What happened
Hugging Face said TorchTPU adds TPU as a native PyTorch backend, so Transformers, Diffusers and Kernels run on TPU by setting device_map="tpu" — no torch_xla or Optimum TPU needed.
Why it matters
Developers can pick a model from the Hub and send it to "tpu" without maintaining a TPU fork, which should lower the barrier to using Hugging Face libraries on TPU hardware.
WHO IT HITSMachine learning engineers and developers who use Hugging Face libraries and want to run models on Google Cloud TPUs can now do so with regular PyTorch code, rather than porting to torch_xla or learning a different framework.
Summaries like this, in your inbox every morning.
Until now, running PyTorch on TPUs typically meant going through PyTorch/XLA (torch_xla), a path with quirks such as lazy tensors and hard-to-reason debugging. Hugging Face had built Optimum TPU to give its models inference and training support on TPUs, but that added another layer. TorchTPU changes the setup by making TPU just another device in PyTorch, so the same code runs without a TPU-specific fork. The body also notes that TPUs are Google's own AI accelerators, available on Google Cloud from a single chip up to pods of thousands of chips, and that the software stack historically grew up around TensorFlow and JAX. Hugging Face is testing the integration: of the tests that run on TPU, 96.5% pass, and the company has reported 34 issues on the TorchTPU repo, with 19 already fixed. Running on TPU also surfaced bugs in Transformers itself, such as in StaticCache and the Qwen3 tensor-parallel plans, which were fixed upstream.
Pick your industry and the AI tools you use, and get news related to your work every day.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.
After six rounds of tuning TypeSafe AI's Jev for note selection, the author benchmarked it on 100 requests aga…
A Zenn author migrating a ten-year-old PHP/Laravel monolith — hundreds of endpoints and more than 600 shared m…
Testing Pi 0.84.2 with claude-sonnet-4-5, a hook returning { block: true, reason } on rm produced a normal too…
OpenAI said GPT-6 Astra passed 87.39% and GPT-6.1 Sol 87.04% of its MultiLangBench test — building a SQLite-co…
DeepSeek released DeepInfer on GitHub under the Apache 2.0 license; it supports only DeepSeek-V4.1-Flash and r…
IBM, Red Hat and Google released llm-d version 1.0, adding multi-tenant quality of service, zero-downtime oper…
