AWS Neuron 2.33 adds native PyTorch support, Neuron Observer, and other new features
AWS Neuron 2.33 introduces TorchNeuron, a native PyTorch backend for Trainium, in public beta, enabling standard PyTorch APIs on Trn2 and Trn3 instances.
AWS Neuron 2.33 is released, providing TorchNeuron, a native PyTorch backend for Trainium, in public beta. It enables standard PyTorch APIs to run models on Amazon EC2 Trn2 and Trn3 instances in eager mode or via torch.compile. Neuron Observer also enters public beta to detect Trainium health issues in under one second and identify affected nodes or devices. The vLLM Neuron plugin upgrades to vLLM 0.29.0 and integrates the vLLM Omni Neuron plugin for Wan2.2 video generation, both supporting native PyTorch compilation through TorchNeuron. New Neuron Deep Learning AMI and containers are available for PyTorch training, including TorchNeuron, PyTorch 2.14, and TorchTitan. Neuron Explorer opens profiles captured with torch.profiler and attributes allocations from PyTorch memory snapshots to Python source lines for off-memory error debugging. Kernels written in Neuron Kernel Interface (NKI) can use full Python language for metaprogramming, with the NKI library adding 25 new experimental kernels such as DeepSeek-V3.2, Qwen3.5, and Kimi Delta Attention.
Why it matters
This update impacts developers and data scientists using Trainium-based AI training and deployment with PyTorch, offering more seamless workflows and debugging capabilities.