Scholarlink

Using large language models (LLMs) to classify genomic and transcriptomic data greatly influences drug discovery. LLMs can handle large datasets and complex biological sequences. However, they are computationally expensive. It is assumed that lowering LLM weight precision would hurt accuracy in tasks like gene classification. However, this work shows a case where the quantized LLM performs better. It also offers a faster, smarter, and more efficient model. We compared a 4-bit model with a 32-bit model for gene classification. The small LLMs can outperform nonquantized larger LLMs. Our goal is to show a model that is faster and achieves high accuracy with limited resources. Quantization allows models to be used across various platforms and on edge devices without GPU limitations. The small LLMs can be adaptable to different memory and GPU allocations.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

LLM Quantization: Enabling Edge Intelligence

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Maryam Heidari,
  • Samira Zad,
  • Kianoosh Boroojeni,
  • Marcia Katzman

摘要

Using large language models (LLMs) to classify genomic and transcriptomic data greatly influences drug discovery. LLMs can handle large datasets and complex biological sequences. However, they are computationally expensive. It is assumed that lowering LLM weight precision would hurt accuracy in tasks like gene classification. However, this work shows a case where the quantized LLM performs better. It also offers a faster, smarter, and more efficient model. We compared a 4-bit model with a 32-bit model for gene classification. The small LLMs can outperform nonquantized larger LLMs. Our goal is to show a model that is faster and achieves high accuracy with limited resources. Quantization allows models to be used across various platforms and on edge devices without GPU limitations. The small LLMs can be adaptable to different memory and GPU allocations.