Programming · by Noah Johansson · 2026-10-20
Google open-sources TorchTPU, a native PyTorch backend for TPUs
github.com
Google has open-sourced TorchTPU, a native PyTorch backend for TPUs. For years using PyTorch on TPUs meant going through PyTorch/XLA, which worked but always felt like a second-class path compared to JAX. A native backend could make TPUs a much more realistic option for the huge number of teams that live in PyTorch. Has anyone tried it yet, and does this change your calculus on GPU vs TPU?
52 sparks · 18 comments
Comments
Amelia Hughes · 11 likes
anyone know what happens to pytorch/xla now? deprecated or coexisting?
Valeria Flores · 12 likes
would like an answer too, migration guides would be nice
Mango Baker · 9 likes
the github org name google-pytorch is funny to me for some reason
Zeynep Demir · 7 likes
if this actually works without the usual xla compile surprises, it's a big deal for anyone priced out of nvidia
Lina Keller · 3 likes
the compile surprises were my whole torch/xla experience. recompiling because a shape changed, pain
Ananya Rao · 0 likes
dynamic shapes on tpu remain my villain origin story
Diego Quispe · 6 likes
tried it for an hour, basic training loop ran. custom ops are where I expect things to break
Isha Menon · 1 likes
custom ops are where everything breaks, on every backend
Mina Chowdhury · 4 likes
the real question is availability. tpus only on google cloud, so this is also a cloud sales funnel
Lina Al Harbi · 0 likes
sure but competition with nvidia is good for everyone even if its funneled
Miguel Santos · 4 likes
for inference this could be interesting too, tpu pricing is decent if you can actually use it without rewriting everything
Rizky Pratama · 3 likes
Google finally admitting the world uses pytorch, not jax
João Pereira · 5 likes
jax is still lovely though, I'll die on that hill
Minji Kim · 0 likes
I'll believe it's native when torch.compile and the profiler both just work. that's my bar