The official implementation for the paper: "A deeper look at depth pruning of LLMs" (https://arxiv.org/abs/2407.16286).
The main scripts are divided into three different categories.
The main experiments are based on block influence where we compute the impact of different block influence estimation techniques for block pruning.
The associated scripts are located in the directory: ./block_influence/.
The evaluate the model based on different block influence techniques, check the script: block_influence/eval.sh.
MMLU shapley focuses on computing the loss shapley directly on the MMLU test set, which serves as an upperbound on the performance that can be achieved by different block pruning techniques on the MMLU benchmark.
The experiments associated with MMLU shapley based block pruning are located in the directory: ./block_influence_mmlu_shapley/.
The evaluate the model based on MMLU loss shapley based block pruning, check the script: block_influence_mmlu_shapley/eval.sh.
We further disect the transformer block into its corresponding feed-forward and self-attention layers, and evaluate their impact separately.
The experiments associated with layer influence estimation are located in the directory: ./layer_influence/.
The evaluate the model based on different layer influence techniques, check the script: layer_influence/eval.sh.
@inproceedings{siddiqui2024deeper,
title={A deeper look at depth pruning of LLMs},
author={Siddiqui, Shoaib Ahmed and Dong, Xin and Heinrich, Greg and Breuel, Thomas and Kautz, Jan and Krueger, David and Molchanov, Pavlo},
booktitle={ICML 2024 Workshop on Theoretical Foundations of Foundation Models}
}
MIT