AI Paper Atlas

Research map / Architectures and efficiency / Quantization

Memory Efficient Optimizers with 4-bit States

Bingrui Li, Jianfei Chen, Jun Zhu

NeurIPS 2023 · Architectures and efficiency › Quantization · approach: Low-precision training formats · 3 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
Replace 32-bit AdamW states with the proposed 4-bit quantized statesAdamW · ImageNet-1k with Swin-TAccuracy: 81.2 ± 0.05 -> 80.8 ± 0.02Hurt
Replace 32-bit AdamW states with the proposed 4-bit quantized statesAdamW · E2E-NLG with GPT-2 MediumBLEU: 67.7 ± 0.67 -> 67.8 ± 0.51No clear effect
Replace 32-bit AdamW states with the proposed 4-bit quantized statesAdamW · WMT14 en-de with Transformer-BaseSacreBleu: 26.61 ± 0.08 -> 26.28 ± 0.05Hurt
Use B128/DE rather than B2048/DE for the first moment (ablation)4-bit optimizer · ImageNet-1k with Swin-TAccuracy: 80.9 ± 0.04 -> 81.0 ± 0.06Improved
Quantize the second moment with Rank-1/Linear (ablation)4-bit optimizer · ImageNet-1k with Swin-TAccuracy: 81.0 ± 0.06 -> 80.8 ± 0.02Hurt
Factorize the second moment (ablation)4-bit optimizer with Rank-1/Linear second moment · ImageNet-1k with Swin-TAccuracy: 80.8 ± 0.02 -> 80.9 ± 0.06Improved
Exclude zero from the second-moment dynamic-exponent mapping (ablation)4-bit optimizer · E2E-NLG with GPT-2 MediumBLEU: 66.6 ± 0.61 -> 67.5 ± 0.97Improved
Use B128/DE rather than B2048/DE for the second moment (ablation)4-bit optimizer · E2E-NLG with GPT-2 MediumBLEU: 66.6 ± 0.61 -> 65.7 ± N/AHurt
Use rank-1 normalization with linear mapping for the second moment (ablation)4-bit optimizer · E2E-NLG with GPT-2 MediumBLEU: 67.5 ± 0.58 -> 67.8 ± 0.51Improved
Use the 4-bit optimizer instead of 32-bit AdamWAdamW · LLaMA-33B instruction tuningMMLU (5-shot): 56.4 -> 54.9Hurt
Use the 4-bit optimizer instead of 32-bit AdamWAdamW · LLaMA-7B trainingTotal memory: 75.40 GB -> 31.87 GBImproved
Use the 4-bit optimizer instead of 8-bit AdamWAdamW · RoBERTa-L RTE trainingTime: 3.38 min -> 5.59 minHurt

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