AWS Neuron 2.32 introduces expanded NKI programming, MXFP8 training kernels, and variable-size collectives for Trn2 and Trn

AWS Neuron 2.32.0 introduces expanded NKI programming capabilities, MXFP8 training kernels, and variable-size collectives for Trn2 and Trn3 instances.

• NKI 0.6.0 adds on-device top-K instruction and variable-length all-gather for differently-sized tensors
• 13 new NKI Library kernels for Mixture of Experts training and sparse attention patterns
• MXFP8 forward and backward pass enables blockwise MoE layers to train end-to-end in MXFP8
• Variable-size all-gather, reduce-scatter, and all-to-all collectives on Trn2 and Trn3
• vLLM Neuron plugin upgraded to 0.24.0 with Neuron Agentic Development skill for transformer porting
• Neuron Compiler adds explicit 64-bit integer control; Explorer adds per-core CPU utilization tracking

These enhancements expand machine learning training and inference capabilities on AWS Trainium and Inferentia instances across all AWS regions.