Turbo Sparse: Achieving LLM SOTA Performance with Minimal Activated Parameters

TL;DR

Turbo Sparse introduces a novel activation function and training strategy to significantly increase activation sparsity in large language models, enabling faster inference with minimal performance loss and practical speedups on mobile devices.

Contribution

The paper proposes a new dReLU activation function and training data mixture ratio, along with leveraging sparse patterns in MoE models, to enhance activation sparsity and inference efficiency in LLMs.

Findings

Achieves 2-5x decoding speedup in large models.

Only 2.5B and 4.3B parameters activated per inference.

Mobile inference speed reaches 11 tokens/sec.

Abstract

Exploiting activation sparsity is a promising approach to significantly accelerating the inference process of large language models (LLMs) without compromising performance. However, activation sparsity is determined by activation functions, and commonly used ones like SwiGLU and GeGLU exhibit limited sparsity. Simply replacing these functions with ReLU fails to achieve sufficient sparsity. Moreover, inadequate training data can further increase the risk of performance degradation. To address these challenges, we propose a novel dReLU function, which is designed to improve LLM activation sparsity, along with a high-quality training data mixture ratio to facilitate effective sparsification. Additionally, we leverage sparse activation patterns within the Feed-Forward Network (FFN) experts of Mixture-of-Experts (MoE) models to further boost efficiency. By applying our neuron sparsification method to the Mistral and Mixtral models, only 2.5 billion and 4.3 billion parameters are activated per inference iteration, respectively, while achieving even more powerful model performance. Evaluation results demonstrate that this sparsity achieves a 2-5x decoding speedup. Remarkably, on mobile phones, our TurboSparse-Mixtral-47B achieves an inference speed of 11 tokens per second. Our models are available at \url{https://huggingface.co/PowerInfer}