CLLMs: Consistency Large Language Models
This addresses the problem of slow LLM inference for users needing faster text generation, representing a strong incremental improvement over existing parallel decoding methods.
The paper tackles the inefficiency of parallel decoding methods like Jacobi decoding in LLM inference by developing a new approach that refines LLMs to consistently predict fixed points, achieving 2.4x to 3.4x speed improvements in generation while maintaining quality across benchmarks.
Parallel decoding methods such as Jacobi decoding show promise for more efficient LLM inference as it breaks the sequential nature of the LLM decoding process and transforms it into parallelizable computation. However, in practice, it achieves little speedup compared to traditional autoregressive (AR) decoding, primarily because Jacobi decoding seldom accurately predicts more than one token in a single fixed-point iteration step. To address this, we develop a new approach aimed at realizing fast convergence from any state to the fixed point on a Jacobi trajectory. This is accomplished by refining the target LLM to consistently predict the fixed point given any state as input. Extensive experiments demonstrate the effectiveness of our method, showing 2.4$\times$ to 3.4$\times$ improvements in generation speed while preserving generation quality across both domain-specific and open-domain benchmarks.