Paper notes · arXiv 2025
Fast-dLLM v2: Efficient Block-Diffusion LLM
Fast-dLLM (v1) sped up LLaDA and Dream without training, but its cache was an approximation. v2 changes the model instead: fine-tune an AR model (Qwen2.5-1.5B/7B) into a block diffusion model, causal across blocks of 32 and bidirectional inside each block. Then:
- Training (§3.2): mask tokens inside each block, train each sample with a mask and its complement, keep the AR token shift. About 1B tokens, against ~500B for Dream.
- Inference (§3.3): earlier blocks are final, so their K/V are an exact cache. Inside the current block, decode sub-blocks of 8 with v1’s confidence threshold and DualCache.
Result: on par with Qwen2.5-7B fine-tuned on the same data with next-token loss (avg 60.3 vs 59.6), and 2.5× its throughput at threshold 0.9.
Loading PDF… open it directly
PDF reproduced unmodified under CC BY-NC-ND 4.0. Highlights and notes are mine.