Reported launch
Google open-sources TorchTPU for native PyTorch on TPUs

- Google open-sourced TorchTPU, a native PyTorch stack for its TPUs, at the PyTorch Conference 2026, co-developed with
Meta.
- Google also announced TPU support in the vLLM and SGLang serving engines, the TorchTitan training framework, Ray, and
Hugging Face's libraries.
- Meta says it already runs LLM, ads and recommendation workloads on TPUs through TorchTPU in key production use cases.
- Google partnered with Hugging Face so that open-weight models on the Hugging Face Hub run natively on TorchTPU, with a simple device_map="tpu" parameter in Transformers v5.
Amazon released its own PyTorch variant, TorchNeuron, earlier the same week, allowing AI apps to run on its Trainium chips with no code rewrite.
Source details
- Latest report
- 19 Oct 2026 13:00 UTC
- First seen here
- 20 Oct 2026 19:25 UTC
- Outlets
- 10