TorchTPU: PyTorch Natively Runs on Google TPU

Episode 1 · Google and Meta Open-Source TorchTPU, Bringing Native PyTorch Support to TPU (2026-10-20, 15 posts)

At PyTorch Conference 2026, Google announced the open-sourcing of TorchTPU, a native PyTorch-on-TPU stack jointly developed by Google and Meta's PyTorch teams. It makes TPU a first-class PyTorch device backend alongside GPU and CPU, allowing existing PyTorch workloads to run directly on TPU with minimal code changes and familiar PyTorch APIs. The release is widely seen as a move by Google, together with Amazon, to challenge NVIDIA's dominance in AI chips.

Confirmed

• TorchTPU was jointly developed by Google and Meta's PyTorch teams, released as an open-source project providing a native, principled TPU backend (@PyTorch, @rseroter).
• The new stack replaces the previous PyTorch/XLA Lazy Tensor-based compiler queuing approach, making TPU a PyTorch device alongside GPU/CPU (@rohanpaulai).
• Existing PyTorch workloads can migrate to TPU with minimal code changes (@rseroter).
• On performance, training Qwen3 8B reached 52% of peak compute (@rohanpaulai).
• The release also comes with a set of other TPU-related updates (@rseroter).
• PyTorch powers roughly 85% of AI applications, so the ecosystem impact is significant (background info relayed by @gaganghotra).

Why it matters

• PyTorch is the dominant framework for AI research and development today, and native TPU support dramatically lowers the barrier for developers to use Google's own accelerators, reducing dependence on the NVIDIA GPU ecosystem.
• Combined with Amazon's earlier moves, cloud providers are competing for the AI compute market through openness at the framework layer, and TorchTPU is a key step in that competitive landscape.

• Google open-sources TorchTPU to run PyTorch workloads on TPUs with minimal code changes — rseroter · 2026-10-20
• Google and Meta launch TorchTPU, a native PyTorch backend for TPUs — PyTorch · 2026-10-21
• Google open-sources TorchTPU: native PyTorch on TPUs hits 52% peak training Qwen3 8B — rohanpaul_ai · 2026-10-21
• Google open-sources TorchTPU: native PyTorch on TPUs, 52% of peak training Qwen 3 8B — rohanpaul_ai · 2026-10-21
• Google open-sources TorchTPU, joining Amazon in loosening Nvidia's grip on PyTorch developers — gaganghotra_ · 2026-10-21
• TorchTPU goes open source as PyTorch officially embraces Google TPUs — beffjezos · 2026-10-21
• TorchTPU makes Google TPUs native PyTorch devices, supported across Hugging Face — GlennCameronjr · 2026-10-21
• Google and Meta launch TorchTPU, an open-source PyTorch backend that ports models to TPUs with one line — lmoroney · 2026-10-21
• Google and Meta open-source TorchTPU, a native PyTorch backend that runs models on TPUs with one line changed — lmoroney · 2026-10-21
• Google open-sources TorchTPU with Meta: up to 52% MFU on 8,000 Ironwood TPUs — 智东西 · 2026-10-21
• Google open-sources TorchTPU: run PyTorch natively on TPU with minimal code changes — nikola_mr64990 · 2026-10-21
• Google open-sources TorchTPU with Meta, letting PyTorch workloads run natively on TPUs — mark_k · 2026-10-21
• Google and Meta open-source TorchTPU: native PyTorch backend for TPUs — PyTorch · 2026-10-22
• Google and Meta launch TorchTPU, a native PyTorch backend for TPUs — PyTorch · 2026-10-22
• Google and Meta open-source TorchTPU, bringing native PyTorch support to TPUs — PyTorch · 2026-10-22

Episode 2 · Meta runs AI inference on Google TPUs in production (2026-10-21, 2 posts)

Google released TPU-native vLLM and SGLang inference engines, with Meta confirming production use of TPUs; the new SGLang-TorchTPU plugin reaches 770 tok/s single-stream decoding in Meta's production.

• Meta engineer confirms Google TPUs now run AI inference in Meta production — dotey · 2026-10-21
• SGLang-TorchTPU launches with Google and Meta, hits 770 tok/s decode in production — ying11231 · 2026-10-22