A professional-grade deployment guide for setting up private NVIDIA Blackwell-based LLM inference servers optimized for Openclaw Skills.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install dgx-spark-setup
2. Manual Installation
Copy the skill folder to one of these locations
~/.openclaw/skills/ <project>/skills/ Priority: Workspace > Local > Bundled
3. Prompt Installation
Copy this prompt to OpenClaw to install it automatically.
Help me install dgx-spark-setup using Clawhub. If Clawhub is not installed, install it first (npm i -g clawhub).
Prefer to download?
Get the raw skill files in a ZIP archive.
What is DGX Spark Setup?
The DGX Spark Setup skill provides a rigorous framework for transforming an NVIDIA DGX Spark (GB10 Blackwell) into a high-performance local LLM inference server. It specializes in handling the unique hardware requirements of the 128GB unified memory architecture, specifically targeting sm_121 GPU compatibility which is often unsupported by standard packages. This Openclaw Skills technical guide ensures developers can run massive models like Nemotron Super 120B using an optimized stack of vLLM and LiteLLM.
By leveraging this skill, teams can establish a secure, multi-user inference environment. The workflow integrates LiteLLM for virtual key management and Tailscale for encrypted remote access, bridging the gap between local high-performance hardware and remote Openclaw Skills workers or MacBook clients without exposing data to the public internet.
DGX Spark Setup Use Cases
- Deploying Nemotron Super 120B (NVFP4) on local hardware for maximum privacy and performance.
- Configuring multi-user LLM access with per-user rate limiting and authentication via LiteLLM.
- Establishing secure remote connections between local DGX servers and remote MacBook clients using Tailscale.
- Troubleshooting complex Blackwell/sm_121 GPU driver and vLLM source-build conflicts.
How DGX Spark Setup Works
- The environment is initialized using uv to manage a Python 3.12 virtual environment tailored for the DGX Spark.
- Custom builds of PyTorch (CUDA 13.0) and Triton (sm_121 support) are compiled to ensure compatibility with Blackwell architecture.
- vLLM is installed from source with specific flashinfer and flashinfer-python version matching to prevent cubin failures.
- The inference server is launched within a persistent tmux session, utilizing NVFP4 quantization for memory-efficient model loading.
- LiteLLM acts as a proxy layer, managing systemd services and providing an OpenAI-compatible API for Openclaw Skills integration.
- Tailscale creates a secure mesh network, allowing the Openclaw host to communicate with the DGX Spark over a private IP.
DGX Spark Setup Setup
Initial Environment
mkdir -p ~/vllm-install && cd ~/vllm-install
uv venv .vllm --python 3.12
source .vllm/bin/activate
Specialized Package Installation
# Install PyTorch with Blackwell support
uv pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu130
# Build Triton from source
git clone https://github.com/triton-lang/triton.git
cd triton && git checkout 4caa0328bf8df64896dd5f6fb9df41b0eb2e750a
pip install -e python/
vLLM Launch
TORCH_CUDA_ARCH_LIST=12.1a VLLM_USE_FLASHINFER_MXFP4_MOE=1 python -m vllm.entrypoints.openai.api_server \
--model /home/jhernandez/models/nemotron-super-120b-nvfp4 --port 8000
DGX Spark Setup Data Schema & Taxonomy
The skill organizes its infrastructure and model storage through a structured hierarchy on the host system:
| Path | Purpose |
|---|---|
~/models/ |
Storage for safetensor shards and model weights (e.g., Nemotron 120B). |
~/vllm-install/ |
Source code and virtual environments for the inference engine. |
~/.config/systemd/user/ |
Service definitions for automated LiteLLM proxy management. |
~/litellm-config.yaml |
Metadata taxonomy for virtual keys, rate limits, and model routing. |
~/litellm.log |
Unified logging for tracking request flow and server health. |
DGX Spark Setup Advanced Features
- Multi-user virtual key routing via LiteLLM for fine-grained access control within Openclaw Skills ecosystems.
- Systemd service integration for automatic recovery and persistence of the inference proxy across reboots.
- Support for NVFP4 quantization, enabling the execution of 120B+ parameter models on unified memory architectures.
- Encrypted networking via Tailscale, allowing for decentralized compute nodes to serve a central Openclaw instance.
SKILL.md
Loading
Related Openclaw Skills
Legal Structural Analysis
A specialized legal intelligence skill that deconstructs complex statutes into a structured seven-dimensional framework for lawyers and law students.
Pumble
Pumble for Openclaw Skills enables AI agents to manage team workspaces, channels, and messages through a secure, automated communication layer.
daily-finance
A professional financial intelligence skill that aggregates top-tier investment news and market trends from multiple authoritative platforms.
ClawTeam
A powerful CLI orchestration tool for deploying and managing self-organizing AI agent swarms with built-in task dependencies and git isolation.
Woodpecker.co Integration for AI Agents
A powerful sales engagement skill that enables AI agents to automate personalized B2B email outreach and manage campaigns through Woodpecker.co.
Extensiv Integration Manager
A comprehensive integration skill for automating data exchange between eCommerce platforms, warehouse management systems, and accounting software.
Featured*
