Native PyTorch TPU engines ship for vLLM and SGLang
On October 20, 2026 Google, RadixArk and Meta released PyTorch-native TPU backends for SGLang (enabled with SGLANG_PLATFORM=tpu_oot) and vLLM via TorchTPU, with Pallas paged-attention and MoE kernels and KV transfer over TPU Sync. Meta says SGLang-TorchTPU runs in production, and a Qwen3-4B GRPO trial on four TPU v7x chips matched a six-H200 setup.
Why it matters
TPUs become a first-class serving target for the two dominant open inference stacks, breaking NVIDIA's default assumption.
Importance: 3/5
notable infrastructure release bringing TPUs to both major open inference stacks