On October 20, 2026, Google announced native PyTorch backends for vLLM and SGLang on its Tensor Processing Units (TPUs). In the announcement, Meta’s PyTorch-TPU lead, Zongwei Zhou, said SGLang-TorchTPU was running in production at Meta.
What Meta is running
The production statement names SGLang-TorchTPU, the SGLang inference engine’s TPU backend. Google also announced vLLM-TorchTPU, a separate backend developed with Inferact. The production statement identifies SGLang-TorchTPU; it does not describe Meta’s use of vLLM-TorchTPU.
What TorchTPU changes for PyTorch developers
TorchTPU lets PyTorch use a TPU as a native device, identified in code as tpu. Google says developers can run PyTorch eagerly or compile it with torch.compile through XLA, while distributed operations use torch.distributed.
Google says most models can retain standard upstream PyTorch definitions, with TPU-specific Pallas kernels and layer implementations providing optimizations. That gives PyTorch developers a route to TPU execution without replacing most model definitions. The SGLang backend was developed with RadixArk; Inferact partnered on vLLM-TorchTPU.
Earlier TPU backends for vLLM and SGLang used JAX. PyTorch/XLA also provided a bridge for PyTorch workloads. TorchTPU brings a native PyTorch device interface to the announced engines.
Cloud TPU access and chip purchases
Google says it launched Cloud TPU service for customers in early 2018. That service provides access to TPU computing capacity through Google Cloud. Cloud access and purchasing physical chips are different ways to use hardware; Zhou’s production statement concerns SGLang-TorchTPU running at Meta.
