ADAPTATION METHOD AND APPARATUS FOR DISTRIBUTED TRAINING IN DEEP LEARNING FRAMEWORK AND AI ACCELERATION CARD

WIPO Patent Application WO/2023/103122

A1

Disclosed in the present invention are an adaptation method and apparatus for distributed training in deep learning framework and AI acceleration card. The method comprises the following steps: S1: a deep learning framework supporting single-card configuration in a newly-added AI acceleration card, the sub-steps comprising: S11: the deep learning framework supporting new hardware; S12, the deep learning framework supporting device threads of the new hardware; S13: the deep learning framework supporting memory operation of the new hardware; and S14: the deep learning framework supporting an operator kernel function of the new hardware; S2: the deep learning framework supporting multi-card configuration in the newly-added AI acceleration card; S3: the deep learning framework supporting tensor segmentation and multi-card distribution; and S4: the deep learning framework supports multi-card set communication in the newly-added AI acceleration card. In the present invention, a relationship between the deep learning framework and the newly-added AI acceleration card is established, the deep learning framework source code is comprehensively introduced to the chip bottom layer software, the hardware capability of the chip is released to the maximum extent, providing the strongest computing power for the end side AI.

WANG HONGSHENG (CN)
HUA WEI (CN)
BAO HUJUN (CN)

PCT/CN2021/142586

June 15, 2023

December 29, 2021

Click for automatic bibliography generation  

ZHEJIANG LAB (CN)

G06F9/50

CN109460827A	2019-03-12
CN113342346A	2021-09-03
CN113076143A	2021-07-06
US20200250525A1	2020-08-06
CN113918351A	2022-01-11

HANGZHOU QIUSHI PATENT OFFICE CO., LTD. (CN)

View/Download PDF      PDF Help

Previous Patent: VOLUTE CASING, AIR PASSAGE ASSEMBLY, AND AIR CONDITIONER

Next Patent: VOLUTE STRUCTURE, AIR DUCT ASSEMBLY AND AIR CONDITIONER