Google and Meta launch TorchTPU, a native PyTorch backend for TPUs
PyTorch · x · 2026-10-21
PyTorch announces TorchTPU, an open-source collaboration between Google and Meta providing a native, principled PyTorch backend for Google TPUs.
- vs. prior efforts: PyTorch/XLA (Lazy Tensors) required intrusive source changes, suffered severe compile latency, and lacked eager-mode support; TorchAX shoehorned PyTorch into JAX. TorchTPU makes eager execution a top priority, letting existing workloads run out of the box on TPUs.
- Capabilities: true eager execution plus a torch.compile backend; fast inference via vLLM and SGLang; near-linear scaling for Llama 3 405B and Qwen 3 MoE training via TorchTitan.
Started at Google in mid-2025, this marks a major opening of the TPU ecosystem to PyTorch users.
Related event: Google and Meta Open-Source TorchTPU for Native PyTorch on TPU(2 posts)→
More from Infra
- ByteDance, Alibaba, Tencent order 3M+ Nvidia RTX PRO 5500 cards worth $35B, Reuters reports — SumitGup · 2026-10-21
- TorchTPU demo: unified vLLM & SGLang backend for LLM serving on TPUs — Rob_Mulla · 2026-10-21
- Google, Red Hat release llm-d 1.0: open K8s-native distributed inference stack — rseroter · 2026-10-21
- acs infra launches free base model API offering Llama 3.1 8B and 405B for LLM psychology researchers — amplifiedamp · 2026-10-21
- Developer Moves AI Agents to a Dedicated Linux Workstation Using Incus System Containers — kevinkern · 2026-10-21
- Nvidia guides 70% growth citing unmet AI demand, while media panics over a $500M Korean neocloud IPO — firstadopter · 2026-10-21