TorchTPU: Running PyTorch Natively on TPUs at Google Scale
mji
105 points
4 comments
April 23, 2026
Related Discussions
Found 5 related stories in 110.5ms across 10,002 title embeddings via pgvector HNSW
- The eighth-generation TPU: An architecture deep dive meetpateltech · 67 pts · April 22, 2026 · 63% similar
- Our eighth generation TPUs: two chips for the agentic era xnx · 427 pts · April 22, 2026 · 61% similar
- Google to sell TPU chips to select customers gmays · 23 pts · May 06, 2026 · 57% similar
- NanoGPT Slowrun: 10x Data Efficiency with Infinite Compute sdpmas · 122 pts · March 19, 2026 · 51% similar
- Google unveils chips for AI training and inference in latest shot at Nvidia wslh · 12 pts · April 22, 2026 · 50% similar
Discussion Highlights (3 comments)
in-silico
This is great to see. I did trained some research models using the existing PyTorch/XLA on TPUs, and it was a mess of undocumented behavior and bugs (silently hanging after 8 hours of training!). If anyone is trying to use PyTorch on TPU before TorchTPU is released, you can check out the training pipeline that I ended up building to support my research: https://github.com/aklein4/easy-torch-tpu
Reubend
Sounds good, but my main question is: is this a fork, or a new backend they're building in (like MPS)?
noracists
Very excited for this.