25.1e NGC CLI: programmatic access to download models and resources¶

Welcome, new engineer! This section introduces you to the NGC CLI — a command-line tool that gives you programmatic access to download models, containers, and resources from the NVIDIA GPU Cloud (NGC) Catalog. Think of it as a direct pipeline from NVIDIA's software hub to your machine, without needing to click through a web interface.

🧭 Context: Why Use the NGC CLI?¶

The NGC Catalog hosts thousands of optimized AI models, containers, Helm charts, and datasets. While you can browse and download these via a web browser, the NGC CLI automates this process — essential for scripting, CI/CD pipelines, and reproducible workflows. Instead of manually clicking "Download," you can fetch exactly what you need with a single command, integrate it into your training scripts, or deploy it to production clusters.

⚙️ What is the NGC CLI?¶

The NGC CLI is a lightweight, cross-platform command-line tool that authenticates you to the NGC Catalog and enables:

• Downloading models (e.g., pre-trained LLMs, vision models)
• Pulling containers (e.g., TensorFlow, PyTorch with NVIDIA optimizations)
• Managing datasets and resources
• Automating workflows via scripts

It replaces manual web downloads with a repeatable, scriptable interface.

🛠️ Core Capabilities¶

• 🔐 Authentication: Log in once, then access private and public resources.
• 📦 Model Downloads: Fetch model files (e.g., .pth, .onnx, .safetensors) directly to your local machine.
• 🐳 Container Management: Pull Docker images from NGC's registry (e.g., nvcr.io/nvidia/tensorflow:latest).
• 📂 Resource Discovery: Search and list available models, containers, and datasets without opening a browser.
• 🔄 Automation: Integrate into shell scripts, Makefiles, or CI/CD pipelines (e.g., Jenkins, GitLab CI).

🧩 Comparison: Web UI vs. NGC CLI¶

📊 Visual Representation: NGC CLI Command Modules¶

This diagram outlines the subcommand groups of the NGC CLI used to download resources programmatically.

🕵️ How It Works (Conceptual Flow)¶

1. Install the NGC CLI on your system (Linux, macOS, or Windows).
2. Authenticate using an API key from your NGC account (stored locally).
3. Search for a resource (e.g., a model named llama-3.1-8b).
4. Download the resource to a specified directory.
5. Use the downloaded files in your AI workflow (training, inference, etc.).

No code blocks needed here — just a logical sequence of steps.

📥 Example Workflow (For Reference)¶

Below is a typical sequence of commands you would run in your terminal. Each line is a separate action.

# For reference only
ngc config set
ngc registry model list nvidia/llama-3.1-8b
ngc registry model download-version nvidia/llama-3.1-8b:latest --dest ./models

📤 Output: After the last command, you would see a confirmation message like:
Downloaded nvidia/llama-3.1-8b:latest to ./models/llama-3.1-8b

🧠 Key Takeaways for New Engineers¶

• The NGC CLI is your programmatic key to NVIDIA's software ecosystem — use it to avoid manual downloads.
• It supports models, containers, datasets, and more — all from the command line.
• Authentication is token-based, making it secure for automated pipelines.
• Always check the latest version of a resource using the list command before downloading.
• For production, store your API key as an environment variable (e.g., NGC_API_KEY) rather than typing it each time.

📚 Next Steps¶

• Explore the NGC Catalog website to see what resources are available.
• Practice listing models with the CLI before downloading.
• Integrate a model download into a simple shell script to automate your setup.

You now have the foundation to programmatically access NVIDIA's accelerated software hub — happy building! 🚀