Methods, devices, and storage media for handling computational cost in large model training

By preprocessing and analyzing performance credentials of large model training datasets, reliable performance credentials are generated and computing power costs are processed. This solves the problems of disconnect between billing and value and lack of green guidance in existing technologies, and realizes cost optimization for large model training and reliable billing in the market ecosystem.

CN122089415APending Publication Date: 2026-05-26SI-TECH INFORMATION TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
SI-TECH INFORMATION TECH CO LTD
Filing Date
2025-12-31
Publication Date
2026-05-26

AI Technical Summary

Technical Problem

The existing computing power billing model cannot adapt to the complexity and optimization needs of large model tasks, resulting in a disconnect between billing and value, a lack of green and efficiency guidance, and a lack of credible evidence chains to prove fee reductions, which can easily lead to disputes.

Method used

By preprocessing the original large model training dataset, performing performance certificate analysis, generating performance certificates, and processing computing power costs based on the certificates, the system includes an import module, a preprocessing module, a performance certificate analysis module, and a computing power cost processing result module. Blockchain technology is used for evidence storage to ensure the fairness and transparency of the incentive mechanism.

Benefits of technology

It reduced computing power expenditure, improved the overall utilization quality of large-scale computing power clusters, significantly reduced the total cost of training large models, increased throughput, and built a trustworthy computing power market ecosystem through green computing power incentives and a trustworthy billing mechanism.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122089415A_ABST
    Figure CN122089415A_ABST
Patent Text Reader

Abstract

This invention provides a method, apparatus, and storage medium for handling computational cost in large-scale model training, belonging to the field of large-scale model processing technology. The method includes: importing the original large-scale model training dataset; preprocessing the original large-scale model training dataset to obtain a preprocessed large-scale model training dataset; performing performance credential analysis on the preprocessed large-scale model training dataset to obtain performance credential analysis results; and performing computational cost analysis on the preprocessed large-scale model training dataset based on the performance credential analysis results to obtain computational cost processing results. This invention reduces computational expenditure, improves the overall utilization quality of large-scale computing clusters, reduces resource idleness and queuing congestion caused by inefficient tasks, improves throughput, and significantly reduces the total cost of large-scale model training.
Need to check novelty before this filing date? Find Prior Art