Large Model Task Execution via Dynamic Modality and Field Routing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Large models, particularly those with Transformer architectures, face challenges in maintaining high processing efficiency while ensuring superior model performance, due to parameter redundancy and the need to run the entire model for each task execution.
Innovation Solution
A task execution method for large models that involves executing modality routing, field routing, and feedforward tasks using a target computing unit, where specific model parameters are read from a storage unit based on recognition results, optimizing the use of model parameters and improving inference efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the entire large model is executed for each task, then model performance is maintained, but processing efficiency deteriorates due to parameter redundancy
Solution Approach 1:
The patent segments the large model into multiple expert modules, each specialized in processing specific modality types (e.g., text, image, audio). Instead of executing the entire model for each task, the system routes tasks to specific expert modules based on modality recognition, thereby maintaining high model performance while significantly improving processing efficiency by avoiding parameter redundancy.
Solution Approach 2:
The patent implements a dynamic routing mechanism that determines which expert modules to activate based on the input task's modality characteristics. The routing system dynamically selects and activates only the necessary expert modules for each task, allowing the system to adapt its computational footprint based on task requirements rather than executing the full model statically.
2Reliability
If the entire large model is executed for each task, then model performance is maintained, but hardware resource requirements and energy consumption increase
Solution Approach 1:
The patent divides the large model into separate expert modules, each handling specific modality types. This segmentation allows the system to activate only the necessary modules for each task, reducing hardware resource requirements and energy consumption compared to executing the entire model, while still maintaining high model performance through specialized processing.
Solution Approach 2:
The patent applies partial action by executing only the necessary expert modules required for each specific task rather than the full model. This partial execution approach reduces energy consumption and hardware resource usage while maintaining sufficient model performance for the given task requirements.
Data Source
AI summary
A task execution method for a large model, an electronic device, and a storage medium are provided, which relate to a field of artificial intelligence technology, particularly to fields of deep learning technology and large model technology. The method includes: executing a modality routing task by using a target computing unit based on a target feature to be processed to obtain a modality recognition result; executing a field routing task by using the target computing unit based on the target feature to be processed and a target field gating model parameter to obtain a field recognition result; and executing a feedforward task by using the target computing unit based on the target feature to be processed and a target feedforward task model parameter to obtain a task execution result