Efficient Speculative Decoding for Llama at Scale: Challenges and Solutions

TL;DR

This paper presents optimized speculative decoding techniques for Llama models, achieving state-of-the-art inference latency and significant speed-ups on GPU hardware, addressing engineering challenges for production deployment.

Contribution

The paper introduces training and inference optimizations enabling EAGLE-based speculative decoding at scale for Llama models, improving speed and efficiency.

Findings

Llama4 Maverick decodes at about 4 ms per token on 8 GPUs

Achieved 10% faster inference than previous methods

Enabled 1.4x to 2.0x speed-up for large batch sizes

Abstract

Speculative decoding is a standard method for accelerating the inference speed of large language models. However, scaling it for production environments poses several engineering challenges, including efficiently implementing different operations (e.g., tree attention and multi-round speculative decoding) on GPU. In this paper, we detail the training and inference optimization techniques that we have implemented to enable EAGLE-based speculative decoding at a production scale for Llama models. With these changes, we achieve a new state-of-the-art inference latency for Llama models. For example, Llama4 Maverick decodes at a speed of about 4 ms per token (with a batch size of one) on 8 NVIDIA H100 GPUs, which is 10% faster than the previously best known method. Furthermore, for EAGLE-based speculative decoding, our optimizations enable us to achieve a speed-up for large batch sizes between 1.4x and 2.0x at production scale.