Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CARVQ: Corrective Adaptor with Group Residual Vector Quantization for LLM Embedding Compression

Dayin Gou, Sanghyun Byun, Nilesh Malpeddi, Gabrielle De Micheli, Prathamesh Vaste, Jacob Song, Woo Seong Chung

TL;DR

A novel compression technique that is compatible with state-of-the-art transformer quantization methods and can be seamlessly integrated into any hardware supporting 4-bit memory to reduce the model's memory footprint in memory-constrained devices is introduced.

Abstract

Large Language Models (LLMs) typically rely on a large number of parameters for token embedding, leading to substantial storage requirements and memory footprints. In particular, LLMs deployed on edge devices are memory-bound, and reducing the memory footprint by compressing the embedding layer not only frees up the memory bandwidth but also speeds up inference. To address this, we introduce CARVQ, a post-training novel Corrective Adaptor combined with group Residual Vector Quantization. CARVQ relies on the composition of both linear and non-linear maps and mimics the original model embedding to compress to approximately 1.6 bits without requiring specialized hardware to support lower-bit storage. We test our method on pre-trained LLMs such as LLaMA-3.2-1B, LLaMA-3.2-3B, LLaMA-3.2-3B-Instruct, LLaMA-3.1-8B, Qwen2.5-7B, Qwen2.5-Math-7B and Phi-4, evaluating on common generative, discriminative, math and reasoning tasks. We show that in most cases, CARVQ can achieve lower average bitwidth-per-parameter while maintaining reasonable perplexity and accuracy compared to scalar quantization. Our contributions include a novel compression technique that is compatible with state-of-the-art transformer quantization methods and can be seamlessly integrated into any hardware supporting 4-bit memory to reduce the model's memory footprint in memory-constrained devices. This work demonstrates a crucial step toward the efficient deployment of LLMs on edge devices.

CARVQ: Corrective Adaptor with Group Residual Vector Quantization for LLM Embedding Compression

TL;DR

A novel compression technique that is compatible with state-of-the-art transformer quantization methods and can be seamlessly integrated into any hardware supporting 4-bit memory to reduce the model's memory footprint in memory-constrained devices is introduced.

Abstract

Large Language Models (LLMs) typically rely on a large number of parameters for token embedding, leading to substantial storage requirements and memory footprints. In particular, LLMs deployed on edge devices are memory-bound, and reducing the memory footprint by compressing the embedding layer not only frees up the memory bandwidth but also speeds up inference. To address this, we introduce CARVQ, a post-training novel Corrective Adaptor combined with group Residual Vector Quantization. CARVQ relies on the composition of both linear and non-linear maps and mimics the original model embedding to compress to approximately 1.6 bits without requiring specialized hardware to support lower-bit storage. We test our method on pre-trained LLMs such as LLaMA-3.2-1B, LLaMA-3.2-3B, LLaMA-3.2-3B-Instruct, LLaMA-3.1-8B, Qwen2.5-7B, Qwen2.5-Math-7B and Phi-4, evaluating on common generative, discriminative, math and reasoning tasks. We show that in most cases, CARVQ can achieve lower average bitwidth-per-parameter while maintaining reasonable perplexity and accuracy compared to scalar quantization. Our contributions include a novel compression technique that is compatible with state-of-the-art transformer quantization methods and can be seamlessly integrated into any hardware supporting 4-bit memory to reduce the model's memory footprint in memory-constrained devices. This work demonstrates a crucial step toward the efficient deployment of LLMs on edge devices.
Paper Structure (28 sections, 5 equations, 3 figures, 6 tables)

This paper contains 28 sections, 5 equations, 3 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Background
  4. Input embedding matrix $\mathcal{M}$
  5. Group Residual Vector Quantization
  6. Group RVQ.
  7. Compression with group RVQ.
  8. CARVQ Embedding Compression
  9. Corrective Adaptor
  10. Defining the map $\sigma_0$.
  11. Defining the map $\sigma_1$.
  12. Compression ratio.
  13. Combining Corrective Adaptor with group RVQ
  14. Experimental results
  15. Implementation Details
  16. ...and 13 more sections

Figures (3)

  • Figure 1: Overview of the CARVQ framework. (a) The embedding matrix $\mathcal{M}$ is partitioned into groups, and each group is compressed using Residual Vector Quantization, with the results stored in a look-up table. LUT is stored in $\kappa$-bit and VQ elements are stored in original precision $p$. (b) At inference time, each token retrieves its corresponding quantized vectors from the look-up table. Simultaneously, it is processed by the corrective adaptor, a lightweight MLP.
  • Figure 2: With transformer layer quantization, the memory ratio of embedding layer increases compared to embedding layers left in FP16.
  • Figure 3: Wikitext-2 Perplexity for LLaMA-3.2-3B and LlaMA-3.1-8B with scalar quantization and CARVQ.

Theorems & Definitions (2)

  • Definition 3.1
  • Definition 3.2