DevelopingHardware
Google ships vLLM and SGLang engines for TPU on TorchTPU
Google announced open-source vLLM-TorchTPU and SGLang-TorchTPU engines that run on TorchTPU, making TPU a native PyTorch device. Google says both run at scale with preview customers. It cites 2.01 cents per million tokens for Qwen 3.5 on Ironwood in SemiAnalysis InferenceX. No pricing or GA date given.
- HEAT
- OUTLETS
- 0
- FIRST SEEN
- LAST SIGNAL
Primary source
Google for Developersdevelopers.googleblog.com/announcing-new-vllm-and-sglang-inference-engines-on-tpu/Coverage · 1 article, oldest first
Tuesday, October 20, 2026
Google for DevelopersPrimary
Announcing New vLLM and SGLang Inference Engines on TPU