
AWS Neuron 2.30.0 is now available with NKI 0.4.0, 22 new NKI Library kernels, and expanded Neuron Agentic Development skills for optimizing ML training and inference workloads on AWS Trainium and Inferentia accelerators.


<div><p>This article announces the general availability of AWS Neuron 2.30.0, a major release for ML developers building custom kernels and optimizing workloads on AWS Trainium and Inferentia accelerators.</p><ul><li>NKI 0.4.0 introduces activate2 Scalar Engine instruction and OCP FP8 input support</li><li>22 new NKI Library kernels added, including segmented attention and KV-parallel prefill</li><li>PyTorch reference implementations now available for 29 kernels</li><li>Neuron Agentic Development adds model porting and numerical equivalence validation skills</li><li>New Neuron DRA Driver enables topology-aware Kubernetes scheduling for accelerators</li><li>Neuron Graph Compiler improves compile-time performance significantly</li><li>Neuron Runtime enables zero-copy host-device transfers by default</li></ul><p>AWS Neuron 2.30.0 enhances ML development capabilities with new kernels, improved compiler performance, and automated model porting tools for Trainium and Inferentia accelerators.</p></div>


Related articles