Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

FinerCut: Finer-grained Interpretable Layer Pruning for Large Language Models

About

Overparametrized transformer networks are the state-of-the-art architecture for Large Language Models (LLMs). However, such models contain billions of parameters making large compute a necessity, while raising environmental concerns. To address these issues, we propose FinerCut, a new form of fine-grained layer pruning, which in contrast to prior work at the transformer block level, considers all self-attention and feed-forward network (FFN) layers within blocks as individual pruning candidates. FinerCut prunes layers whose removal causes minimal alternation to the model's output -- contributing to a new, lean, interpretable, and task-agnostic pruning method. Tested across 9 benchmarks, our approach retains 90% performance of Llama3-8B with 25% layers removed, and 95% performance of Llama3-70B with 30% layers removed, all without fine-tuning or post-pruning reconstruction. Strikingly, we observe intriguing results with FinerCut: 42% (34 out of 80) of the self-attention layers in Llama3-70B can be removed while preserving 99% of its performance -- without additional fine-tuning after removal. Moreover, FinerCut provides a tool to inspect the types and locations of pruned layers, allowing to observe interesting pruning behaviors. For instance, we observe a preference for pruning self-attention layers, often at deeper consecutive decoder layers. We hope our insights inspire future efficient LLM architecture designs.

Yang Zhang, Yawei Li, Xinpeng Wang, Qianli Shen, Barbara Plank, Bernd Bischl, Mina Rezaei, Kenji Kawaguchi• 2024

Related benchmarks

TaskDatasetResultRank
Commonsense ReasoningWinoGrande
Accuracy69.51
1581
Question AnsweringARC Challenge
Accuracy (ARC)40.61
631
Question AnsweringARC Easy--
597
Multitask Language UnderstandingMMLU
Accuracy33.5
568
Multi-task Language UnderstandingMMLU
MMLU Accuracy38.73
456
Commonsense ReasoningPIQA
Accuracy78.73
400
Multiple-choice Question AnsweringARC Easy
Accuracy68.39
269
Reading ComprehensionBoolQ
Accuracy (BoolQ)71.99
258
Question AnsweringARC Easy
Accuracy65.78
246
Question AnsweringOpenBookQA
Accuracy41.4
157
Showing 10 of 19 rows

Other info

Follow for update