GRASPrune: Global Gating for Budgeted Structured Pruning of Large Language Models
For practitioners deploying LLMs, GRASPrune offers a budget-aware pruning framework that reduces memory and latency costs without requiring full model fine-tuning, though the gains are incremental over existing pruning methods.
GRASPrune introduces a structured pruning method for LLMs that jointly prunes FFN channels and KV head groups under a single global budget, achieving 50% parameter removal on LLaMA-2-7B with 12.18 perplexity on WikiText-2 and competitive zero-shot accuracy, using only 512 calibration sequences and no full fine-tuning.
Large language models (LLMs) are expensive to serve because model parameters, attention computation, and KV caches impose substantial memory and latency costs. We present GRASPrune, a structured pruning framework applied after pretraining that jointly prunes FFN channels and KV head groups under a single global budget. Instead of learning importance scores without constraints and applying the budget only after training, GRASPrune learns lightweight gate scores with a projected straight-through estimator that enforces a hard mask satisfying the budget at every step while keeping the backbone weights frozen. After the mask is fixed, we calibrate scaling factors on the retained units to mitigate scale mismatch caused by pruning, and fold these factors into the pruned weights to obtain a smaller dense checkpoint with no extra parameters at inference. On LLaMA-2-7B, GRASPrune removes 50% of parameters and achieves 12.18 perplexity on WikiText-2 while maintaining competitive average zero-shot accuracy on five benchmarks, using four epochs on 512 unlabeled calibration sequences on a single NVIDIA A100 80GB GPU without any full model fine-tuning.