A model training acceleration method and system based on computing power resource scheduling

By constructing a computing network collaborative scheduling model, resource consumption is monitored in real time, and computing power resource allocation is dynamically adjusted, which solves the problem of resource supply and demand mismatch in existing technologies and improves model training efficiency and resource utilization.

CN122363906APending Publication Date: 2026-07-10NINGBO INST OF INFORMATION TECH APPL CHINESE ACAD OF SCI
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
NINGBO INST OF INFORMATION TECH APPL CHINESE ACAD OF SCI
Filing Date
2026-04-15
Publication Date
2026-07-10

AI Technical Summary

Technical Problem

Existing computing resource scheduling methods lack awareness of the inherent characteristics of model training tasks, resulting in resource supply and demand mismatch, low utilization, and insufficient training efficiency.

Method used

By acquiring the structural features of the model, a computing network collaborative scheduling model is constructed using graph neural networks and reinforcement learning. This model monitors resource consumption data in real time, dynamically adjusts scheduling strategies, and achieves refined matching of computing resources.

Benefits of technology

It improves the resource utilization and efficiency of model training, and enhances the robustness and scalability of the system in dynamic environments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122363906A_ABST
    Figure CN122363906A_ABST
Patent Text Reader

Abstract

This invention discloses a method and system for accelerating model training based on computing power resource scheduling, applied in the field of artificial intelligence model training technology. The method includes the following steps: acquiring model training task requests and the structural features of the target model; inputting the model structural features into a pre-trained computing power demand prediction model; generating a distribution sequence of computing power resource demands at different stages of the target model's training process; fusing the computing power node information and network path information in the current computing power network; outputting a collaborative scheduling decision based on the fused features; matching and scheduling computing power resources with network paths according to the demand sequence and scheduling decision; executing the model training task; and monitoring the running data in real time during training to dynamically adjust the scheduling decision. This invention can accurately predict the dynamic demands of different models at each stage of training and continuously monitor and provide feedback during training, ensuring the stability and efficiency of the training process.
Need to check novelty before this filing date? Find Prior Art