Torch-FL: Ending the Fragmentation Nightmare of Heterogeneous AI Chips
Date:
FlagOS livestream (replay) on Torch-FL (Pytorch-Plugin-FL). Different accelerators currently need different torch_* plugins, which fragments training and inference. Torch-FL aims at one PyTorch experience across backends, using FlagOS libraries such as FlagGems and FlagCX.
- Video (Bilibili)
- Related conference talk: Torch-FL: A Virtual Device Layer for PyTorch
中文概要
国产 AI 芯片各自维护不同的 Torch 插件，科研和商用训推都要重复适配。这场 FlagOS 直播介绍 Torch-FL：对外提供统一的 PyTorch 使用体验，并借助 FlagGems / FlagCX 等 FlagOS 组件，把硬件差异挡在插件层后面。