Towards Device-agnostic PyTorch: Building Unified Infrastructure for a Multi-Backend Ecos...

on YouTube ↗05m

" summary

Cambricon describes hardening PyTorch for PrivateUse1 backends by contributing device-agnostic fixes upstream: over 100 pull requests across PyTorch, NiteI, and torchvideo with 85 merged, moves that make torch.mlu feel native with eager, torch.compile and distributed training supported. Changes touch Inductor to simplify backend codegen, add PrivateUse1 support to FSDP and device-agnostic checkpointing, and generalize low‑level abstractions so profilers, FSDP and Inductor work out of the box for multiple accelerators.

AI-written summary · about

" related talks

• 18mtorch.accelerator: A Unified, Device-Agnostic Runtime API for Stream-Based AcceleratorsYu Guangye · PyTorch Day China 2025
• 13mLightning Talk: Challenges and Standardization in PyTorch Ecosystem...Zesheng Zong, Ashok Emani · PyTorch Conference 2025
• 23mHugging Face Accelerate: Making Device-Agnostic ML Training and Inference Easy...Zachary Mueller · AI_dev: Open Source GenAI & ML Summit Europe 2024
• 08mSponsored Session: Lightning Talk: PyTorch in Production: Boosting LLM Training and...F. Hua · PyTorch Conference 2025
• 12mLightning Talk: State of PyTorchAlban Desmaison · PyTorch Conference 2023
• 48mFireside ChatSoumith Chintala, Ibraham Haddad · Emerging OS Forum @ Open Source Summit Europe 2022
• 24mSponsored Session: PyTorch Distributed and Fault ToleranceTristan Rice · PyTorch Conference 2025
• 47mGet to Know PyTorch - A Cutting Edge Open Source AI Framework for Deep LearningSahdev P. Zala · Open AI + Data Forum @ Open Source Summit North America 2024