Memory information processing method, hardware acceleration module, equipment and storage medium
Through the first dynamic link library rewriting technology of the hardware acceleration module, the third dynamic link library of deep learning applications is accessed to obtain its memory usage information, which solves the memory overflow and resource waste caused by excessive GPU memory usage during deep learning model training, and realizes precise management of memory usage.
Patent Information
- Application Number
- CN202311284248.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-09-28
- Publication Date
- 2025-03-28
AI Technical Summary
During the deep learning model training process, model parameters, sample data used for training and intermediate result data occupy a lot of GPU memory, resulting in other processes being unable to apply for sufficient GPU memory, resulting in memory overflow or resource waste problems.
The first dynamic link library rewrite provided by the hardware acceleration module realizes access to the third dynamic link library owned by the application, and calls its target memory management function to obtain the application's current memory usage information, so as to accurately obtain the memory usage without intrusion of the application.
It solves the memory overflow and resource waste caused by unreasonable application memory usage, and realizes precise management of GPU memory usage.