DAILY BRIEF / HARDWARE
SGLang-Neuron brings SGLang serving to AWS Trainium.
What changed
A public-preview plugin runs SGLang on Trainium2 and Trainium3. HiCache L2 lifts GPT-OSS-120B output throughput 72-75% at 32K context.
Why it matters
Trainium becomes a candidate for SGLang workloads, including Kimi-K3 with prefill-decode split across three Trn3 nodes.
Scope and limitations
Preview only, supported through the AWS Deep Learning Container. No measured gains are given for the async runtime or megakernels.
Go to the source
A sourced technical brief published 18 Nov 2026. The implication is turingfold’s interpretation. We have not independently benchmarked this change.