Google and Meta open-source TorchTPU for native PyTorch on TPUs

At PyTorch Conference North America 2026 in San Jose, Google and Meta open-sourced TorchTPU, a native PyTorch backend for Tensor Processing Units. Meta plans to move larger-scale LLM workloads onto TPUs, and four repositories were released, including vLLM-TorchTPU and SGLang-TorchTPU.

InfoQ

Open source

TorchTPU replaces PyTorch/XLA and needs only model.to("tpu") instead of model.to("cuda")
Qwen3 235B MoE ran on 8,192 TPU v7x chips at 28% model FLOPs utilization
TorchTPU, vLLM-TorchTPU, SGLang-TorchTPU and TorchTitan were open sourced
AWS released a similar TorchNeuron backend for Trainium in public beta

#Google#Meta#PyTorch#TPU

Google TPUs now run AI workloads in Meta's data centers

AI digest: OpenAI air-gaps training, Anthropic books up to 1M TPUs

MemTensor open-sources MemT 1.0 memory-native models with fixed-size state

Security

Google and Meta open-source TorchTPU for native PyTorch on TPUs

Google TPUs now run AI workloads in Meta's data centers

AI digest: OpenAI air-gaps training, Anthropic books up to 1M TPUs

MemTensor open-sources MemT 1.0 memory-native models with fixed-size state

Open-source keras-model-scan and pickle-model-scan check ML models without loading them