KV-Dict: Sparse KV Cache Compression with Universal Dictionaries

Junhyuck Kim, Jongho Park, Jaewoong Cho, Dimitris Papailiopoulos

Keywords: transformer, kv cache, compression, quantization

TL;DR: We find a universal dictionary for LLM KV cache compression.

Abstract:

Transformer has become the de facto architecture for Large Language Models (LLMs), yet its substantial memory required for long contexts make it costly to deploy. Managing the memory usage of the key-value (KV) cache during inference has become a pressing challenge, as the cache grows with both model size and input length, consuming significant GPU memory.

We introduce a novel post-training KV cache compression method using KV-Dict, a universal dictionary that can accurately decompose and reconstruct key-value states. Unlike traditional quantization methods, KV-dict leverages sparse dictionary learning, allowing for flexible memory usage with minimal performance loss through fine-grained controls of sparsity levels. Moreoever, we retain competitive performance in the low memory regimes that 2-bit compression struggles to offer.

KV-Dict is remarkably universal, as it uses a small, input-agnostic dictionary that is shared across tasks and batches without scaling memory. This universality, combined with the ability to control sparsity for different memory requirements, offers a flexible and efficient solution to the KV cache bottleneck, maintaining strong performance on complex reasoning tasks, such as LongBench and GSM8k.

Primary Area: foundation or frontier models, including LLMs

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 2039