• Sources: primary, discussion
  • Summary: The torch_tpu repository provides ATen kernels for TPU, a compilation cache for XLA computations, data and tensor parallelism, and torch.compile integration. Version 0.2.0 installs from PyPI and supports Python 3.11 and later. It includes examples for Llama, Qwen, ResNet and DLRM.
  • Why it matters: PyTorch code gets a first-party route to TPUs through ATen and torch.compile, beside PyTorch/XLA.