Request allocation method, system and equipment based on heterogeneous computing system and medium
By obtaining inference task information and heterogeneous computing power equipment performance parameters in heterogeneous computing system, determining the key-value cache reading time and selecting appropriate equipment, the problem of poor balance in inference request allocation is solved, and the saving of computing power resources and cost reduction is achieved.
Patent Information
- Application Number
- CN202510221607.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-27
- Publication Date
- 2025-03-25
- Estimated Expiration
- 2045-02-27
AI Technical Summary
In heterogeneous computing systems, the allocation balance of inference requests is poor, resulting in a lot of wasted computing resources, especially when calling separate memory or local memory.
By obtaining inference task information and performance parameters of heterogeneous computing power equipment, the reading time required for key-value cache is determined, and the appropriate heterogeneous computing power equipment is selected based on the reading time, the preset allocation number is determined based on the target reading time and performance parameters, and the request allocation result is determined based on the preset allocation number and the concurrent request number.
It improves the allocation balance of inference requests, saves computing resources, reduces costs, and speeds up the execution of inference tasks.