From Cloud to Edge: A Comprehensive Review of Optimization Strategies for On-Device Large Language Models

Resumen

Making artificial intelligence move to the edge is a promise of more privacy, lower latency, and offline, but it has to confront the realities of consumer hardware resources. The given paper represents a thorough taxonomy of the on-device LLM optimization technique, which summarizes the discoveries of 2023-2025. A granular review of optimization methods allows us to note a severe bifurcation in the research field: an excessive use of the Llama model family to benchmark against an increasing fragmentation of hardware targets, which is ranging between Raspberry pi to specialized Jetson modules. Our research underscores the fact that optimization that fits all should be left in the past and that to be successful, deployment should be constraint-based. The literature is proposed by us to compare the methods in terms of compression, accuracy, speedup, and sparsity. The results show that the most uniform energy advantages are obtained with quantization, although hybrid hardware-aware methods are the only way to break existing latency bottlenecks on the heterogeneous systems.