Hungry Hungry Hippos: Towards Language Modeling with State Space Models

TL;DR

This paper advances language modeling by developing a new SSM layer called H3 that improves expressivity, and a novel training algorithm FlashConv that enhances efficiency, resulting in models that outperform Transformers on several benchmarks.

Contribution

Introduces H3, a new SSM layer designed for language tasks, and FlashConv, a fast FFT-based algorithm, to close the gap between SSMs and attention-based models.

Findings

H3 matches attention on synthetic tasks and approaches Transformer perplexity on OpenWebText.

Hybrid H3-attention models outperform Transformers in perplexity on OpenWebText.

FlashConv accelerates training and inference, enabling larger models with better performance.

Abstract

State space models (SSMs) have demonstrated state-of-the-art sequence modeling performance in some modalities, but underperform attention in language modeling. Moreover, despite scaling nearly linearly in sequence length instead of quadratically, SSMs are still slower than Transformers due to poor hardware utilization. In this paper, we make progress on understanding the expressivity gap between SSMs and attention in language modeling, and on reducing the hardware barrier between SSMs and attention. First, we use synthetic language modeling tasks to understand the gap between SSMs and attention. We find that existing SSMs struggle with two capabilities: recalling earlier tokens in the sequence and comparing tokens across the sequence. To understand the impact on language modeling, we propose a new SSM layer, H3, that is explicitly designed for these abilities. H3 matches attention on the synthetic languages and comes within 0.4 PPL of Transformers on OpenWebText. Furthermore, a hybrid 125M-parameter H3-attention model that retains two attention layers surprisingly outperforms Transformers on OpenWebText by 1.0 PPL. Next, to improve the efficiency of training SSMs on modern hardware, we propose FlashConv. FlashConv uses a fused block FFT algorithm to improve efficiency on sequences up to 8K, and introduces a novel state passing algorithm that exploits the recurrent properties of SSMs to scale to longer sequences. FlashConv yields 2$\times$ speedup on the long-range arena benchmark and allows hybrid language models to generate text 2.4$\times$ faster than Transformers. Using FlashConv, we scale hybrid H3-attention language models up to 2.7B parameters on the Pile and find promising initial results, achieving lower perplexity than Transformers and outperforming Transformers in zero- and few-shot learning on a majority of tasks in the SuperGLUE benchmark.