_version_ 1866911719405649920
author Lab, Mind
:
Cao, Song
Cao, Vic
Chen, Andrew
Chen, Kaijie
Cheng, Cleon
Chiang, Steven
Fan, Kaixuan
Feng, Hera
Feng, Huan
Fu, Arthur
Gao, Jun
Gu, Hongquan
Guan, Aaron
Ho, Nolan
Hong, Mutian
Hou, Hailee
Hua, Peixuan
Huang, Charles
Jiang, Miles
Jiang, Nora
Jiang, Yuyi
Jin, Qiuyu
Kong, Fancy
Lei, Andrew
Lei, Kyrie
Li, Alexy
Li, Lucian
Li, Ray
Li, Theo
Li, Zhihui
Lin, Jiayi
Liu, Kairus
Liu, Kieran
Liu, Logan
Liu, Xiang
Lu, Irvine
Luo, Maeve
Lv, Runze
Ma, Pony
Niu, Verity
Qiu, Anson
Wang, Vincent
Yang, Rio
Yao, Maxwell
Ye, Carrie
Ye, Regis
Ye, Wenlin
Ying, Josh
Zeng, Danney
Zhan, Yuhan
Zhang, Anya
Zhang, Di
Zhang, Ruijia
Zhang, Sueky
Zhang, Ya
Zhao, Wei
Zhou, Ada
Zhou, Changhai
Zhou, Yuhua
Zhu, Xinyue
Zhuang, Murphy
author_facet Lab, Mind
:
Cao, Song
Cao, Vic
Chen, Andrew
Chen, Kaijie
Cheng, Cleon
Chiang, Steven
Fan, Kaixuan
Feng, Hera
Feng, Huan
Fu, Arthur
Gao, Jun
Gu, Hongquan
Guan, Aaron
Ho, Nolan
Hong, Mutian
Hou, Hailee
Hua, Peixuan
Huang, Charles
Jiang, Miles
Jiang, Nora
Jiang, Yuyi
Jin, Qiuyu
Kong, Fancy
Lei, Andrew
Lei, Kyrie
Li, Alexy
Li, Lucian
Li, Ray
Li, Theo
Li, Zhihui
Lin, Jiayi
Liu, Kairus
Liu, Kieran
Liu, Logan
Liu, Xiang
Lu, Irvine
Luo, Maeve
Lv, Runze
Ma, Pony
Niu, Verity
Qiu, Anson
Wang, Vincent
Yang, Rio
Yao, Maxwell
Ye, Carrie
Ye, Regis
Ye, Wenlin
Ying, Josh
Zeng, Danney
Zhan, Yuhan
Zhang, Anya
Zhang, Di
Zhang, Ruijia
Zhang, Sueky
Zhang, Ya
Zhao, Wei
Zhou, Ada
Zhou, Changhai
Zhou, Yuhua
Zhu, Xinyue
Zhuang, Murphy
contents We present MindLab Toolkit (MinT), a managed infrastructure system for Low-Rank Adaptation (LoRA) post-training and online serving. MinT targets a setting where many trained policies are produced over a small number of expensive base-model deployments. Instead of materializing each policy as a merged full checkpoint, MinT keeps the base model resident and moves exported LoRA adapter revisions through rollout, update, export, evaluation, serving, and rollback, hiding distributed training, serving, scheduling, and data movement behind a service interface. MinT scales this path along three axes. Scale Up extends LoRA RL to frontier-scale dense and MoE architectures, including MLA and DSA attention paths, with training and serving validated beyond 1T total parameters. Scale Down moves only the exported LoRA adapter, which can be under 1% of base-model size in rank-1 settings; adapter-only handoff reduces the measured step by 18.3x on a 4B dense model and 2.85x on a 30B MoE, while concurrent multi-policy GRPO shortens wall time by 1.77x and 1.45x without raising peak memory. Scale Out separates durable policy addressability from CPU/GPU working sets: a tensor-parallel deployment supports 10^6-scale addressable catalogs (measured single-engine sweeps through 100K) and thousand-adapter active waves at cluster scale, with cold loading treated as scheduled service work and packed MoE LoRA tensors improving live engine loading by 8.5-8.7x. MinT thus manages million-scale LoRA policy catalogs while training and serving selected adapter revisions over shared 1T-class base models.
format Preprint
id arxiv_https___arxiv_org_abs_2605_13779
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle MinT: Managed Infrastructure for Training and Serving Millions of LLMs
Lab, Mind
:
Cao, Song
Cao, Vic
Chen, Andrew
Chen, Kaijie
Cheng, Cleon
Chiang, Steven
Fan, Kaixuan
Feng, Hera
Feng, Huan
Fu, Arthur
Gao, Jun
Gu, Hongquan
Guan, Aaron
Ho, Nolan
Hong, Mutian
Hou, Hailee
Hua, Peixuan
Huang, Charles
Jiang, Miles
Jiang, Nora
Jiang, Yuyi
Jin, Qiuyu
Kong, Fancy
Lei, Andrew
Lei, Kyrie
Li, Alexy
Li, Lucian
Li, Ray
Li, Theo
Li, Zhihui
Lin, Jiayi
Liu, Kairus
Liu, Kieran
Liu, Logan
Liu, Xiang
Lu, Irvine
Luo, Maeve
Lv, Runze
Ma, Pony
Niu, Verity
Qiu, Anson
Wang, Vincent
Yang, Rio
Yao, Maxwell
Ye, Carrie
Ye, Regis
Ye, Wenlin
Ying, Josh
Zeng, Danney
Zhan, Yuhan
Zhang, Anya
Zhang, Di
Zhang, Ruijia
Zhang, Sueky
Zhang, Ya
Zhao, Wei
Zhou, Ada
Zhou, Changhai
Zhou, Yuhua
Zhu, Xinyue
Zhuang, Murphy
Machine Learning
Artificial Intelligence
Distributed, Parallel, and Cluster Computing
We present MindLab Toolkit (MinT), a managed infrastructure system for Low-Rank Adaptation (LoRA) post-training and online serving. MinT targets a setting where many trained policies are produced over a small number of expensive base-model deployments. Instead of materializing each policy as a merged full checkpoint, MinT keeps the base model resident and moves exported LoRA adapter revisions through rollout, update, export, evaluation, serving, and rollback, hiding distributed training, serving, scheduling, and data movement behind a service interface. MinT scales this path along three axes. Scale Up extends LoRA RL to frontier-scale dense and MoE architectures, including MLA and DSA attention paths, with training and serving validated beyond 1T total parameters. Scale Down moves only the exported LoRA adapter, which can be under 1% of base-model size in rank-1 settings; adapter-only handoff reduces the measured step by 18.3x on a 4B dense model and 2.85x on a 30B MoE, while concurrent multi-policy GRPO shortens wall time by 1.77x and 1.45x without raising peak memory. Scale Out separates durable policy addressability from CPU/GPU working sets: a tensor-parallel deployment supports 10^6-scale addressable catalogs (measured single-engine sweeps through 100K) and thousand-adapter active waves at cluster scale, with cold loading treated as scheduled service work and packed MoE LoRA tensors improving live engine loading by 8.5-8.7x. MinT thus manages million-scale LoRA policy catalogs while training and serving selected adapter revisions over shared 1T-class base models.
title MinT: Managed Infrastructure for Training and Serving Millions of LLMs
topic Machine Learning
Artificial Intelligence
Distributed, Parallel, and Cluster Computing
url https://arxiv.org/abs/2605.13779