Understand this story
Google collaborated with Meta's PyTorch team to launch the open-source project TorchTPU, providing a native PyTorch backend for Google TPU; migration only requires changing model.to("cuda") to model.to("tpu"). The project supports both eager and torch.compile modes, adds a new DeferAndFuse eager mode, covers about 450 Aten operators, and has been verified to run out of the box on 4,917 Hugging Face models, with a Tier 1 pass rate of 98% and Tier 2 numerical consistency relative to GPU of 97%. The official statement says it has been used in the inference and training production workloads of some Google Cloud major customers, and is integrated with vLLM, SGLang, TorchTitan, Ray, and TorchRec. Subsequent reports said that at PyTorch Conference North America 2026, Google open-sourced TorchTPU, whose runtime no longer depends on JAX, and existing workloads can be migrated with only minor code changes; PyTorch noted that TorchTPU's default Skinny Collectives scheme suffers roughly 30–40% end-to-end performance loss under distributed workloads compared with optimized schemes, and achieving peak performance still requires Pallas kernels and vLLM-T
AI-generated from public reports
Latest development · Report
PyTorch Conference 2026 held in San Jose, announces progress on TorchTPU and TorchNeuronDevelopments
1 confirmed developments
- Source:PyTorch Blog
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Oct 23
PyTorch BlogFirst-party
PyTorch Conference 2026 held in San Jose, announces progress on TorchTPU and TorchNeuron
PyTorch Conference North America 2026 was held on October 20–21 in San Jose, attracting more than 3,500 attendees and focusing on accelerating open intelligence and projects including PyTorch, vLLM, DeepSpeed, Ray, Helion, and Safetensors. The event announced TorchTPU, a native PyTorch TPU backend developed in collaboration between Google and Meta, as well as TorchNeuron, a native backend for AWS Trainium, which entered public testing under Apache 2.0 on October 20. Shopify shared a case study of building a continuous learning loop with PyTorch and vLLM, saying it compresses production failures into model weights every day and has reduced serving costs by 96%.
Oct 22
The New Stack
Google open-sources TorchTPU, AWS launches TorchNeuron public beta
At PyTorch Conference North America 2026, Google open-sourced TorchTPU, a native PyTorch stack for TPU, while AWS released the open-source backend TorchNeuron public beta for Trainium; both are already available on GitHub. TorchTPU was developed by Google in collaboration with Meta's PyTorch team, and its runtime no longer depends on JAX, so existing workloads can be migrated with only minor code changes; TorchNeuron supports Trainium2 and Trainium3 instances, and migration only requires switching to .to('neuron') for eager execution. The PyTorch side points out that TorchTPU's default Skinny Collectives scheme suffers roughly 30–40% end-to-end performance loss on distributed workloads compared with optimized schemes, and achieving peak performance still requires Pallas kernels and production-grade stacks such as vLLM-TorchTPU and SGLang-TorchTPU.
Oct 21
PyTorch BlogFirst-partySelected
Google and Meta jointly launch TorchTPU, enabling PyTorch to run natively on TPUs
Google and Meta's PyTorch teams have collaborated to launch the open-source project TorchTPU, providing a native PyTorch backend for Google TPUs; migration only requires changing model.to("cuda") to model.to("tpu"). The project supports both eager and torch.compile modes, adds a new DeferAndFuse eager mode, covers about 450 Aten operators, and has been verified to run out of the box on 4,917 Hugging Face models, with a Tier 1 pass rate of 98% and Tier 2 numerical consistency relative to GPUs of 97%. According to the official announcement, it has already been used in the inference and training production workloads of some major Google Cloud customers, and is integrated with vLLM, SGLang, TorchTitan, Ray, and TorchRec.