AI Paper Atlas

Research map / Architectures and efficiency / Pruning and efficient tuning

ZipLM: Inference-Aware Structured Pruning of Language Models

Eldar Kurtic, Elias Frantar, Dan Alistarh

NeurIPS 2023 · Architectures and efficiency › Pruning and efficient tuning · approach: Layer-wise pruning of LLMs · 7 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
prune with ZipLM instead of DistilGPT2 at similar speedup (throughput)GPT2 · WikiText-103perplexity: 43.0 -> 35.4Improved
prune for latency with ZipLM instead of DistilGPT2GPT2 · WikiText-103perplexity: 43.0 -> 41.2Improved
ZipLM one-shot vs Kwon et al., 2.0x speedupBERT-base · SQuADF1: 76.5 -> 84.1Improved
ZipLM one-shot vs Kwon et al., 2.0x speedupBERT-base · QQPaccuracy: 83.9 -> 84.8Improved
ZipLM one-shot vs Kwon et al., 2.0x speedupBERT-base · MNLIaccuracy: 78.1 -> 78.2No clear effect
prune for speedup instead of for sparsity (ablation)ZipLM · BERT-baseaccuracy: up to 10 pointsImproved
increase calibration samples from 4 to 32 (ablation)ZipLM one-shot · SQuAD1.1, 2.0xF1: 48.4 -> 82.6Improved
increase calibration samples from 32 to 4096 (ablation)ZipLM one-shot · SQuAD1.1, 1.5xF1: 86.8 -> 87.6Improved
add layer-wise token distillation to ZipLM (ablation)ZipLM · SST-2accuracy: 89.2 -> 91.7Improved
add layer-wise token distillation to ZipLM (ablation)ZipLM · QNLIaccuracy: 86.5 -> 88.6Improved
add layer-wise token distillation to ZipLM (ablation)ZipLM · SQuADF1: 85.7 -> 85.7No clear effect
use ZipLM instead of CoFi with distillationBERT-base · SQuADF1: 82.6 -> 85.7Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy