MinT: Managed Infrastructure for Training and Serving Millions of LLMs

2026-05-13

AI Summary

This work addresses the substantial storage and serving overhead incurred by fine-tuning large language models with massive numbers of LoRA adapters by introducing the MindLab Toolkit (MinT), a hosting framework that shares a common base model while transmitting only lightweight LoRA adapters and managing their full lifecycle uniformly. Key innovations include support for catalogs of up to one million LoRA strategies, adapter compression to less than 1% of the base model size, decoupling of persistent storage from computational address spaces, and integration of tensor parallelism, GRPO-based concurrent multi-strategy training, batched MoE-LoRA loading, and cold-start-aware scheduling. Experiments demonstrate an 18.3× inference speedup on a 4B dense model and a 2.85× speedup on a 30B MoE model; a single engine can traverse 100,000 strategies, clusters support over a thousand concurrent requests, and MoE loading efficiency improves by 8.5–8.7×.

Abstract

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.

Problem

• LoRA
• large language models
• model serving
• scalable infrastructure
• parameter-efficient fine-tuning

Innovation

• LoRA
• managed infrastructure
• model serving
• MoE
• parameter-efficient fine-tuning