Request processing method and apparatus, computer device, and readable storage medium
By dynamically adjusting the bundle width of the sequence generation model and merging input operations according to the number of requests in the request queue, the trade-off between server throughput and generated sequence quality is resolved, thereby improving server throughput and computing power allocation efficiency.
Patent Information
- Application Number
- CN202210805040.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-07-08
- Publication Date
- 2026-09-11
- Estimated Expiration
- 2042-07-08
AI Technical Summary
When deploying a sequence generation model on a server, how can we improve server throughput while ensuring the quality of the generated sequences, especially the trade-off between bundle width and sequence length in terms of inference speed and computing power requirements?
By obtaining the real-time number of requests to be processed in the request queue, the bundle width of the sequence generation model is dynamically adjusted. When the number of requests is greater than or equal to a preset number, the preset bundle width is used. When the number is less than the preset number, the temporary bundle width is determined based on the real-time number of requests. The input sequence generation model is then merged for operation.
It achieves improved server throughput and optimized server computing power allocation to adapt to different request volumes while ensuring output quality and inference speed.