APAR: LLMs Can Do Auto-Parallel Auto-Regressive Decoding
This addresses the problem of slow text generation for LLM users, offering a significant but incremental improvement in serving efficiency.
The paper tackles the inefficiency of auto-regressive decoding in large language models (LLMs) by introducing APAR, a method that enables parallel generation, achieving up to 2x speed-up alone and 4x with speculative decoding, along with throughput increases of 20-70% and latency reductions of 20-35%.
The massive adoption of large language models (LLMs) demands efficient deployment strategies. However, the auto-regressive decoding process, which is fundamental to how most LLMs generate text, poses challenges to achieve efficient serving. In this work, we introduce a parallel auto-regressive generation method. By instruct-tuning on general domain data that contains hierarchical structures, we enable LLMs to independently plan their generation process and perform auto-parallel auto-regressive (APAR) generation, significantly reducing the number of generation steps. APAR alone can achieve up to 2x speed-up, and when combined with speculative decoding, the speed-up can reach up to 4x. In addition, APAR reduces the key-value cache consumption and attention computation during generation. This leads to a throughput increase of 20-70% and a latency reduce of 20-35% in high-throughput scenarios, compared to state-of-the-art serving frameworks.