PyTorch
Tracking PyTorch itself as it evolves: release changes, new hardware backends, the torch.compile path and distributed training, so developers can judge what an upgrade really changes in their training and inference workflows.
PyTorch, the open-source framework stewarded by Meta, is where most AI research happens: new models and training libraries typically land on it first. Its focus is now shifting toward hardware portability. Through PrivateUse1, the official extension point for custom accelerators, third-party chips register as standard devices — and TorchNeuron, the native backend AWS just built for its Trainium chips, is the latest example.
