Google and Meta Open-Source TorchTPU Native PyTorch Backend for Cloud TPUs

Google and Meta announced the open-source release of TorchTPU on October 20, 2026, at the PyTorch Conference. This native backend enables PyTorch models to run directly on Google Cloud TPUs without requiring the previous PyTorch/XLA infrastructure for new development projects.

The release includes four open repositories, with the main TorchTPU package available as version 0.2.0 on PyPI under the Apache 2.0 license. It requires Python 3.11 or newer and integrates via PyTorch’s PrivateUse1 interface, treating TPUs as standard devices alongside GPUs and CPUs.

Technical Architecture and Performance

TorchTPU operates differently from its predecessor by using ordinary PyTorch tensors without subclasses or wrappers, preserving tensor views and in-place mutations. It supports three execution modes: Strict for standard semantics, Debug for synchronous bounds checking, and DeferAndFuse for optimizing HBM traffic.

For production throughput, developers can use torch.compile with the TPU backend. This captures graphs via TorchDynamo and lowers them through StableHLO to OpenXLA, emitting cached TPU machine code. Custom Pallas kernels and prebuilt Tokamax kernels also function as standard PyTorch functions within this stack.

Google reports that TorchTitan, part of the suite, achieved 52% model flops utilization while training Qwen 3 8B on Ironwood chips. The system scaled nearly linearly across 8,192 TPU7x chips. Serving capabilities include support for FP8, MXFP4, and NVFP4 checkpoint formats.

Developer Impact and Migration

The new backend replaces PyTorch/XLA, which relied on lazy-tensor graphs and required developers to mark compilable regions. PyTorch/XLA also supported only pure SPMD code, whereas TorchTPU handles divergent MPMD execution, offering greater flexibility for complex model architectures and distributed training scenarios.

Existing PyTorch code can target TPUs using the standard torch.device("tpu") command. Google recommends TorchTPU for all new workloads, noting that it simplifies the development process by aligning TPU execution more closely with standard PyTorch behaviors and debugging workflows.

Distribution features include the tpu_dist process group, supporting DDP, FSDP2, and DTensor. The system also provides a topology-aware mesh for TPU interconnects, facilitating efficient communication across large-scale clusters without requiring significant code refactoring from existing distributed training setups.

What should you do?

Developers using PyTorch for TPU workloads should evaluate whether migrating to TorchTPU improves their specific performance metrics. Teams should review the four open repositories to understand integration requirements and compatibility with their current model architectures and training pipelines.

What to do and how to stay safe: Google

• Review the Apache 2.0 licensed TorchTPU v0.2.0 documentation to assess compatibility with existing PyTorch codebases.
• Test the DeferAndFuse execution mode in non-production environments to measure potential HBM traffic reductions.
• Evaluate torch.compile performance with the TPU backend for production workloads requiring high throughput.
• Monitor the four open repositories for updates on MPMD execution support and distributed training features.

Step-by-step guide: Cloud Misconfiguration Response: Containment, Recovery and Prevention Steps

General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

Does TorchTPU replace PyTorch/XLA?

Yes, Google states that TorchTPU replaces PyTorch/XLA and recommends it for all new workloads on Cloud TPUs.

What Python version is required for TorchTPU?

TorchTPU requires Python 3.11 or newer and is available as version 0.2.0 on PyPI.

Can TorchTPU handle MPMD execution?

Yes, unlike PyTorch/XLA which only supported pure SPMD code, TorchTPU handles divergent MPMD execution.