AWS Neuron 2.33 is released, featuring TorchNeuron, a native PyTorch backend for AWS Trainium, in public beta. Users can now run standard PyTorch models on Amazon EC2 Trn2 and Trn3 instances using eager mode or torch.compile. This update also introduces Neuron Observer in public beta to quickly detect and diagnose Trainium health issues. The vLLM Neuron plugin is upgraded to vLLM 0.29.0, and the vLLM Omni Neuron plugin for video generation is now integrated. Both plugins leverage TorchNeuron for native PyTorch compilation. New Neuron Deep Learning AMIs and Containers include TorchNeuron, PyTorch 2.14, and TorchTitan for PyTorch training. Neuron Explorer now attributes memory allocations to specific Python source lines for debugging out-of-memory errors. Kernels written using the Neuron Kernel Interface (NKI) can now utilize the full Python language for metaprogramming. The NKI Library has added 25 new experimental kernels, including those for recent large language models and attention mechanisms. These enhancements aim to simplify PyTorch training on AWS Trainium and improve debugging capabilities.