Latest: Google Open-Sources TorchTPU as Native PyTorch Backend for TPUs
Native PyTorch backends for TPUs reduce friction for developers by enabling familiar step-by-step debugging rather than compiler-oriented workflows. The TorchTPU integration with Hugging Face extends this accessibility to a broad base of ML practitioners.
When stories like this move, I write about them in an email each morning.
One free email each morning, around 8 AM US Eastern.
I have sent you a confirmation link. One click and you are in.
Not in your inbox? Look in spam or promotions. If it's there, mark it "Not spam" so future editions land in your inbox.
Google open-sourced TorchTPU, a native PyTorch backend for TPUs that replaces the older PyTorch/XLA layer. The previous approach queued operations for a compiler, making TPU debugging behave differently from standard GPU workflows. TorchTPU runs each operation eagerly for debugging, then switches to torch.compile for production performance. In training Qwen 3 8B, TorchTPU reached up to 52% of peak TPU compute. Hugging Face integrated TorchTPU into its ecosystem, making TPU usage accessible alongside existing Hugging Face tooling. Amazon also offers a native PyTorch backend that allows workloads to run outside Nvidia GPUs, though item coverage of the Amazon side is sparse.
Hugging Face published Part 1 of a multi-part profiling series for TorchNeuron, the native PyTorch backend for AWS Trainium it built with AWS to simplify deploying Hugging Face models on that accelerator.
The first installment covers system-level profiling techniques for models running on Trainium.
Google and Amazon both released native PyTorch backends for non-Nvidia hardware on October 19-20, 2026.
Google open-sourced TorchTPU, which replaces the older PyTorch/XLA layer by running operations eagerly for debugging then switching to torch.compile for production; in training Qwen 3 8B, TorchTPU reached up to 52% of peak TPU compute. Separately, Hugging Face announced TorchNeuron, built with AWS to simplify deploying Hugging Face models on AWS Trainium accelerators.
BackgroundFrontier models, open models, and how they are compared
When stories like this move, I write about them in an email each morning.
One free email each morning, around 8 AM US Eastern.
I have sent you a confirmation link. One click and you are in.
Not in your inbox? Look in spam or promotions. If it's there, mark it "Not spam" so future editions land in your inbox.