➡️
继续阅读
-
Airbnb Cuts Authentication Code by 60% with Server Driven Architecture
Airbnb redesigned its authentication architecture around server driven flows ...
-
Presentation: From S3 to GPU in One Copy: Rethinking Data Loading for ML Training
Onur Satici explains how Vortex, an open-source columnar file format under th...
-
Tether: Apple Continuity Like Experience Between iOS and Linux Desktop Machines
Zack Bartel has developed Tether, an open-source project designed to integrat...
-
显存账本:每张卡上住着什么,搬走付什么
本文介绍训练大模型时显存管理的通用记账方法。显存分为五类:权重、梯度、优化器状态、激活和通信缓冲。前三类由参数量和并行度决定,激活随输入长度增长,优化空间...
-
看 NocoBase 的分支更新:AI 无代码平台进生产前,先把风险边界想清楚
NocoBase 的更新摘要提到 main、next、develop 三个分支。比起新功能本身,我更关心它进入真实项目后的升级、回滚、权限、审计和维护成本...
-
数据并行与 ZeRO:三级分片各省多少
本文介绍数据并行与ZeRO优化技术。ZeRO通过三级分片(优化器状态、梯度、权重)减少显存占用,前两级不增通信量,第三级需额外all-gather。ZeR...