苏乐
Employers on the program
- 2026支付宝1 talk
Events1
Conferences1
Talks1
Years active2026
Upcoming appearances
Who’s booked ahead? →Event dates shown; only events starting today or later count. 1 upcoming event.
-
AICon 全球人工智能开发与应用大会 | 北京 – · Beijing, China
- 大模型后训练：AReno 训推一体架构设计 speaker Program source ↗
No upcoming appearances are currently recorded.
The year around now
Talks in the past 12 months and announced talks ahead, one row per conference, January to December. Hover a mark for the talk.
Appearances by year
- AICon 全球人工智能开发与应用大会 (AICon Global AI Development and Application Conference)
Talks
| Date | Event | Conference | Role | Title | Links |
|---|---|---|---|---|---|
| Dec 18, 2026 | AICon 全球人工智能开发与应用大会 | 北京 | AICon 全球人工智能开发与应用大会 (AICon Global AI Development and Application Conference) | speaker | 大模型后训练：AReno 训推一体架构设计 as 支付宝 大模型后训练面临训练与推理衔接复杂、显存占用高及计算路径差异影响策略优化等问题。本次分享介绍 AReno 训推一体架构的设计与技术选型，通过共享参数与计算组件，统一支持监督微调、偏好优化和在线强化学习。重点解析 MoE 路由记录与重放、变长样本打包、分页缓存，以及低比特优化器、QLoRA 与 MST…Show less大模型后训练面临训练与推理衔接复杂、显存占用高及计算路径差异影响策略优化等问题。本次分享介绍 AReno 训推一体架构的设计与技术选型，通过共享参数与计算组件，统一支持监督微调、偏好优化和在线强化学习。重点解析 MoE 路由记录与重放、变长样本打包、分页缓存，以及低比特优化器、QLoRA 与 MST 分块重计算的组合方案。在 DGX Spark 上，Ling-3.0-tiny 模型 16K 上下文实测中，QLoRA 结合 MST 将训练峰值显存从 37.785 GiB 降至 11.973 GiB，降低 68.31%。同时介绍 AReno Flow 如何贯… |
Frequent co-speakers
No one else spoke at two or more of the same events.
Speakers with the same name are merged into one entry, so this page may combine more than one person. Affiliations are shown as each conference listed them.