Timezone: »

 
Accelerating LLM Inference with Staged Speculative Decoding
Benjamin F Spector · Christopher Re

@
in ES-FoMo: Efficient Systems for Foundation Models »
Event URL: https://openreview.net/forum?id=RKHF3VYjLK »

Recent advances with large language models (LLM) illustrate their diverse capabilities. We propose a novel algorithm, staged speculative decoding, to accelerate LLM inference in small-batch, on-device scenarios. We address the low arithmetic intensity of small-batch inference by improving upon previous work in speculative decoding. First, we restructure the speculative batch as a tree, which reduces generation costs and increases the expected tokens per batch. Second, we add a second stage of speculative decoding. Taken together, we reduce single-batch decoding latency by 3.16x with a 762M parameter GPT-2-L model while perfectly preserving output quality.

Author Information

Benjamin F Spector (Stanford University)
Christopher Re (Stanford University)

More from the Same Authors