DeepSeek Demo In Jupyter Notebook

This repository provides a Jupyter Notebook environment for using the DeepSeek-AI model (and any other models available on Hugging Face) to generate code responses based on user prompts. It includes a GPU-accelerated setup and a CPU-only branch for local execution.

For a detailed walkthrough of this repository and its features, check out my video on my YouTube channel.

Features

• Uses the DeepSeek-Coder-1.3B-Instruct model
• Supports both GPU-accelerated and CPU-only execution
• Provides a Gradio-based UI for interactive code generation
• Includes VS Code DevContainer support for easy reproducibility

Setup Instructions

1. Clone the Repository

git clone https://github.com/Perpetue237/deepseek_demo.git
cd deepseek-demo

2. Setup Environment

Ensure you have a Hugging Face API token. Create a .devcontainer/devcontainer.env file and add:

HF_TOKEN=<your_huggingface_token>

3. Choose a Branch

GPU-Accelerated Setup (Default)

The default branch uses PyTorch with CUDA for GPU acceleration. This setup is defined in the following DevContainer:

{
  "name": "DeepSeek Notebook",
  "image": "pytorch/pytorch:2.6.0-cuda12.6-cudnn9-devel",
  "postCreateCommand": "pip install jupyter ipykernel transformers accelerate sentencepiece gradio ipywidgets",
  "customizations": {
    "vscode": {
      "extensions": [
        "ms-python.python",
        "ms-toolsai.jupyter",
        "RSIP-Vision.nvidia-smi-plus"
      ]
    }
  },
  "runArgs": ["--env-file=.devcontainer/devcontainer.env", "--gpus", "all"],
  "mounts": [
    "source=${localWorkspaceFolder},target=/workspace,type=bind"
  ]
}

To use this version, simply checkout the main branch and open the folder in VS Code.

git checkout main

CPU-Only Setup (Local Execution)

If you don’t have a GPU, switch to the CPU-only branch, which uses a lighter DevContainer:

git checkout cpu-only

The CPU-only DevContainer is defined as follows:

{
    "name": "DeepSeek local Demo",
    "image": "python:3.11-bullseye",
    "postCreateCommand": "pip install torch jupyter ipykernel transformers accelerate sentencepiece gradio ipywidgets",
    "runArgs": [
        "--env-file=.devcontainer/devcontainer.env"
    ],
    "customizations": {
        "vscode": {
            "extensions": [
                "RSIP-Vision.nvidia-smi-plus",
                "ms-python.python",
                "ms-toolsai.jupyter"
            ]
        }
    }
}

4. Start the Notebook

After setting up the environment, open the DevContainer in VS Code and start the Jupyter Notebook:

jupyter notebook

Running the Model

The deepseek_notebook.ipynb file contains the necessary code to load the DeepSeek model and generate responses. Here’s a summary:

1. Load the model: The script automatically downloads the model (if not already saved) or loads it from disk.
2. Generate responses: Uses transformers to generate AI-powered code solutions.
3. Gradio UI: Provides an interactive chat interface for code generation.

Run the notebook cells to generate AI-assisted Python code snippets!

Commit Guidelines

Since models are large, add the following to .gitignore:

models/
.gradio/
.env

Before pushing, ensure you don’t accidentally commit your Hugging Face token:

git add .
git commit -m "DeepSeek Demo with Gradio"
git push origin main

Conclusion

This repository provides an environment for experimenting with DeepSeek-AI and any other models available on Hugging Face. You can use GPU acceleration for faster performance or switch to the CPU-only branch if you lack a compatible GPU.

Enjoy coding with DeepSeek-AI! 🚀