A task execution method, device, storage medium and computer program product
CN122173259APending Publication Date: 2026-06-09ALIBABA CLOUD COMPUTING CO LTD
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ALIBABA CLOUD COMPUTING CO LTD
- Filing Date
- 2024-12-06
- Publication Date
- 2026-06-09
AI Technical Summary
Technical Problem
The serial processing of probability sorting and cumulative probability calculation in the kernel sampling task leads to insufficient execution efficiency, affecting the task execution efficiency.
Method used
A parallel search method is used to filter out target confidence scores greater than the candidate threshold in the second candidate set. The target confidence scores that meet the kernel sampling conditions are found by iteratively adjusting the candidate threshold, thus eliminating the traditional serial sorting and cumulative probability calculation steps.
Benefits of technology
It improves the efficiency and accuracy of kernel sampling tasks, reduces processing time, and enhances the control over the diversity and coherence of generated text by large language models.
✦ Generated by Eureka AI based on patent content.
Smart Images
Figure CN122173259A_ABST
Abstract
Embodiments of the present application provide a task execution method and device, a storage medium and a computer program product. In the embodiments of the present application, the traditional kernel sampling scheme is abandoned, and a new kernel sampling scheme is proposed. In the embodiments of the present application, on the basis of introducing a candidate boundary value, a parallel search method can be used to filter out a part of target confidence degrees that may meet the kernel sampling condition in the second candidate set. In this way, by adjusting the candidate boundary value in a loop, the target confidence degree that can meet the kernel sampling condition can be found, and then the target element can be sampled from the first candidate set. In each round of filtering the target confidence degree, the parallel search method can be used, and therefore, the processing efficiency of the internal part of this link can be effectively improved, and the execution efficiency of the kernel sampling task can be improved.
Need to check novelty before this filing date? Find Prior Art