Jove
Visualize
联系我们
JoVE
x logofacebook logolinkedin logoyoutube logo
关于 JoVE
概览领导团队博客JoVE 帮助中心
作者
出版流程编辑委员会范围与政策同行评审常见问题投稿
图书馆员
用户评价订阅访问资源图书馆顾问委员会常见问题
研究
JoVE JournalMethods CollectionsJoVE Encyclopedia of Experiments存档
教育
JoVE CoreJoVE BusinessJoVE Science EducationJoVE Lab Manual教师资源中心教师网站
使用条款与条件
隐私政策
政策

相关概念视频

Improving Translational Accuracy02:07

Improving Translational Accuracy

2.7K
2.7K
Per-Unit Sequence Models01:26

Per-Unit Sequence Models

119
An ideal Y-Y transformer, grounded through neutral impedances, displays per-unit sequence networks akin to those of a single-phase ideal transformer when subjected to balanced positive- or negative-sequence currents. These currents do not produce neutral currents, and their associated voltage drops.
Zero-sequence currents, which are identical in magnitude and phase, generate a neutral current, resulting in voltage drops across the neutral impedance and the low-voltage winding. If the...
119
Choosing Between z and t Distribution01:25

Choosing Between z and t Distribution

2.9K
The z and the Student t distribution estimate the population mean using the sample mean and standard deviation. However, to decide which distribution to use for a calculation, one needs to determine the sample size, the nature of the distribution, and whether the population standard deviation is known. If the population standard deviation is known and the population is normally distributed, or if the sample size is greater than 30, the z distribution is preferred. The Student t distribution is...
2.9K
Linear Approximation in Frequency Domain01:26

Linear Approximation in Frequency Domain

136
Linear systems are characterized by two main properties: superposition and homogeneity. Superposition allows the response to multiple inputs to be the sum of the responses to each individual input. Homogeneity ensures that scaling an input by a scalar results in the response being scaled by the same scalar.
In contrast, nonlinear systems do not inherently possess these properties. However, for small deviations around an operating point, a nonlinear system can often be approximated as linear....
136
Upsampling01:22

Upsampling

319
Managing signal sampling rates is essential in digital signal processing to maintain signal integrity. A decimated signal, characterized by a reduced frequency range due to its lower sampling rate, can be upsampled by inserting zeros between each sample. This upsampling process expands the original spectrum and introduces repeated spectral replicas at intervals dictated by the new Nyquist frequency. To refine this zero-inserted sequence, it is passed through a lowpass filter with a cutoff...
319
Sample Size Calculation01:19

Sample Size Calculation

3.8K
Knowledge of the sample size is the first requirement to conduct random sampling or an experiment. The sample size is the total number of units, observations, or groups (in some cases) used to get the data to estimate a population parameter. As the name suggests, the sample size is that of the sample drawn from the population and differs from the population size.
The sample size for the given experiment or sampling effort is fundamental to any study design. Sample size decides the number of...
3.8K

您也可能阅读

相关文章

通过共同作者、期刊和引用图与本文相关的文章。

排序
Same author

An injector testbed based on a direct current gun and an interchangeable very high frequency gun for superconducting continuous-wave free-electron lasers.

Journal of synchrotron radiation·2025
Same author

FEbeam: Cavity and electron emission data conversion, processing, and analysis. A freeware toolkit for rf injectors.

The Review of scientific instruments·2021
Same author

Transition Radiation in Photonic Topological Crystals: Quasiresonant Excitation of Robust Edge States by a Moving Charge.

Physical review letters·2019
Same author

Generation of High-Power, Reversed-Cherenkov Wakefield Radiation in a Metamaterial Structure.

Physical review letters·2019
Same author

In Situ Observation of Dark Current Emission in a High Gradient rf Photocathode Gun.

Physical review letters·2016
Same author

Observation of Field-Emission Dependence on Stored Energy.

Physical review letters·2016

相关实验视频

Updated: Sep 15, 2025

Augmenting Large Language Models via Vector Embeddings to Improve Domain-Specific Responsiveness
03:14

Augmenting Large Language Models via Vector Embeddings to Improve Domain-Specific Responsiveness

Published on: December 6, 2024

691

PT-BitNet:扩大1-Bit大语言模型的规模,使用后训练量化.

Yufei Guo1, Zecheng Hao2, Jiahang Shao3

  • 1Intelligent Science & Technology Academy of CASIC, China.

Neural networks : the official journal of the International Neural Network Society
|July 16, 2025
PubMed
概括

对于大语言模型 (LLM) 来说,PT-BitNet可以实现三元量子化,最大可达70B的参数,无需重新训练. 这种后训练方法可以显著减少模型大小和推断时间,同时保持高准确度.

关键词:
有效的推理推理.大型语言模型.培训后的量化定量化三级量化定量化是第三级的.

更多相关视频

Large Scale Energy Efficient Sensor Network Routing Using a Quantum Processor Unit
05:30

Large Scale Energy Efficient Sensor Network Routing Using a Quantum Processor Unit

Published on: September 8, 2023

664
Objectification of Tongue Diagnosis in Traditional Medicine, Data Analysis, and Study Application
05:56

Objectification of Tongue Diagnosis in Traditional Medicine, Data Analysis, and Study Application

Published on: April 14, 2023

2.6K

相关实验视频

Last Updated: Sep 15, 2025

Augmenting Large Language Models via Vector Embeddings to Improve Domain-Specific Responsiveness
03:14

Augmenting Large Language Models via Vector Embeddings to Improve Domain-Specific Responsiveness

Published on: December 6, 2024

691
Large Scale Energy Efficient Sensor Network Routing Using a Quantum Processor Unit
05:30

Large Scale Energy Efficient Sensor Network Routing Using a Quantum Processor Unit

Published on: September 8, 2023

664
Objectification of Tongue Diagnosis in Traditional Medicine, Data Analysis, and Study Application
05:56

Objectification of Tongue Diagnosis in Traditional Medicine, Data Analysis, and Study Application

Published on: April 14, 2023

2.6K

科学领域:

  • 人工智能的人工智能
  • 计算机科学 计算机科学
  • 机器学习 机器学习

背景情况:

  • 大型语言模型 (LLM) 面临硬件和成本限制.
  • 量化技术为减少LLM资源需求提供了一个解决方案.
  • 比特网证明了三元量子化的有效性,但需要从头开始培训.

研究的目的:

  • 介绍PT-BitNet,这是LLMs的培训后量化方法.
  • 将BitNet的三元量子化优势扩展到大规模模型 (高达70B参数).
  • 减少LLM大小和推断延迟,以最小的性能降低.

主要方法:

  • 开发了一种两阶段的训练后定量化算法.
  • 第一个阶段:为了量子化兼容性,转换重量分布.
  • 第二阶段:以区块智能的方式优化量化权重.

主要成果:

  • PT-BitNet成功量化了高达70B个参数的LLM.
  • 实现了模型大小和推断时间的大幅减少.
  • 保持高任务性能,例如,70B模型下游任务的准确率为61%.

结论:

  • PT-BitNet有效地将三元量子化扩展到大型LLM.
  • 在大型模型上表现优于之前的方法,如BitNet b.158 (51.2%准确率).
  • 提供了一个可行的解决方案,用于部署高效和高性能的LLMs.