Skip to content

1634594707/VerDee

Repository files navigation

VerDee

Vertical Deep Network — 基于 Mamba-2 的纵向分层 LoRA、动态早退路由与领域专家;支持 16GB 单卡 Pilot(370M)与 Production(2.7B)实验。

VerDee: depth-staged LoRA (shallow/mid/deep), early-exit routing, and domain adapters on Mamba-2 — reproducible research for consumer GPUs.

文档

文件 说明
ARCHITECTURE.md 整体架构与目录说明
VerDee.md 完整方案(Production / 5060 Ti)
VerDee-Experiment.md 实验对照、评测、Pilot 流程

配置

文件 模型
configs/pilot_370m.yaml 370M,先跑这个
configs/production_2.7b.yaml 2.7B,Pilot 通过后

目录速览

configs/      超参
src/verdee/   核心代码包
scripts/      命令行入口
data/         数据集
eval/         评测集与报告
adapters/     LoRA 输出
checkpoints/  训练 checkpoint
logs/         日志

环境(已创建 .venv

PowerShell(在项目根目录):

.\scripts\activate.ps1
# 或:  .\.venv\Scripts\Activate.ps1

CMDscripts\activate.bat

快速开始

# 1. 激活环境后
python scripts/00_check_gpu.py      # GPU / sm_120
python scripts/01_verify_model.py   # 下载并验证 mamba2-370m

# 2. 准备数据(首次需联网下载 HF 数据集,约 10–30 分钟)
python scripts/prepare_data_pilot.py

# 3. Pilot 训练(按顺序)
python scripts/03_train_stage.py --stage shallow
python scripts/03_train_stage.py --stage mid
python scripts/03_train_stage.py --stage deep
# python scripts/03_train_stage.py --stage traditional_b   # 对照 B

# 3. 评测对比(待实现)
python scripts/run_benchmark.py --runs A,B,C1,C3,M --suite pilot

硬件

  • RTX 5060 Ti 16GB
  • PyTorch cu128(Blackwell sm_120)
  • 推荐 WSL2 Ubuntu 24.04

About

VerDee: Vertical Deep Network on Mamba-2 — staged LoRA (shallow/mid/deep), early-exit routing, and domain experts. Pilot on 370M; 16GB GPU friendly. Research/experiment repo.

Topics

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

No releases published

Packages

 
 
 

Contributors