跳到正文
🚨 AI News | TestingCatalog· @testingcatalog · X·· 4 天前AI 评分48
AI 导读

SGLang-Neuron 以亚马逊的 Neuron 软件命名,目前处于公开预览阶段。 Trainium 版本旨在运行 GPT-OSS、Kimi-K3、GLM 和 Qwen 等开源模型。当芯片内存满时,它会将较早的聊天记录存储在常规服务器内存中,使 GPT-OSS-120B 的输出量增加超过 70%。 请看 👀 https://www.lmsys.org/blog/2026-11-12-sglang-neuron-v0.1.0

正文 · 原文

SGLang-Neuron is named after Amazon's Neuron software and is in public preview.

The Trainium version is designed to run open models like OpenAI's GPT-OSS, Kimi-K3, GLM, and Qwen. When chip memory fills up, it stores older chat history in regular server memory, giving GPT-OSS-120B over 70% more output.

Take a look 👀
https://www.lmsys.org/blog/2026-11-12-sglang-neuron-v0.1.0

来源:🚨 AI News | TestingCatalog · x.com