Google open-sources TorchTPU: native PyTorch on TPUs hits 52% peak training Qwen3 8B
rohanpaul_ai · x · 2026-10-21
At PyTorch Conference 2026, Google open-sourced TorchTPU, a native PyTorch backend for TPUs that replaces the PyTorch/XLA compiler-queue approach. It executes ops eagerly for GPU-like step-by-step debugging, then switches to torch.compile for production speed; training Qwen3 8B reached up to 52% of peak TPU compute. Co-developed with Meta's PyTorch team, it offers custom ATen kernels, compilation caching, and distributed training, and works with vLLM, SGLang, llm-d, TorchTitan, Ray and Hugging Face. With Amazon offering a similar backend, PyTorch code can now leave Nvidia GPUs natively on both clouds.
Related event: Google and Meta Open-Source TorchTPU for Native PyTorch on TPU(4 posts)→
More from Infra
- Anthropic's draft IPO filing: $16.3B H1 revenue, $80B net loss, $750B compute plans — kimmonismus · 2026-10-21
- dphnAI's distributed inference network served 32B DeepSeek tokens in 24h, ~6% of Venice's daily volume — Ronangmi · 2026-10-21
- Jensen Huang: every company will have two factories, one real and one digital — round · 2026-10-21
- DeepSeek V4 called 'god-tier infra, 4th-tier ML': 7 tricks stacked, 20 patches to fix them — teortaxesTex · 2026-10-21
- HSBC says CoreWeave shares to drop 30%+ as ROIC of 8.9% lags borrowing costs — GaryMarcus · 2026-10-21
- RunInfra launches $10/mo coding plan: fastest open models across 19 agents incl. Claude Code — ycombinator · 2026-10-21