MinT enables efficient training and serving of millions of LLM policy variants using LoRA adapters over shared base models. It achieves 18.3× faster handoff, 1.77× wall time reduction, and 8.7× loading speedup through time-sliced training, distributed export, and three-tier caching.
论文原始摘要
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.
Paper Collector 中文速览
百万级LoRA策略训练与托管基础设施
方法概述
采用Scale Up扩展至前沿架构，Scale Down仅传输1%大小的适配器，Scale Out分离持久化地址与工作集；通过tensor-parallel部署、GRPO并发多策略训练、MoE LoRA张量打包等技术，实现高效训练和服务
核心贡献
提出MinT系统，通过保持基础模型驻留、仅传输LoRA适配器，实现百万级策略目录的训练与服务，支持1T参数规模的Dense和MoE架构
原始来源与核验范围
核验仅覆盖论文正文中的主张、证据片段和参考文献关系。NGJOO 未独立运行作者代码、重做实验或验证真实部署效果；页面中的可复现性分数是文档完整度评估，不是复现实验结果。
核心问题
How can infrastructure efficiently train and serve millions of LLM policy variants over shared base models using LoRA adapters, addressing scalability across large model sizes, training-serving handoff efficiency, and policy-population management?
核心方法
MinT uses LoRA adapters as basic policy units, keeping base models resident while adapters represent different behaviors. The system implements time-sliced multi-LoRA training over shared bases, distributed export converting sharded training layouts to serving layouts, and a three-tier cache hierarchy (catalog, CPU, GPU) for multi-LoRA serving via vLLM.
方法组件
- One adapter type, export format, and serving path carry SFT loss, DPO reward margin, and GRPO train accuracy without any per-paradigm tooling change.
- SFT experiments show large held-out gains: +36 points on FinEval, +19 on FPB, +31 on TFNS over base Qwen3-4B.
- GRPO on Qwen3-8B-base achieves 20/30, 19/30, and 18/30 problems solved on AIME24, AIME25, and AIME26 eval sets respectively.
- MoE experiments confirm the adapter/base split survives distributed placement across tensor-parallel and expert-parallel workers on Qwen3-235B-A22B and Kimi K2 1T.
- AutoResearch uses proxy LawBench tasks to screen recipe candidates, with the maintained v23 weighted-aligned recipe reaching 0.5079 on full LawBench versus base 0.4628.
论点验证
可复现性评估
较低可复现性 (0%)
缺失的复现细节
- No code available - MinT infrastructure implementation not released
- No data available - training datasets, benchmarks, and evaluation data not released
- Missing core hyperparameters: learning rates, batch sizes, number of training epochs/steps for most experiments
- Missing LoRA configuration details: rank, alpha, dropout, target modules
- Missing optimizer settings: optimizer type, weight decay, gradient clipping, warmup schedule
- Missing random seeds for training runs (only seed 42 mentioned for one evaluation)
- Missing detailed model architectures and configurations for Qwen3 variants
- Missing data preprocessing and tokenization details
- Missing exact evaluation metric implementations (mean@1, reward margin calculations)
- Missing MinT system architecture details (adapter lifecycle, checkpoint format, export mechanisms)
局限与证据边界
- MinT currently lacks replay for every DSA indexer selection
- That mitigation keeps unstable tokens out of the gradient term; it does not reconstruct the exact sparse-attention token set selected by the inference engine
- Directly loading the MinT 1M packed adapter layout into native vLLM fails because the packed directory layout is not a standard PEFT tensor layout
- After the two-phase rollout result, we tested a more aggressive SLO-aware activation scheduler that changes prewarm concurrency from recent warm-traffic TTFT. The result is intentionally not used as a positive claim. It shortens rollout wall time, but the aggressive variants hurt warm tails
本分析由 PDF 阅读助手 自动生成，仅供参考，不构成学术评审意见。验证结论和可复现性评估基于论文文本自动分析，可能存在偏差。原始论文请参阅 arXiv。
分析时间：2026-05-27T07:10:51+00:00 · 数据来源：Paper Collector