SparkD: The missing dashboard for spark-vllm-docker

Accelerated Computing DGX Spark / GB10 User Forum DGX Spark / GB10
1

Fellow GB10 owners — built this for myself after burning a weekend on SSH sessions, hand-edited YAMLs, and tail -f across tmux panes just to get one vLLM serve running. Open-sourced so the next person doesn’t lose the same weekend I did.

sparkd is a localhost dashboard that wraps eugr/spark-vllm-docker and turns the “I have a Spark, now what?” workflow into a few clicks. The headline idea: let AI do the part that actually makes the Spark feel intimidating — reading model cards, sizing recipes for your hardware, and splitting work across nodes correctly on the first try.

🤖 What the AI co-pilot actually does for you

  • Reads the Hugging Face model card automatically — architecture, parameter count, context length, dtype. You don’t paste anything.

  • Sees your hardware — per-node GPU model, VRAM, network interface, plus the total cluster topology if you’ve grouped boxes.

  • Writes a tuned vLLM recipe — correct --tensor-parallel-size, --pipeline-parallel-size, --gpu-memory-utilization, --max-model-len, --distributed-executor-backend=ray, and NCCL_SOCKET_IFNAME pinned to your IB iface.

  • Optimizes existing recipes for throughput, latency, or whatever goal you give it — with a side-by-side diff and rationale for each change.

  • Suggests fixes when launches fail — hand it the error log, get a proposed mods/ patch.

  • Bring your own provider — Anthropic, OpenAI, Gemini, Mistral, Groq, OpenRouter, Together, or a local vLLM endpoint. Keys live in your OS keyring.

🛠️ What the dashboard does around it

  • One-click launches on a single box or a multi-Spark cluster (tag boxes with the same cluster=<name> and they group automatically).

  • Multi-node “just works” — sparkd hands off to upstream launch-cluster.sh, which scps the launch script to workers, starts Ray, and runs vLLM across the cluster. No new orchestration code — just no more keyboard-yoga.

  • Smart recipe filtering — pick a target, the recipe dropdown narrows itself to ones whose tp×pp matches your node count. Toggle the chip row to override.

  • Live log streaming, pause / inspect / restart / stop without leaving the dashboard.

  • Recipe versioning with side-by-side diff and one-click revert. Every save is a new version.

  • Full Hugging Face browser baked in — filter by task / library, sort by downloads / likes / recent, drop your HF token for gated repos.

  • Mods support — sync from upstream or author your own.

  • No cloud. No telemetry. No account. Runs on your laptop, talks to your Sparks over SSH, stores everything in ~/.sparkd/.

If it saved me a weekend, I figured it might save someone else one too. MIT licensed. Feedback, issues, PRs all welcome:

👉 https://github.com/mchenetz/sparkd

2

This looks very cool! Can’t wait to give it a try after work..

3

This looks very useful, I will give it a try. I have been working on something similar, making it easy to run recipe’s, create recipe’s, and fix/optimize using claude or codex. I also included a section for chat and a canvas section, I can prompt the model to create a game and show it on canvas right there in the UI. I had the same idea make it user friendly, I can also run the llama benchy and get the results right on my app. It’s far from perfect but hoping to get it there.

4

Let me know how you like it. I am looking to make it truly useful for the community.

5

You solution looks interesting too. I will have to take a look at it. I love how the community is coming together to make things better for all. I have a few plans for the future.

  1. Integrate with monitoring tools on the Spark Boxes in order to create a full optimization loop that can kind of live optimize.
  2. integrate with some of the kernel tuning tools
  3. create a more dynamic dashboard of training cycles and changes

Let me know if you have any feedback.

Mike