The case for 4-bit precision: k-bit Inference Scaling Laws
Authors:
- Tim DettmersTDTim Dettmers
- Luke ZettlemoyerLZLuke Zettlemoyer
Preprint
Abstract
Quantization methods reduce the number of bits required to represent each parameter in a model, trading accuracy for smaller memory footprints and inference latencies. However, the final model size depends on both the number of parameters of the original model and the rate of compression. For example, a 30B 8-bit model and a 60B 4-bit model have the same number of bits but may have very different zero-shot accuracies. In this work, we study this trade-off by developing inference scaling laws of zero-shot performance in Large Language Models (LLMs) to determine the bit-precision and model size that maximizes zero-shot performance. We run more than 35,000 experiments with 16-bit inputs and k-bit parameters to examine which zero-shot quantization methods improve scaling for 3 to 8-bit precision at scales of 19M to 176B parameters across the LLM families BLOOM, OPT, NeoX/Pythia, and GPT-2. We find that it is challenging to improve the bit-level scaling trade-off, with the only improvements being the use of a small block size -- splitting the parameters into small independently quantized blocks -- and the quantization data type being used (e.g., Int vs Float). Overall, our findings show that {4-bit} precision is almost universally optimal for total model bits and zero-shot accuracy.Keywords
All Related Versions
This publication has 12 references indexed in Scilit:
- Outlier Dimensions that Disrupt Transformers are Driven by Frequency2022
- GPT-NeoX-20B: An Open-Source Autoregressive Language Model2022
- WinoGrandeCommunications of the ACM, 2021
- All Bark and No Bite: Rogue Dimensions in Transformer Language Models Obscure Representational Quality2021
- Understanding and Overcoming the Challenges of Efficient Transformer Quantization2021
- Positional Artefacts Propagate Through Masked Language Model Embeddings2021
- BERT Busters: Outlier Dimensions that Disrupt Transformers2021
- Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference2018
- XNOR-Net: ImageNet Classification Using Binary Convolutional Neural Networks2016
- The LAMBADA dataset: Word prediction requiring a broad discourse context2016