What changed
AWS and RadixArk released SGLang-Neuron 0.1.0 as a public preview container and out-of-tree plug-in. It runs SGLang serving and reinforcement-learning rollouts on Trainium through native PyTorch, with asynchronous execution and prefill-decode disaggregation on supported configurations.
Why it matters
Open-model teams can evaluate an established serving interface on AWS's accelerator rather than rewriting application code for a chip-specific stack. The preview also widens competition for large-model inference and rollout infrastructure.
What to do
Run the published container on one supported model, pin the exact SGLang commit and Neuron SDK, then compare throughput, latency, accuracy and total instance cost with the current GPU path before committing infrastructure.
Caveat
This is a 0.1.0 public preview with a pinned SGLang revision, specific Python and Neuron dependencies and a limited validated model and instance matrix. AWS publishes no independent price-performance result in the launch.