pipelm 0.1.2

A lightweight, CLI modular tool for running Large Language Models (LLMs) & SLMs from Hugging Face.

PipeLM provides an intuitive CLI interface for interactive chat and a robust FastAPI server to integrate LLMs seamlessly into your applications.

---

Overview

PipeLM simplifies interaction with AI models, allowing you to:

• 📥 Download and manage models from Hugging Face.
• 🌐 Serve models through a standardized REST API.
• 💬 Test prompts via an interactive chat interface & a client
• 📜 Maintain conversation history.
• 🔄 Easily switch models with single command.

---

Features

• Interactive CLI Chat: Chat directly from your terminal.
• FastAPI Server: REST APIs with health monitoring.
• Efficient Model Management: Download and manage models easily.
• Support for different models: text2text and image2text models are supported
• Easy Authentication: Easily access HF models using only HF_TOKEN
• Client-Server Architecture: Model deployed on server can be accessed by client
• Docker Support: Containerize your models for better isolation.
• GPU Acceleration: Automatically utilize available GPUs.
• Model Quantization: Reduce memory usage (4-bit and 8-bit).
• Conversation History: Persistent chat context.
• Rich Terminal Interface: Enhanced CLI with markdown rendering.
• Robust Error Handling: Good handling of issues.

---

Installation

💻 From Source (Recommended)

Step 1: Create a Python Virtual Environment

# Clone the source repository
git clone https://github.com/kashyaprparmar/PipeLM
cd PipeLM

# Create virtual environment with Python's venv
python -m venv .venv

# Activate the environment
# On Linux/macOS:
source .venv/bin/activate

# On Windows:
.venv\Scripts\activate

Step 2: Install uv Within the Virtual Environment

# Install uv package manager
pip install uv

Step 3: Install dependencies using uv

# Install the package with uv (recommended)
uv pip install -e .

# Install just the dependencies from requirements.txt
uv pip install -r requirements.txt

📦 From PyPI (in development)

pip install pipelm

🐳 With Docker

git clone https://github.com/kashyaprparmar/PipeLM
cd PipeLM

docker build -f docker/Dockerfile -t pipelm .

docker run -p 8080:8080 -v pipelm_data:/root/.pipelm -e HF_TOKEN=your_token -e MODEL_NAME=HuggingFaceTB/SmolLM2-1.7B-Instruct pipelm

---

Usage

Login with HF_TOKEN

pipelm login
# Enter your HF_TOKEN to terminal

Download a Huggingface Model

# Download a model with particular allowed patterns
pipelm download HuggingFaceTB/SmolLM2-1.7B-Instruct --include ['*.json','*.safetensors']

List all Downloaded Models

pipelm list

Interactive Chat

# (Streaming is ENABLED BY DEFAULT)
# Start chatting with model 
pipelm chat HuggingFaceTB/SmolLM2-1.7B-Instruct

# Chatting with local model ( by default loads to port 8080)
pipelm chat /path/to/local/model

# Chatting using quantization
pipelm chat HuggingFaceTB/SmolLM2-1.7B-Instruct --quantize 4bit

# Chatting with model (Streaming DISABLED)
pipelm chat HuggingFaceTB/SmolLM2-1.7B-Instruct --no-stream

🚀 Start API Server

pipelm server HuggingFaceTB/SmolLM2-1.7B-Instruct --port 8080

# Using local model
pipelm server /path/to/local/model --port 8080

# With quantization
pipelm server HuggingFaceTB/SmolLM2-1.7B-Instruct --quantize 8bit

Making Client Requests to model on Server

# Making client request on a text2text model
pipelm client HuggingFaceTB/SmolLM2-1.7B-Instruct 
  --prompt "Summarize the benefits of solar energy." 

# Making client request on a text2text model with some model config
pipelm client HuggingFaceTB/SmolLM2-1.7B-Instruct 
  --prompt "Summarize the benefits of solar energy." 
  --max-tokens 80   
  --top-p 0.95   

# Making client request on a image2text model with some model config
pipelm client HuggingFaceTB/SmolVLM-500M-Instruct   
  --port 8080   
  --model-type image2text   
  --image "/home/kashyap/Desktop/mywork/PipeLM/assets/logo.png"   
  --prompt "Describe what you see."   
  --max-tokens 80   
  --top-p 0.95   
  --no-stream

🐳 Docker Compose

export HF_TOKEN=your_token
docker-compose up -d pipelm

---

API Endpoints

Quick Commands

Check Server Health:

curl http://localhost:8080/health

Script for Testing a Sample Prompt

# Checks the server health and runs a sample prompt in the server
python client.py

Send a Sample Prompt (cURL):

curl -X POST http://localhost:8080/generate \
  -H "Content-Type: application/json" \
  -d '{
    "messages": [{"role": "user", "content": "Explain the difference between AI and machine learning."}],
    "max_tokens": 200,
    "temperature": 0.7,
    "top_p": 0.9
  }'

GET /health

Health status of server and model.

{
  "status": "healthy",
  "model": "HuggingFaceTB/SmolLM2-1.7B-Instruct",
  "uptime": 42.5
}

GET /docs

Swagger UI for API documentation.

POST /generate

Generate text from conversation history.

Request:

{
  "messages": [
    {"role": "user", "content": "What is artificial intelligence?"}
  ],
  "max_tokens": 1024,
  "temperature": 0.7,
  "top_p": 0.9
}

Response:

{
  "generated_text": "Artificial intelligence (AI) refers to the simulation of human intelligence in machines..."
}

---

Chat Commands

• /exit or /quit – Exit chat
• /clear – Clear conversation history
• /info – Display current model information

---

Environment Variables

• HF_TOKEN: Your Hugging Face token (required).
• MODEL_DIR: Local model directory.
• PORT: Server port (default: 8080).

---

Project Structure

PipeLM/
├── pipelm/                 # Main package
│   ├── __init__.py
│   ├── cli.py
│   ├── server.py
│   ├── client.py
│   ├── downloader.py
│   ├── chat.py
│   └── utils.py
├── docker/                 # Docker setup
│   ├── Dockerfile
│   └── docker-compose.yml
├── setup.py
├── README.md
└── requirements.txt

---

Requirements

• Python 3.8+
• Torch (GPU support recommended)
• 16+ GB RAM (model-dependent)
• CUDA-compatible GPU (recommended)

---

Troubleshooting

Model Download Issues

If you encounter issues downloading models:

1. Check your Hugging Face token:
   • Create or verify your token at https://huggingface.co/settings/tokens (MAKE SURE TO GENERATE A READ/WRITE HF ACCESS TOKEN)
   • Set it in your environment as export HF_TOKEN=your_token_here
   • Or store it in .env file as HF_TOKEN=your_token_here
2. Network issues:
   • Check your internet connection
   • Verify you have permissions to download the model

Server Startup Issues

If the server fails to start:

1. Check if another process is using port 8080:
   • Use a different port: pipelm server HuggingFaceTB/SmolLM2-1.7B-Instruct --port 8081
2. Verify Python dependencies:
   • Ensure all required packages are installed: pip install -r requirements.txt

Memory Issues

If you encounter memory errors:

1. Choose a smaller model
2. Try quantization: pipelm chat HuggingFaceTB/SmolLM2-1.7B-Instruct --quantize 4bit
3. Ensure you have enough RAM and GPU VRAM if using CUDA

Model Storage

Models are downloaded to:

• Linux/Mac: ~/.pipelm/models/[sanitized_model_name] OR .local/share/pipelm/models OR /home/kashyap/snap/code/190/.local/share/pipelm/models
• Windows: C:\Users\[username]\AppData\Local\pipelm\pipelm\models\[sanitized_model_name]

When using Docker, models are stored in /root/.pipelm/models/ in the container, typically mapped to a persistent volume.

---

Contributing

Contributions are welcome! Submit a Pull Request.

---

License

MIT License. See LICENSE for details.