GRAPH EXECUTION PIPELINE PARALLELISM METHOD AND APPARATUS FOR NEURAL NETWORK MODEL COMPUTATION

WIPO Patent Application WO/2023/082575

A1

Provided in the present invention are a graph execution pipeline parallelism method and apparatus for neural network model computation, and provided are a graph execution pipeline parallelism method and apparatus for neural network model computation in a deep learning training system. The method comprises a graph execution flow during a process for neural network model computation, and a process during which functional modules work cooperatively. According to the graph execution pipeline parallelism method for neural network model computation, graph execution bodies on a local machine are created according to a physical computational graph generated by means of compiling a deep learning framework, and a scheme for allocating a plurality of free memory blocks to each graph execution body is designed, such that the entire computational graph simultaneously participates in deep learning training tasks of different batches of data by means of pipeline parallelism, thereby fully increasing the utilization rate of a memory and the parallelism rate of data.

WANG HONGSHENG (CN)
TAN BOWEN (CN)
BAO HUJUN (CN)
CHEN GUANG (CN)

PCT/CN2022/092481

May 19, 2023

May 12, 2022

Click for automatic bibliography generation  

ZHEJIANG LAB (CN)

G06N3/04; G06F12/02

CN114548383A	2022-05-27
CN114237918A	2022-03-25
CN114139702A	2022-03-04
CN114186687A	2022-03-15
CN112884086A	2021-06-01
US20190362227A1	2019-11-28

BEIJING ZHILIN HENGYUAN INTELLECTUAL PROPERTY AGENCY CO., LTD. (CN)

View/Download PDF      PDF Help

Previous Patent: METHOD FOR SYNTHESIZING SELAMECTIN

Next Patent: DISTRIBUTED DATA ROUTING MODELING METHOD AND APPARATUS FOR NEURAL NETWORK COMPUTATION