Hardware-Aware Parallel Prompt Decoding for Memory-Efficient Acceleration of LLM Inference

TL;DR

This paper introduces a hardware-aware parallel prompt decoding method for LLM inference that significantly reduces memory and training costs while improving speed and prediction accuracy, adaptable to various GPU architectures.

Contribution

It proposes a novel parallel decoding technique inspired by human language generation, with a dynamic sparse tree optimization, enabling efficient training and inference on standard GPUs.

Findings

Achieves up to 2.49× speedup in inference.

Requires only 0.0002% trainable parameters, enabling quick training.

Maintains minimal memory overhead of 0.0004%."],

Abstract

The auto-regressive decoding of Large Language Models (LLMs) results in significant overheads in their hardware performance. While recent research has investigated various speculative decoding techniques for multi-token generation, these efforts have primarily focused on improving processing speed such as throughput. Crucially, they often neglect other metrics essential for real-life deployments, such as memory consumption and training cost. To overcome these limitations, we propose a novel parallel prompt decoding that requires only $0.0002$% trainable parameters, enabling efficient training on a single A100-40GB GPU in just 16 hours. Inspired by the human natural language generation process, $PPD$ approximates outputs generated at future timesteps in parallel by using multiple prompt tokens. This approach partially recovers the missing conditional dependency information necessary for multi-token generation, resulting in up to a 28% higher acceptance rate for long-range predictions. Furthermore, we present a hardware-aware dynamic sparse tree technique that adaptively optimizes this decoding scheme to fully leverage the computational capacities on different GPUs. Through extensive experiments across LLMs ranging from MobileLlama to Vicuna-13B on a wide range of benchmarks, our approach demonstrates up to 2.49$\times$ speedup and maintains a minimal runtime memory overhead of just $0.0004$%. More importantly, our parallel prompt decoding can serve as an orthogonal optimization for synergistic integration with existing speculative decoding, showing up to $1.22\times$ further speed improvement. Our code is available at https://github.com/hmarkc/parallel-prompt-decoding.