该仓库封存 ServingROM 第一代 POC 的两个正式数据集,用于研究 1P2D 大模型推理系统的全阶状态重建、POD/DMDc 降阶建模、运行时路由控制和 Control-aware ROM。
实验系统为 Qwen3.6-27B-w8a8,拓扑为 Prefill TP2 + Decode A TP2 + Decode B TP2,通过 Mooncake 传输 KV。Decode 使用 FULL_DECODE_ONLY 图模式和 async scheduling。数据不包含 prompt、对话正文、模型输出正文、模型权重或原始逐事件 telemetry。
配套代码、部署文件、Schema 与实验报告位于 panxy1019/ServeringROM。
rom-dataset-v1.1/
train/{X,X_next,D,Y,MU}.npy
validation/{X,X_next,D,Y,MU}.npy
test/{X,X_next,D,Y,MU}.npy… See the full description on the dataset page:
https://huggingface.co/datasets/panxy1019/ServeringROM.