TechPrufer
Sign inGet started

Blogs›Together AI›Consistency Diffusion Language Models

Consistency Diffusion Language Models

Consistency Diffusion Language Models

1
posts
2026

Together AI introduces Consistency Diffusion Language Models (CDLM), a post-training recipe that significantly accelerates inference for Diffusion Language Models (DLMs). CDLM combines consistency-based multi-token finalization with block-wise KV caching, achieving up to 14.5x latency speedups on math and coding tasks. This method addresses the inefficiencies of standard DLMs by enabling exact block-wise KV caching and reducing the high number of refinement steps required for quality generation. CDLM trains a block-wise causal student model by minimizing distillation loss for newly unmasked positions, consistency loss for still-masked positions, and an auxiliary DLM masked-denoising loss. At inference, CDLM decodes in a block-wise autoregressive manner with KV caching for finalized blocks and parallel finalization within each block. System-level analysis shows that block-wise DLMs occupy a balanced operating point for small-batch settings, offering improved arithmetic intensity over autoregressive models while remaining efficient.

2026

Consistency diffusion language models: Up to 14x faster inference without sacrificing quality

2/19/2026

This post introduces Consistency Diffusion Language Models (CDLM), a novel training-based acceleration scheme for Diffusion Language Models (DLMs). CDLM enables faster inference by combining consistency modeling with block-wise KV caching. Key technical contributions include: 1) A post-training recipe that makes fewer-step inference reliable while enabling exact block-wise KV caching. 2) Training objectives: distillation loss for newly unmasked positions, consistency loss for still-masked positions, and an auxiliary DLM masked-denoising loss. 3) A block-wise causal student architecture that attends to the prompt, previously completed blocks, and the current decoding block. 4) Inference strategy using block-wise autoregressive decoding with KV cache reuse and confidence-thresholded parallel finalization within blocks. The post also provides system-level analysis of arithmetic intensity and roofline placement for CDLM compared to AR and vanilla DLMs, explaining its efficiency at small batch sizes.