Ghostdecoding: leveraging random-feature kernels for error-aware and training-free KV cache selection

Key-value (KV) cache is essential for efficient inference in large language models (LLMs) by storing intermediate representations to reduce redundant computations. However, as sequence lengths grow, it becomes a major bottleneck due to increasing computational and memory demands. Existing KV cache compression methods mitigate this issue by pruning or selecting critical entries, but often suffer from several limitations, including unpredictable errors, limited dynamism, and strong assumptions about context relevance. In this work, we propose GhostDecoding, a novel training-free KV cache selection mechanism that dynamically reduces the effective sequence length while maintaining error awareness for evicted entries. Using a random feature softmax kernel, our method estimates the attention score of an arbitrary number of evicted positions with O(1) time and space overhead, and selectively recomputes them when necessary. We develop efficient sparse CUDA kernels to support our algorithm. Experimental results demonstrate that GhostDecoding achieves up to 1.6× more computation reduction compared to H2O, leading to up to 1.9× decoding speed-up compared to full attention on long sequences.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Illinois Urbana-Champaign

Year dc:date

2025

Author dc:creator

• Guo, Hao

Contributors dc:contributor

• Mendis, Charith

• Large Language Models
• Efficient ML
• KV Cache

Statement dc:rights

• Copyright 2025 Hao Guo

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/129705

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F129705