means. In teacher forcing, the usual way to train AR models and block diffusion, each block is fed a clean prefix. Diffusion forcing (Chen et al., 2024, from video) feeds a noised one: masks rise from left to right, so earlier blocks are partly masked, but less than the current one. That is exactly what the pipeline looks like at inference, so the next block can start before the previous one is finished. A teacher-forced model has never seen an unfinished prefix, so it has to wait.

Training. Pick a target block. The student sees only the blocks up to it, and those are still noisy. Compare with block diffusion, which gives the earlier blocks clean.

target block
block 1 · t1block 2 · t2block 3 · t3block 4 · t4Teacherpre-trained dLLM, full attention (frozen)dLLMsarenewofmodelideasStudentD2F, block-causal attentiondLLMsarenewofmodelideasFor contrastblock diffusion: teacher forcing, clean prefixdLLMsarenewofmodelideas

Blocks are masked more and more from left to right (t1 < t2 < t3 < t4). The student (block-causal) must match the teacher (full attention, frozen) on the masked tokens of the target block, but sees only blocks up to it, with their masks still in. Block diffusion's teacher forcing would show it a clean prefix instead.

Inference. Step through Algorithm 2 and compare with decoding one block at a time.

block 1block 2block 3block 4+tok

4 blocks of 6 tokens, one row per step. A new block is appended, semi-activated, once the last block passes τ_add. It becomes fully activated once its predecessor passes τ_act. Finished blocks are cached exactly and never recomputed. Toy confidence: a fully activated block commits one token per step, a semi-activated one every other step. In the paper, fully activated blocks commit every token above τ_conf, often several per step.

Loading PDF… open it directly

PDF reproduced unmodified under CC BY 4.0. Highlights and notes are mine.