§ feed · storyline
Consistency diffusion models achieve 14x faster inference without
Consistency diffusion language models (CDLM) achieve up to 14.5x faster inference by enabling block-wise KV caching and reducing refinement steps via a post-training method.
Standard diffusion language models can't use KV caching and need too many refinement steps to be practical. CDLM fixes both with a post-training recipe that enables exact block-wise KV caching and trajectory-consistent step reduction — delivering up to 14.5x latency improvements
§ sources1 publication · timeline below