TorchTPU: Running PyTorch Natively on TPUs at Google Scale
by mji on 4/23/2026, 8:53:27 PM
https://developers.googleblog.com/torchtpu-running-pytorch-natively-on-tpus-at-google-scale/
Comments
by: in-silico
This is great to see.<p>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!).<p>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: <a href="https://github.com/aklein4/easy-torch-tpu" rel="nofollow">https://github.com/aklein4/easy-torch-tpu</a>
4/23/2026, 11:30:14 PM
by: aldielshala1
Looks like a new backend, similar to MPS for Apple Silicon?
4/24/2026, 12:26:09 AM
by: Reubend
Sounds good, but my main question is: is this a fork, or a new backend they're building in (like MPS)?
4/23/2026, 11:50:08 PM