⬢github C++ · 433 ★ +330 since we first saw it · pushed 9 min ago · NOASSERTION
google-pytorch/torch_tpu
TorchTPU is an open-source PyTorch backend for Google's Tensor Processing Units, letting you run PyTorch models directly on TPUs. It implements custom ATen kernels, caches XLA compilations for faster execution, supports distributed training (data and tensor parallelism), integrates with torch.compile(), and ships examples for models like Llama, Qwen, ResNet, and minGPT.
Why now: Google just open-sourced it, which is notable because TPUs previously required the XLA-based PyTorch/XLA bridge; a native backend for running PyTorch on TPUs is a big deal for ML practitioners.
Who it is for: ML researchers and engineers who want to train or run PyTorch models on Google TPU hardware.
Stars over our 59 snapshots: 103 to 433, since 15 h ago.
Where people talked about it
API: https://socialmediatrends-api.osmike.com/v1/repos/google-pytorch/torch_tpu