Research

Paper

AI LLM March 24, 2026

Sparser, Faster, Lighter Transformer Language Models

Authors

Edoardo Cetin, Stefano Peluchetti, Emilio Castillo, Akira Naruse, Mana Murakami, Llion Jones

Abstract

Scaling autoregressive large language models (LLMs) has driven unprecedented progress but comes with vast computational costs. In this work, we tackle these costs by leveraging unstructured sparsity within an LLM's feedforward layers, the components accounting for most of the model parameters and execution FLOPs. To achieve this, we introduce a new sparse packing format and a set of CUDA kernels designed to seamlessly integrate with the optimized execution pipelines of modern GPUs, enabling efficient sparse computation during LLM inference and training. To substantiate our gains, we provide a quantitative study of LLM sparsity, demonstrating that simple L1 regularization can induce over 99% sparsity with negligible impact on downstream performance. When paired with our kernels, we show that these sparsity levels translate into substantial throughput, energy efficiency, and memory usage benefits that increase with model scale. We will release all code and kernels under an open-source license to promote adoption and accelerate research toward establishing sparsity as a practical axis for improving the efficiency and scalability of modern foundation models.

Metadata

arXiv ID: 2603.23198
Provider: ARXIV
Primary Category: cs.LG
Published: 2026-03-24
Fetched: 2026-03-25 06:02

Related papers

Raw Data (Debug)
{
  "raw_xml": "<entry>\n    <id>http://arxiv.org/abs/2603.23198v1</id>\n    <title>Sparser, Faster, Lighter Transformer Language Models</title>\n    <updated>2026-03-24T13:43:27Z</updated>\n    <link href='https://arxiv.org/abs/2603.23198v1' rel='alternate' type='text/html'/>\n    <link href='https://arxiv.org/pdf/2603.23198v1' rel='related' title='pdf' type='application/pdf'/>\n    <summary>Scaling autoregressive large language models (LLMs) has driven unprecedented progress but comes with vast computational costs. In this work, we tackle these costs by leveraging unstructured sparsity within an LLM's feedforward layers, the components accounting for most of the model parameters and execution FLOPs. To achieve this, we introduce a new sparse packing format and a set of CUDA kernels designed to seamlessly integrate with the optimized execution pipelines of modern GPUs, enabling efficient sparse computation during LLM inference and training. To substantiate our gains, we provide a quantitative study of LLM sparsity, demonstrating that simple L1 regularization can induce over 99% sparsity with negligible impact on downstream performance. When paired with our kernels, we show that these sparsity levels translate into substantial throughput, energy efficiency, and memory usage benefits that increase with model scale. We will release all code and kernels under an open-source license to promote adoption and accelerate research toward establishing sparsity as a practical axis for improving the efficiency and scalability of modern foundation models.</summary>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.LG'/>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.CL'/>\n    <published>2026-03-24T13:43:27Z</published>\n    <arxiv:comment>Code and checkpoints available at: https://github.com/SakanaAI/sparser-faster-llms</arxiv:comment>\n    <arxiv:primary_category term='cs.LG'/>\n    <author>\n      <name>Edoardo Cetin</name>\n    </author>\n    <author>\n      <name>Stefano Peluchetti</name>\n    </author>\n    <author>\n      <name>Emilio Castillo</name>\n    </author>\n    <author>\n      <name>Akira Naruse</name>\n    </author>\n    <author>\n      <name>Mana Murakami</name>\n    </author>\n    <author>\n      <name>Llion Jones</name>\n    </author>\n  </entry>"
}