This repository offers a curated collection of step-by-step playbooks for configuring and deploying AI/ML workloads on NVIDIA DGX Spark systems, specifically those equipped with Blackwell architecture. It targets engineers and researchers aiming to accelerate AI development and deployment on high-performance NVIDIA hardware.

How It Works

The project provides detailed, actionable playbooks covering installation, configuration, inference, and development environment setup for various AI frameworks and models. Each playbook includes prerequisites, step-by-step instructions, troubleshooting, and example code, designed to leverage the specialized capabilities of DGX Spark devices.

Quick Start & Requirements

• Primary Requirement: NVIDIA DGX Spark devices with Blackwell architecture.
• Setup: Follow individual playbook instructions for specific configurations.
• Resources:
  • Documentation: https://www.nvidia.com/en-us/products/workstations/dgx-spark/
  • Developer Forum: https://forums.developer.nvidia.com/c/accelerated-computing/dgx-spark-gb10

Highlighted Details

• Supports numerous AI frameworks and tools: NVIDIA Comfy UI, Dreambooth LoRA, LLaMA Factory, NeMo, Ollama, vLLM, TRT LLM, SGLang, Unsloth.
• Encompasses advanced AI tasks: multi-agent chatbots, multi-modal inference, RAG applications, speculative decoding, video search/summarization agents.
• Includes playbooks for optimizing inference, fine-tuning models, and setting up development environments (e.g., VS Code, AI Workbench).
• Covers system-level configurations: local network access, multi-Spark NCCL communication, Tailscale setup.

Maintenance & Community

Community support and discussion are available via the NVIDIA Developer Forum. Specific details on project maintainers, sponsorships, or roadmap are not provided in the README snippet.

Licensing & Compatibility

Licensing information is detailed in the repository's LICENSE and LICENSE-3rd-party files. Specific terms, restrictions, or compatibility notes for commercial use are not elaborated upon in the provided text.

Limitations & Caveats

The playbooks are strictly limited to NVIDIA DGX Spark hardware featuring the Blackwell architecture, posing a significant hardware dependency. Setup may require advanced knowledge of AI infrastructure and NVIDIA-specific technologies.