add docker feature to the project

[wild-deer]

1. High-Level Summary (TL;DR)

• Impact Level: [Medium] - Introduces a standardized Docker container runtime environment for the project, simplifying environment setup (especially CUDA and PyTorch dependencies), with no effect on existing local runtime logic.
• Key Changes:
  • ✨ Added Dockerfile: Based on PyTorch and CUDA 13.0 images, automatically configures Alibaba Cloud APT mirror and pre-installs the flashinfer library along with system dependencies.
  • ✨ Added docker-compose.yml: Provides one-click startup configuration, supporting full GPU mounting, 32GB shared memory, and persistent model caching.
  • ✨ Updated README.md: Added comprehensive Docker usage instructions, including build/run commands and guidance for adjusting CUDA versions to match local machines.
  • 🔧 Added .dockerignore: Excludes .git, .venv, caches, and build artifacts to optimize image build speed.

2. Visual Overview (Code & Logic Map)

graph TD
    %% Style definition (high contrast)
    classDef host fill:#bbdefb,color:#0d47a1,stroke:#0d47a1
    classDef docker fill:#c8e6c9,color:#1a5e20,stroke:#1a5e20
    classDef volume fill:#fff3e0,color:#e65100,stroke:#e65100

    subgraph "Host Machine"
        User["Developer"]:::host
        Source["Local code directory (./)"]:::host
    end

    subgraph "Docker Environment"
        Compose["docker-compose.yml"]:::docker
        Container["lingbot-map container (CUDA 13.0)"]:::docker
        
        subgraph "Volumes"
            HFCache["hf_cache"]:::volume
            TorchCache["torch_cache"]:::volume
        end
    end

    User -->|"docker compose run"| Compose
    Compose -->|"Build & Start"| Container
    Source -.->|"Mount workspace (/workspace)"| Container
    Container -.->|"Persistent read/write"| HFCache
    Container -.->|"Persistent read/write"| TorchCache

Loading

3. Detailed Change Analysis

🐳 Docker Containerization Configuration

Description: Added full Docker support configuration, making it easy to replicate complex GPU/CUDA environments across different machines (source files: Dockerfile, docker-compose.yml, .dockerignore).

Base Environment & Dependencies (Dockerfile):

Item	Details	Notes
Base Image	spxiong/pytorch:2.11.0-py3.10.19-cuda13.0.2-ubuntu22.04	Pre-installed PyTorch 2.11 and CUDA 13.0.2
APT Mirror	https://mirrors.aliyun.com/ubuntu	Uses Alibaba Cloud mirror by default for faster builds in China (overridable via build arg UBUNTU_MIRROR)
System Dependencies	ffmpeg, libgl1, libsm6, etc.	Provides low-level libraries for vision processing (e.g., OpenCV)
Python Dependencies	flashinfer-python, flashinfer-jit-cache	Installs FlashInfer wheels for cu130 and the main project package .[vis]

Environment & Resource Mounting (docker-compose.yml):

Key	Value / Target Path	Notes
Volume: Code	.:/workspace	Mounts local code directory as container workspace, supporting hot-reload
Volume: HF Cache	hf_cache:/root/.cache/huggingface	Persists HuggingFace models to avoid repeated downloads
Volume: Torch Cache	torch_cache:/root/.cache/torch	Persists PyTorch cache data
Ports	8080:8080	Maps internal service port to host
Resource: gpus	all	Exposes all available host NVIDIA GPUs to the container
Resource: shm_size	32gb	Increases shared memory to prevent out-of-memory issues in multi-process data loading (e.g., DataLoader)

📝 Documentation Update

Description: Updated user guide to support the new Docker workflow (source file: README.md).

• Added a new "Docker / Docker Compose" section.
• Provided clear startup commands (docker compose build & docker compose run --rm --service-ports lingbot-map).
• Added an example for running the demo script demo.py inside the container.
• High-value guidance: Included concrete Diff examples for modifying the Dockerfile to support other CUDA versions (e.g., downgrading from CUDA 13.0 to the more common CUDA 12.8).

4. Impact & Risk Assessment

• ⚠️ Breaking Changes: None. All changes are new files and documentation updates, with no impact on existing local development/runtime workflows using conda/pip.
• 🧪 Testing Recommendations:
  • Build Test: Run docker compose build on a machine with NVIDIA drivers and Docker installed, verifying successful Alibaba Cloud mirror switching and dependency installation (network variations may cause APT/PIP download failures).
  • GPU Passthrough Test: After starting the container, run nvidia-smi and python -c "import torch; print(torch.cuda.is_available()) in the terminal to confirm GPU detection.
  • Business Logic Test: Run demo.py inside the container to verify flashinfer and model inference work as expected without errors.