Hey everyone! 👋 I recently got an Asus Ascent GX10 (DGX Spark) and quickly realized that going from unboxing to actually serving an LLM was a surprisingly manual process: system updates, GPU checks, Docker, NGC auth, HuggingFace, Tailscale, SSH keys, vLLM container config… I work from a MacBook Pro, so I also needed reliable remote access from day one. After going through all of that, I figured every DGX Spark owner would face the exact same steps, so I automated the entire thing into a single Bash script. What I'm most proud of is the auto-profiler: instead of figuring out the right vLLM flags for each model, spark reads the model's config.json and calculates everything automatically, tool-call support, context window, multimodal detection, MoE architecture, and memory utilization based on your actual VRAM. The whole thing is zero-dependency (just bash + curl), remote-first (Tailscale built in), and auto-updating. I'd love feedback from anyone running local LLMs, especially other DGX Spark owners. What's missing? What would make your workflow easier? PRs are very welcome.

spark cli

One bash script to set up and serve LLMs on DGX Spark

One bash script to set up and serve LLMs on DGX Spark