vLLM CLI

A command-line interface tool for serving Large Language Models using vLLM. Provides both interactive and command-line modes with features for configuration profiles, model management, and server monitoring.

Interactive terminal interface with GPU status and system overview
Tip: You can customize the GPU stats bar in settings

• 🎯 Interactive Mode - Rich terminal interface with menu-driven navigation
• ⚡ Command-Line Mode - Direct CLI commands for automation and scripting
• 🤖 Model Management - Automatic discovery of local models with HuggingFace and Ollama support
• 🔧 Configuration Profiles - Pre-configured and custom server profiles for different use cases
• 📊 Server Monitoring - Real-time monitoring of active vLLM servers
• 🖥️ System Information - GPU, memory, and CUDA compatibility checking
• 📝 Advanced Configuration - Full control over vLLM parameters with validation

Quick Links: 📖 Docs | 🚀 Quick Start | 📸 Screenshots | 📘 Usage Guide | ❓ Troubleshooting | 🗺️ Roadmap

The Multi-Model Proxy is a new experimental feature that enables serving multiple LLMs through a single unified API endpoint. This feature is currently under active development and available for testing.

What It Does:

• Single Endpoint - All your models accessible through one API
• Live Management - Add or remove models without stopping the service
• Dynamic GPU Management - Efficient GPU resource distribution through vLLM's sleep/wake functionality
• Interactive Setup - User-friendly wizard guides you through configuration

Note: This is an experimental feature under active development. Your feedback helps us improve! Please share your experience through GitHub Issues.

For complete documentation, see the 🌐 Multi-Model Proxy Guide.

New built-in profiles specifically optimized for serving GPT-OSS models on different GPU architectures:

• gpt_oss_ampere - Optimized for NVIDIA A100 GPUs
• gpt_oss_hopper - Optimized for NVIDIA H100/H200 GPUs
• gpt_oss_blackwell - Optimized for NVIDIA Blackwell GPUs

Based on official vLLM GPT recipes for maximum performance.

Save and quickly launch your favorite model + profile combinations:

vllm-cli serve --shortcut my-gpt-server

• Automatic discovery of Ollama models
• GGUF format support (experimental)
• System and user directory scanning

• Environment Variables - Universal and profile-specific environment variable management
• GPU Selection - Choose specific GPUs for model serving (--device 0,1)
• Enhanced System Info - vLLM feature detection with attention backend availability

See CHANGELOG.md for detailed release notes.

⚠️ Binary Compatibility Warning: vLLM contains pre-compiled CUDA kernels that must match your PyTorch version exactly. Installing mismatched versions will cause errors.

vLLM-CLI will not install vLLM or Pytorch by default.

# Install vLLM -- Skip this step if you have vllm installed in your environment
uv venv --python 3.12 --seed
source .venv/bin/activate
uv pip install vllm --torch-backend=auto
# Or specify a backend: uv pip install vllm --torch-backend=cu128

# Install vLLM CLI
uv pip install --upgrade vllm-cli
uv run vllm-cli

# If you are using conda:
# Activate the environment you have vllm installed in
pip install vllm-cli
vllm-cli

# Install vLLM CLI + vLLM
pip install vllm-cli[vllm]
vllm-cli

git clone https://github.com/Chen-zexi/vllm-cli.git
cd vllm-cli
pip install -e .

⚠️ Compatibility Note: pipx creates isolated environments which may have compatibility issues with vLLM's CUDA dependencies. Consider using uv or conda (see above) for better PyTorch/CUDA compatibility.

# If you do not want to use virtual environment and want to install vLLM along with vLLM CLI
pipx install "vllm-cli[vllm]"

# If you want to install pre-release version
pipx install --pip-args="--pre" "vllm-cli[vllm]"

• Python 3.9+
• CUDA-compatible GPU (recommended)
• vLLM package installed
• For dependency issues, see Troubleshooting Guide

# Interactive mode - menu-driven interface
vllm-cli
# Serve a model
vllm-cli serve --model openai/gpt-oss-20b

# Use a shortcut
vllm-cli serve --shortcut my-model

For detailed usage instructions, see the 📘 Usage Guide and 🌐 Multi-Model Proxy Guide.

vLLM CLI includes 7 optimized profiles for different use cases:

General Purpose:

• standard - Minimal configuration with smart defaults
• high_throughput - Maximum performance configuration
• low_memory - Memory-constrained environments
• moe_optimized - Optimized for Mixture of Experts models

Hardware-Specific (GPT-OSS):

• gpt_oss_ampere - NVIDIA A100 GPUs
• gpt_oss_hopper - NVIDIA H100/H200 GPUs
• gpt_oss_blackwell - NVIDIA Blackwell GPUs

See 📋 Profiles Guide for detailed information.

• Main Config: ~/.config/vllm-cli/config.yaml
• User Profiles: ~/.config/vllm-cli/user_profiles.json
• Shortcuts: ~/.config/vllm-cli/shortcuts.json

• 📘 Usage Guide - Complete usage instructions
• 🌐 Multi-Model Proxy - Serve multiple models simultaneously
• 📋 Profiles Guide - Built-in profiles details
• ❓ Troubleshooting - Common issues and solutions
• 📸 Screenshots - Visual feature overview
• 🔍 Model Discovery - Model management guide
• 🦙 Ollama Integration - Using Ollama models
• ⚙️ Custom Models - Serving custom models
• 🗺️ Roadmap - Future development plans

vLLM CLI uses hf-model-tool for model discovery:

• Comprehensive model scanning
• Ollama model support
• Shared configuration

src/vllm_cli/
├── cli/           # CLI command handling
├── config/        # Configuration management
├── models/        # Model management
├── server/        # Server lifecycle
├── ui/            # Terminal interface
└── schemas/       # JSON schemas

Contributions are welcome! Please feel free to open an issue or submit a pull request.

MIT License - see LICENSE file for details.