Multi-Objective Structured Pruning for Edge LLM Deployment

A two-stage hardware-aware framework jointly optimizes layers, attention heads, and MLP dimensions. This method targets specific latency and model size constraints to improve LLM execution on resource-constrained edge devices.