Skip to content
Hot storiesContinuously updated

研究团队提出面向训练集群的LEAP吞吐量预测系统

1 reports1 reporting sourcesUpdated 16 hours ago

Start here

AI overview

2026年11月,研究团队提出面向生产训练集群硬件选择的学习型吞吐量预测服务LEAP。该系统可在作业提交时预测各类候选硬件的吞吐量,并按单位成本吞吐量排序以优化硬件路由选择。实际应用中,LEAP每天跨20多种硬件类型路由数万次训练作业,经30次夜间重训后准确率达77%,表现优于查表法及微调语言模型。在上线运行的6个月内,其覆盖的训练GPU工时比例已提升至95%,预估可节省约3%的训练计算成本。

AI-generated from reports · Updated 16 hours ago

Report timeline

Follow the reports to see different sides of the story.

Nov 4
  1. arXiv 人工智能
    LEAP:面向生产训练集群硬件选择的学习型吞吐量预测系统

    机器学习性能预测服务 LEAP 可在作业提交时预测各候选硬件类型的吞吐量,并按单位成本吞吐量排序以优化硬件路由选择。LEAP 每天跨 20 多种硬件类型路由数万次训练作业,经 30 次夜间重训后准确率达 77%,表现优于查表法与微调语言模型。在上线 6 个月内,其覆盖的训练 GPU-hours 比例提升至 95%,预估节省约 3% 的训练计算成本。

Story heat trend

Current heat 6·Comparable peak 10(Nov 4 14:00)·Comparable change, last 24 hours–

02.557.510Nov4Nov4Nov5Nov5

Trends compare only the same parties observed continuously, so the scope can be narrower than current heat counts. Move the pointer or tap the chart to see hourly heat; use the left and right arrow keys on a keyboard.