🧰 PyTorch XLA

The items that name it 4

• Amazon’s Trainium Takes a Strike at the Training King
  article2026-10-23It is written here as PyTorch/XLA.

  Older route to run PyTorch on Trainium using lazy tensor graph capture. It had no first-class eager mode and no native torch.distributed, DTensor, SimpleFSDP or DDP support.

• Hugging Face ❤️ TorchTPU: TPUs, Natively
  article2026-10-20It is written here as PyTorch/XLA (torch_xla).

  The older way to run PyTorch on TPUs. It used lazy tensors, could trigger graph recompilations on small changes, and was hard to debug.

• TorchNeuron: running HF models on AWS Trainium is easier than ever
  article2026-10-19

  Compiler-style backend underlying the earlier Neuron stack: lazy tensors, static shapes, recompilation on graph breaks.

• TPU Inference Externalization Full Steam Ahead - InferenceX
  article2026-09-07It is written here as PyTorch/XLA.

  First vLLM TPU prototype backend, using lazy execution to build graphs for XLA.