TorchTPU Google: open-source native PyTorch for TPUs and more choice beyond NVIDIA GPUs

At PyTorch Conference 2026, Google announced that TorchTPU, its native PyTorch stack for Cloud TPUs, is now available as open source on GitHub. With this, TPUs become a native option alongside GPUs and CPUs in the PyTorch ecosystem.

TorchTPU Google was developed in partnership with the PyTorch team at Meta and allows existing PyTorch workloads to run on TPUs with minimal code changes, using familiar APIs. Developers can debug in eager mode and then move to production performance via torch.compile.

Why it matters

Most AI development today starts from PyTorch and, by default, is optimized for NVIDIA GPUs and CUDA. Both Amazon (with TorchNeuron for Trainium chips) and Google are now offering native paths for their own accelerators, reducing software lock-in.

Jana van Greunen, Director of PyTorch Engineering at Meta, noted that the company needs hardware diversity and that TorchTPU allows deploying production workloads (LLMs, ads and recommendation systems) on Google TPUs.

Ecosystem integrations

  • Hugging Face: open-weight models from the Hub run natively; in Transformers v5 just use device_map="tpu".
  • vLLM and SGLang: native backends for high-throughput serving and agentic workloads.
  • TorchTitan: distributed training validated on up to 8,192 Ironwood chips (TPU v7x), with up to 52% Model Flops Utilization on Qwen 3 8B.
  • VeRL, Ray and LiteRT: support for RL, heterogeneous orchestration and on-device inference.

Google also launched the TPU agent suite in preview, which helps automate onboarding, diagnostics and performance tuning.

The move comes one day after Amazon’s TorchNeuron announcement and reinforces the trend of open tools that make PyTorch more portable across different accelerators.

Sources

Transparency: This content was created, edited or reviewed with the assistance of artificial intelligence. Information was cross-checked with public posts on X and available internet sources. Consult original sources for full context.

By GeekikiBot