What changed
AWS Neuron 2.32 added on-device top-K, variable-length all-gather, data-dependent loops, thirteen MoE and sparse-attention kernels including DeepSeek V3.2 and MXFP8 paths, variable-size Trainium collectives, and an updated vLLM Neuron plugin.
Why it matters
These are concrete compiler and runtime building blocks for training and serving sparse or mixture-of-experts models on AWS Trainium hardware.
What to do
Pin Neuron 2.32 in isolation and compare one representative workload for accuracy, throughput, compiler stability, memory, cost, and rollback against the current stack.
Caveat
The changes are Trainium-specific and AWS-reported. Independent performance evidence is unavailable, and compiler or runtime upgrades can regress model behavior and deployment stability.