EdgeLLM-Nano evaluates key-cache compression on small on-device language models
EdgeLLM-Nano: Auditing Polar-Guided Key-Cache Compression and Consistency-Anchored Pruning on a 33.6 MB On-Device Language Model
Computation and Language
Summary
Running language models on small devices is hard because they have limited memory and power. The authors tested ways to shrink a part of the model’s memory called the KV-cache to see if it helps. Surprisingly, some popular compression methods didn’t work well and could harm performance. They introduced a new way to measure how well these methods work, showing some flaws that usual tests missed. Their tests on a Raspberry Pi also found using compression can cost more energy than just running the model normally.
What this means in practice
- •For mobile app developers: Improve memory use and speed by testing new key-cache compression methods on small on-device language models before deploying.
- •For embedded systems engineers: Benchmark KV-cache compression impacts on efficiency and accuracy for deploying transformer models on resource-limited hardware like Raspberry Pi.
Authors
S M Mahbub Murshed, Hayden Bright, Rezaul Karim, Ali Shiri Sichani
Abstract
Deploying language models (LMs) on the edge presents unique memory, latency, and power challenges absent in high-performance computing. KV-cache compression emerges as a promising approach to address these challenges. However, the effectiveness of current KV-cache compression methods has not been thoroughly examined on edge devices. Moreover, we show that existing evaluation methods do not always involve the KV-cache under test. We propose EdgeLLM-Nano, a measurement study of KV-cache compression using a nano-scale decoder-only LM (transformer) testbed that has 8.39M parameters, i.e., 33.6 MB, to enable end-to-end examination of each backend, seed, and power sample. Our results give a negative verdict for a polar-guided key-cache quantizer (round-to-nearest (RTN) control causes an order of magnitude less damage), as well as a consistency-anchored progressive weight pruner (one-shot magnitude-based pruning performs as well as, if not better). We also introduce Long-Context Decode Fidelity (LCDF) evaluation, a token-by-token examination of teacher-forced log-losses using the KV-cache under test. Unlike perplexity that gives the same 5.461 score to all methods, LCDF distinguishes different KV-cache compression backends thanks to its paired statistics across seeds and effect sizes. Our companion Raspberry Pi 5 measurement shows all compression backends we tested to be less energy-efficient per token compared to uncompressed FP16 inference. Except for KIVI that is close to lossless in terms of LCDF, our findings indicate KV-cache compression can be improved. LCDF offers a simple and effective way of shedding light on KV-cache compression methods, whose negative impacts have been overlooked by perplexity.