Model publishing method, device and equipment, and storage medium

Abstract

The invention discloses a model publishing method. According to the method, a local + distributed storage model release scheme is provided; a complete deep learning model with a large space is segmented into a dense part and a sparse part; the sparse part occupying large space is deployed in the distributed cluster, the dense part occupying small space is deployed in the local computing cluster, and the model dispersion part is dispersedly pulled, so that the occupation of a memory by the model is reduced; the memory pressure of the local computing cluster is relieved when the model is loaded,and the memory consumption of computing nodes is greatly reduced. The invention further provides a model publishing device and equipment and a readable storage medium, which have the above beneficialeffects.

Description

technical field [0001] The present application relates to the technical field of system communication, and in particular to a model release method, device, equipment and a readable storage medium. Background technique [0002] With the increase of artificial intelligence application scenarios (such as voice, semantics, image, video, search, network, etc.), the application of deep learning models is becoming more and more extensive. [0003] At present, the commonly used model release method when releasing the deep learning model is: deploying the complete deep learning model to the local node, and then calling the model through a specific communication protocol, such as the http protocol. [0004] However, for large-scale deep learning models, such as those used in recommendation systems, this model release method can take up hundreds of gigabytes of space. There is a severe shortage of storage resources. For example, the memory of a server is 256G, and the scale of the de...

AI Technical Summary

Problems solved by technology

[0004] However, for large-scale deep learning models, such as those used in recommendation systems, this model release method can take up hundreds of gigabytes of space. Serious shortage of storage resources

For example, the memory of a server is 256G, and the scale of the deep learning model is 200G. At this time, the above model deployment method consumes a lot of memory during the loading process, but at the same time, the consumption of CPU and GPU may not be so large, so it will cause serious problems. CPU and GPU computing resources are wasted, and memory resources become a bottleneck limiting computing power; moreover, the server can only deploy one model instance, and the training efficiency is very low

Images