Model Deployment Method via Layer Segmentation for Chip Efficiency
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The complexity of current machine learning models makes it difficult to deploy them efficiently on chips due to high computing resource requirements and complex deployment processes.
Innovation Solution
The method involves dividing a machine learning model into multiple modules based on property information corresponding to its layers, generating a deployment file for each module, and sending this file to a chip for deployment, allowing for efficient allocation of computing resources and parallel processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a complete machine learning model is deployed directly to a chip, then the model can perform comprehensive tasks, but the data amount and computing resource requirements become excessively large
Solution Approach 1:
The patent divides a machine learning model into multiple sub-models based on layer property information, where each sub-model contains a subset of layers with similar characteristics. This segmentation reduces the data amount required for deploying each individual sub-model to the chip, while the collection of all sub-models maintains the complete model's functionality. The deployment file organizes layer information into multiple sub-model definitions, enabling modular deployment.
2Adaptability or versatility
If a complete machine learning model is deployed directly to a chip, then the model can perform comprehensive tasks, but the deployment process becomes overly complex and consumes excessive computing resources
Solution Approach 1:
The patent segments the model deployment process into manageable sub-models, making the deployment process less complex. Each sub-model can be independently configured and deployed, simplifying the overall deployment workflow compared to deploying a monolithic model. The deployment file structure with sub-model definitions enables organized management of model components.
Solution Approach 2:
The patent applies local quality by grouping layers with similar property information together to form sub-models. This allows each sub-model to have optimized characteristics suited for specific hardware resources on the chip, enabling more efficient resource allocation and reducing deployment complexity. Each sub-model can be tailored to match specific computational units or memory configurations.
3Productivity
If model layers are grouped into sub-models based on property information, then resource allocation becomes more efficient, but the model structure becomes more segmented
Solution Approach 1:
The patent accepts the structural segmentation as a necessary trade-off for improved deployment efficiency. The segmentation is performed systematically based on layer property information, creating a structured organization that facilitates efficient resource allocation on the chip. The deployment file maintains clear definitions of each sub-model and its associated layers, providing manageability despite the increased structural complexity.
Data Source
AI summary
The present disclosure applied to a model deployment field provides a model deployment method, a model deployment device and a terminal equipment. The model deployment method includes: dividing a preset to-be-deployed model into at least two to-be-deployed modules, according to property information respectively corresponding to at least two layers of the preset to-be-deployed model; generating a deployment file according to each of the at least two to-be-deployed modules, wherein the deployment file includes deployment information configured for deploying each of the at least two to-be-deployed modules; and sending the deployment file to a specified chip to indicate the specified chip to deploy the to-be-deployed model according to the deployment file. The present disclosure can efficiently deploy an algorithm model to a chip.