相关实验视频
Updated: Sep 10, 2025

06:42
Generation and Coherent Control of Pulsed Quantum Frequency Combs
Published on: June 8, 2018
9.1K
ModuLoRA:通过与模块化量化器集成,对消费者GPU进行微调2位LLM
Junjie Yin1, Jiahao Dong2, Yingheng Wang3
1Department of Computer Science, Johns Hopkins University.
概括
我们介绍了ModuLoRA,这是一个使用单个GPU精确调整大型语言模型 (LLM) 的存储效率高的算法. 这种方法可以实现先进的低精度微调,在减少内存使用的情况下实现竞争性性能.
科学领域:
- 人工智能
- 机器学习
- 自然语言处理
背景情况:
- 大型语言模型 (LLM) 需要大量的计算资源进行微调.
- 现有的微调方法往往需要高端硬件,限制了可访问性.
- 低精度量子化提供了减少记忆足迹的途径,但也带来了微调挑战.
研究的目的:
- 为LLM开发一个内存高效的微调算法.
- 在消费级GPU上使用65B参数的LLM进行微调.
- 整合任意重量定量器与低级调整以实现灵活的微调.
主要方法:
- 提出了ModuLoRA (模块化低级调整),这是一个新的微调方法.
- 实现了量子化无关的向后传递,用于适应性低精度重量物质化.
- 集成的最先进的2位QuIP#和3位OPTQ量子化方法.
主要成果:
- 在单个24GB GPU上使用2/3/4位精度,成功调整了65B参数的LLM.
- 在文本分类,自然语言推断和指令执行任务方面取得了竞争性表现.
- 在总结任务中超越了最先进的ROUGE分数,超过了现有的4位和8位方法.
结论:
- ModuLoRA显著降低了LLM微调的内存需求.
- 该方法首次使高精度LLM (2位,3位) 的微调成为可能.
- 通过LLMTools库发布ModuLoRA和低精度模型,以实现更广泛的访问.
相关概念视频
Downsampling
252
When considering a sampled sequence with zero values between sampling instants, one can replace it by taking every N-th value of the sequence. At these integer multiples of N, the original and sampled sequences coincide. This process, known as decimation, involves extracting every N-th sample from a sequence, thereby creating a more efficient sequence.
The Fourier transform of the decimated sequence reveals a combination of scaled and shifted versions of the original spectrum. This...
The Fourier transform of the decimated sequence reveals a combination of scaled and shifted versions of the original spectrum. This...
252
Linear Approximation in Frequency Domain
131
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....
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....
131
Upsampling
309
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...
309
Fineness Modulus
784
The fineness modulus (FM) of aggregate is a numerical index that measures the coarseness or fineness of the particles. It is calculated by adding the cumulative percentages of aggregate retained on each of a specified series of sieves and dividing the sum by 100.
Consider performing sieve analysis on sand through a set of ASTM sieves. The weight of aggregate retained in each sieve and pan placed at the bottom is recorded, as given in Column B of Table 1.
To determine the fineness modulus of...
Consider performing sieve analysis on sand through a set of ASTM sieves. The weight of aggregate retained in each sieve and pan placed at the bottom is recorded, as given in Column B of Table 1.
To determine the fineness modulus of...
784
Per-Unit Sequence Models
116
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...
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...
116
Improving Translational Accuracy
2.7K
2.7K

