Google open-sources TorchTPU for native PyTorch on Tensor Processing Units

Google has released TorchTPU as open-source software, providing PyTorch developers with a native path to execute models on Google's Tensor Processing Units rather than relying on translation layers or proprietary tools. The release accompanies expanded TPU support across multiple widely-used AI infrastructure components.

Google also announced that TPU support is now available in the vLLM and SGLang inference serving engines, the TorchTitan distributed training framework, the Ray distributed computing project, and Hugging Face's libraries. These integrations suggest Google is positioning TPU as a first-class citizen in the open-source AI tooling stack rather than a vendor-locked alternative to NVIDIA's CUDA ecosystem.

What TorchTPU enables and where it fits

• Native PyTorch backend: TorchTPU provides direct PyTorch execution on Google's TPUs without requiring developers to convert models through JAX or TensorFlow intermediaries.
• Inference engine integration: vLLM and SGLang, two popular open-source LLM serving engines, now support TPU as a compute backend for production inference workloads.
• Training framework support: The TorchTitan distributed training framework gains TPU compatibility, enabling large-scale model training on Google's cloud hardware.
• Ecosystem breadth: Ray and Hugging Face libraries now include TPU support, lowering friction for developers already using these tools to adopt Google's accelerators.

Why it matters

This release reduces vendor lock-in friction for AI developers who prefer PyTorch but were previously steered toward NVIDIA GPUs or Google's own JAX stack. For Linux-based AI infrastructure and cloud deployments, native PyTorch TPU support creates genuine hardware competition in the training and inference market, though Google has not yet disclosed whether TorchTPU's license permits independent cloud providers to offer TPU instances or if it remains tied to Google Cloud's proprietary hardware availability.