Research map / Architectures and efficiency / Pruning and efficient tuning
Deep Compression of Pre-trained Transformer Models
Naigang Wang, Chi-Chun Liu, Swagath Venkataramani, Sanchari Sen, Chia-Yu Chen, Kaoutar El Maghraoui, Viji Srinivasan, Leland Chang
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 changed | Where | Measured | Effect |
|---|---|---|---|
| add zero alignment to weight quantization (ablation) | sparse INT4 models · BERT-base SQuAD1.1 | F1 (%): 86.06 -> 87.07 | Improved |
| add zero alignment to weight quantization (ablation) | sparse INT4 models · Wav2vec2.0 Librispeech | WER (%): 5.76 -> 4.53 | Improved |
| add zero alignment to weight quantization (ablation) | sparse INT4 models · ViT ImageNet1k | accuracy (%): 83.25 -> 83.60 | Improved |
| use PACT instead of MinMax activation quantizer (ablation) | INT8 models · BERT-base SQuAD1.1 | F1 (%): 88.35 -> 87.89 | Hurt |
| use PACT instead of MinMax activation quantizer (ablation) | INT8 models · Wav2vec2.0 Librispeech | WER (%): 3.85 -> 4.05 | Hurt |
| use PACT instead of MinMax activation quantizer (ablation) | INT8 models · ViT ImageNet1k | accuracy (%): 84.47 -> 83.98 | Hurt |
| initialize INT4 from fine-tuned FP32 instead of pre-trained model (ablation) | INT4 models · BERT-base SQuAD1.1 | F1 (%): 86.68 -> 87.86 | Improved |
| initialize INT4 from fine-tuned FP32 instead of pre-trained model (ablation) | INT4 models · Wav2vec2.0 Librispeech | WER (%): 5.02 -> 4.56 | Improved |
| initialize INT8 from fine-tuned FP32 instead of pre-trained model (ablation) | INT8 models · ViT ImageNet1k | accuracy (%): 83.05 -> 84.47 | Improved |
| initialize sparse INT4 from sparse INT8 instead of FP32 fine-tuned model (ablation) | sparse INT4 models · BERT-base SQuAD1.1 | F1 (%): 86.75 -> 87.07 | Improved |
| initialize sparse INT4 from sparse INT8 instead of FP32 fine-tuned model (ablation) | sparse INT4 models · Wav2vec2.0 Librispeech | WER (%): 5.09 -> 4.65 | Improved |
| initialize sparse INT4 from sparse INT8 instead of FP32 fine-tuned model (ablation) | sparse INT4 models · ViT ImageNet1k | accuracy (%): 82.66 -> 83.60 | Improved |
Closest papers by meaning
- Train Big, Then Compress: Rethinking Model Size for Efficient Training and Inference of Transformers (ICML 2020 · similarity 0.91)
- ZeroQuant: Efficient and Affordable Post-Training Quantization for Large-Scale Transformers (NeurIPS 2022 · similarity 0.90)
- Transformers Get Stable: An End-to-End Signal Propagation Theory for Language Models (ICML 2024 · similarity 0.89)
- FrameQuant: Flexible Low-Bit Quantization for Transformers (ICML 2024 · similarity 0.89)
- HybridNorm: Towards Stable and Efficient Transformer Training via Hybrid Normalization (NeurIPS 2025 · similarity 0.89)
- Trainable Transformer in Transformer (ICML 2024 · similarity 0.88)
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.
