Ascend/ascend-model-migration

Ascend Model Migration Suite

Complete model migration and training suite for Ascend NPU environment. This suite provides end-to-end workflow from SSH connection to model training.

When to Invoke

• User wants to migrate open-source models to Ascend NPU
• User needs to setup training environment on Ascend NPU
• User wants to train models (BEVFormer, etc.) on Ascend NPU
• User asks about model deployment on Ascend platform

Suite Structure

ascend-model-migration/
├── SKILL.md              # Main skill (this file)
├── ssh-connection/       # SSH connection skill (from ssh-dev-suite)
│   └── SKILL.md
├── model-migration/      # Model migration skill
│   └── SKILL.md
└── model-training/       # Model training skill
    └── SKILL.md

# External Dependencies (in AgentSkills/skills/)
└── ascend-mmlab-install-suite/  # MMLab environment setup
    └── SKILL.md

Workflow Overview

┌─────────────────┐     ┌─────────────────┐     ┌─────────────────┐     ┌─────────────────┐
│  ssh-connection │ ──► │ mmlab-install   │ ──► │ model-migration │ ──► │ model-training  │
│                 │     │    -suite       │     │                 │     │                 │
│  Connect to     │     │  Install mmcv,  │     │  Clone repo,    │     │  Execute        │
│  Ascend server  │     │  mmdet, mmdet3d │     │  apply patches  │     │  training script│
└─────────────────┘     └─────────────────┘     └─────────────────┘     └─────────────────┘

Complete Workflow

Phase 0: DrivingSDK Preparation

DrivingSDK仓库地址：https://gitcode.com/Ascend/DrivingSDK

在开始任何工作前，确认DrivingSDK已克隆到工作目录：

cd <working_directory>
git clone https://gitcode.com/Ascend/DrivingSDK

重要：DrivingSDK包含所有必要的patch文件和训练脚本，无需向用户询问patch位置。

Phase 1: Environment Information Collection

Ask user for the following information:

1. Server IP address (e.g., 192.168.13.151)
2. SSH username (e.g., root)
3. Docker container name (if applicable, e.g., qianqian_0312)
4. Conda environment name (e.g., qianqian_2)
5. Working directory path (e.g., /home/zqq/0312_trae_new)

Use ssh-connection skill (this suite) to establish connection.

Phase 2: Environment Setup

Use AgentSkills/skills/ascend-mmlab-install-suite/ to install dependencies:

1. Verify base environment (torch, torch_npu, CANN)
2. Install MMLab dependencies:
   • mmcv-full (with NPU operators)
   • mmdetection
   • mmsegmentation
   • mmdetection3d
   • detectron2

Phase 3: Model Migration

Use model-migration skill (this suite) to prepare model code:

1. Clone open-source repository
2. Apply NPU adaptation patches
3. Setup dataset and weight links

Phase 4: Model Training

Use model-training skill (this suite) to launch training:

1. Select training mode (performance/accuracy)
2. Execute training script
3. Monitor training progress

Supported Models

| Model | Status | DrivingSDK Path | |-------|--------|-----------------| | BEVFormer | ✅ Supported | model_examples/BEVFormer | | More models | 🔄 Coming soon | - |

Quick Start Example

User: "I want to train BEVFormer on Ascend NPU"

Agent workflow:
1. [ssh-connection] Ask for server info → Connect
2. [ascend-mmlab-install-suite] Check environment → Install MMLab deps
3. [model-migration] Clone BEVFormer → Apply patch → Link data
4. [model-training] Launch training → Monitor progress

Key Principles

NPU Compatibility Check

After installation, verify NPU compatibility:

python -c "import site; print([p for p in site.getsitepackages() if 'cuda' in p.lower()])"

Should return empty list [] for NPU environment.

Network Strategy

| Situation | Strategy | |-----------|----------| | Server can access GitHub | Direct git clone on server | | Server network timeout | Configure proxy or clone locally → scp to server | | Proxy available | Configure proxy: export http_proxy=http://proxy:port |

重要：所有GitHub仓库克隆都应使用GitHub官方链接，不使用镜像站点。如果网络超时，优先配置代理。

Error Handling

Each sub-skill handles its own errors and provides recovery suggestions. Common issues:

1. SSH connection failed: Check IP, username, password/key
2. NPU device not found: Check environment variables
3. CUDA paths conflict: Reinstall with NPU support
4. Patch application fails: Check code version

Environment Variables Reference

Common Ascend NPU environment variables:

export ASCEND_SLOG_PRINT_TO_STDOUT=0
export ASCEND_GLOBAL_LOG_LEVEL=3
export TASK_QUEUE_ENABLE=2
export COMBINED_ENABLE=1
export HCCL_WHITELIST_DISABLE=1
export HCCL_CONNECT_TIMEOUT=1200

Note: ASCEND_RT_VISIBLE_DEVICES may be required in specific environments.

Reference Files

• SSH connection: ssh-connection/SKILL.md (this suite)
• MMLab install suite: AgentSkills/skills/ascend-mmlab-install-suite/SKILL.md
• DrivingSDK path: DrivingSDK/model_examples/
• Model migration: model-migration/SKILL.md (this suite)
• Model training: model-training/SKILL.md (this suite)