开源项目

Dynamo贯通有界决策评分到OpenAI与System One接口

feat(decisions): expose compatible decision and systemone endpoints (#15854) / feat(decisions): coordinate bounded scoring with pinned worker placement (#15853) / feat(decisions): add bounded admission and local cancellation primitives (#15852) / feat(decisions): bind scoring capabilities to admitted model revisions (#15851)

AI 基础设施模型推理判别式推理与决策推理服务

概述

Dynamo本日合并决策服务链:注册支持因果评分的SGLang能力,把模型身份、tokenizer/模板、worker incarnation和放置选择绑定;首次分支使用配置的router,其后问题分支固定兼容worker/rank,限制并发和截止时间,并隔离请求内缓存。基础设施错误按整次请求失败,取消时本地分支排空前仍保留必要permit。 显式开启后提供POST /v1/decisions与/v1/systemone,分别按OpenAI和Jev/TypeSafe契约,普通chat仍可用。第一阶段拒绝所有nvext顶层及SGLang-native公共形状,后者仅作内部执行原语;作者披露小型真实Qwen模型与SDK检查,未完成27B质量、性能扩展、认证gateway或物理GPU结算资格验证。