A self-hosted control panel that turns an NVIDIA RTX Spark (N1X) laptop — or any GB10-class Linux box such as a DGX Spark — into a 24/7 local-AI server. It replaces the monitoring and one-click AI tooling of NVIDIA's DGX OS, adds remote GPU servers, a full-screen app for Claude Code and Codex, and browser terminals — and everything sits behind one port (11000), so a single SSH tunnel or a Tailscale link is all you need.
Personal project, tested on one machine: an HP OmniBook X 14 (RTX Spark N1X, 128 GB) running Omarchy (Arch Linux, aarch64). Not affiliated with NVIDIA, HP, Anthropic, OpenAI, or any of the projects it integrates.
- Overview — live GPU / CPU / unified memory / storage / network charts, GPU processes (with an End button: your own processes get SIGTERM → SIGKILL, container processes stop the container, systemd services are stopped so they don't respawn), active models and containers, and a tile per GPU on every connected server. Quick actions: Open WebUI, JupyterLab, VS Code, opencode; desktop on/off (frees GPU memory), real screen-off for OLED panels (works on the console and under GNOME / Hyprland), restart, shut down.
- Models
- Ollama — installed models (load / unload / delete / chat, stop a stuck load), live search of ollama.com with per-tag sizes and one-click download.
- vLLM — search Hugging Face (formats vLLM can't serve are filtered out), resumable downloads, one-click deploy on NVIDIA's
nvcr.io/nvidia/vllmcontainer. Per-model defaults: the memory budget is worked out from the weights + a KV-cache budget + runtime overhead and checked against free memory; flags come from the model card's single-nodevllm servecommand (DGX / RTX Spark sections first) and are checked against the flags, parsers and architectures of the installed vLLM image; tool / reasoning parsers are read from the chat template; hybrid (Mamba / DeltaNet) models get a sane--max-num-seqs; the context length auto-fits the budget. - TensorFold — runs Qwen3.8-Flash-Next, which is too large for vLLM in 128 GB, with the MiaAI-Lab single-Spark TensorFold recipe.
- Model router — an OpenAI-compatible endpoint at
http://127.0.0.1:11020/v1that lists every running vLLM / TensorFold model, local or on a server, and routes requests by model name. Point Open WebUI (or any OpenAI client) at it once and new deployments simply show up in the model menu. - Servers — add GPU servers over SSH (a one-time password installs a dashboard-only key; the password is never stored). Live per-GPU status (available / mine / in use by others), end your own GPU processes, and deploy vLLM or Ollama on a server: it installs into a scratch folder you choose (no root, no Docker), downloads the model, starts it in the background and forwards the port back, so it appears in the model router.
- Agents — Claude Code and Codex on this machine and on every server: versions, login state, binary detection (including the copies bundled with VS Code extensions), and tmux-backed browser terminals.
- Spark Agents (
/agents/) — a full-screen chat app for Claude Code and Codex, driving the CLIs the same way the official IDE extensions do (Claude'sstream-jsoncontrol protocol, Codex'sapp-serverJSON-RPC), locally or over SSH: streaming replies with Markdown, collapsible thinking, a tool-call timeline with live run times, inline diffs, permission / plan / question cards, mode, model and reasoning-effort pickers, a plan / changes / usage side panel, searchable history with resume and delete, automatic reconnection after network drops or restarts, and a force take-over for Codex conversations held open by another window. - Terminal — several browser terminals at once (tabs, split in two or four), on this machine or any server. They live in tmux, so closing the browser keeps them running; reopening reconnects.
- Downloads that survive anything — Hugging Face downloads use a small resumable downloader (HTTP Range, SHA-256 check, standard HF cache layout). Every download runs as its own systemd unit and resumes after network drops, dashboard restarts, or reboots.
- Docker — containers (start / stop / logs / open published ports), images (pull with progress), volumes, and a "run container" form with a GPU switch.
- Network — Tailscale on/off and login, MagicDNS, Tailscale SSH, exit nodes, a device list with direct-vs-relay status and ping, a connectivity check, local interfaces, and HTTPS sharing of the dashboard, Open WebUI and opencode on your tailnet (
tailscale serve). - Developer tools — JupyterLab with CUDA PyTorch, VS Code for the Web (the official
code serve-web), and opencode with your running models (local and remote) configured as its backend automatically. - English / 中文 UI, dark and light themes.
![]() |
![]() |
![]() |
![]() |
![]() |
![]() |
![]() |
![]() |
![]() |
![]() |
browser ──► :11000 Caddy (127.0.0.1 only)
├─ / dashboard backend :11010 (python3, stdlib only)
├─ /agents/ Spark Agents :11005 (python3, stdlib only)
├─ /term/ browser terminals :11004 (ttyd + tmux)
├─ /jupyter/ JupyterLab :11001
├─ /code/ VS Code for the Web :11002
├─ webui.localhost Open WebUI :8080
├─ opencode.localhost opencode :4096
└─ p<port>.localhost any local port (e.g. a container on :8188)
model router 127.0.0.1:11020/v1 ──► local vLLM / TensorFold, and servers via ssh -L
servers ssh (dashboard key, ControlMaster) ──► GPU probe, remote deploys, Claude Code / Codex, terminals
*.localhost names resolve to your own machine in Chrome, Edge and Firefox, so forwarding port 11000 alone (ssh -L 11000:localhost:11000 host) reaches every service. Over Tailscale, the Network tab can publish the dashboard at https://<machine>.<tailnet>.ts.net.
Requirements: Linux with an NVIDIA GPU and driver, python3, docker with nvidia-container-toolkit, and caddy. Optional: ollama, tailscale, opencode, ttyd + tmux (Terminal tab), claude / codex (Agents).
git clone https://github.com/zyyangzy/spark-dashboard.git
cd spark-dashboard
./scripts/install.shThe installer copies the dashboard to ~/.local/share/spark, installs the systemd units (it uses sudo for the system-level ones), generates access tokens in ~/.config/spark, and starts the entry point on http://localhost:11000, Spark Agents, and — if ttyd is installed — the browser terminals. On a laptop it also installs the lid/screen service: closing the lid turns the screen off without suspending.
Optional pieces
- JupyterLab:
python3 -m venv ~/.local/share/spark/jupyter-venv, then installjupyterlaband a CUDA build of PyTorch into it, thensystemctl --user enable --now spark-jupyter. - VS Code for the Web: place the official VS Code CLI (arm64) at
~/.local/share/spark/vscode-cli, thensystemctl --user enable --now spark-code. - opencode:
systemctl --user enable --now spark-opencode. Models are written to~/.config/spark/opencode.jsonand merged with your own opencode config. - Power menu (desktop on/off, restart a stuck Ollama, restart, shut down from the dashboard):
./scripts/install-power-control.sh. It allows exactly the fixed actions indeploy/bin/spark-ctlthrough a sudoers rule. - Servers: add them in the Servers tab. Remote deploys need nothing on the server except Python 3 and an NVIDIA driver.
- HP OmniBook X 14 only:
./extras/hp-omnibook-x14/install.sh— see below. extras/ubuntu-style-gnome.sh: a GNOME desktop with Ubuntu's look, for those who prefer it over Hyprland.
- Every service listens on
127.0.0.1only. Reach them through an SSH tunnel or your tailnet. - The dashboard has no login of its own, and it can start containers, open terminals and shut the machine down. Its POST endpoints require a custom header and only accept localhost /
*.localhost/*.ts.nethosts, which blocks cross-site requests and DNS rebinding; the terminal WebSocket rejects other origins. Even so, only share it on a tailnet you fully control. - JupyterLab, VS Code and opencode are protected by tokens or passwords stored in
~/.config/spark(mode 600), and the dashboard passes them along when you open each service. - Server passwords are used once to install the dashboard's SSH key and are never written to disk.
- GPU throttling on the HP OmniBook X 14. Under sustained load the HP firmware lowers the SoC power limit (PL1) from ~80 W to 45 W at 0.1 W/s, and the GPU follows it from ~1,800 to ~1,100 MHz (NVIDIA/open-gpu-kernel-modules#1521). On this machine a bf16 GEMM fell from 79 to 49 TFLOP/s in 5 minutes. Re-sending HP's performance thermal profile restores PL1, but the firmware's own
\_SB.WMID.STPSaborts on Linux because\_SB.FFA0.RBUFdeclares a field past the end of its buffer.extras/hp-omnibook-x14replays the same EC mailbox writes from a small SSDT (with a fixed copy ofRBUF) loaded throughacpi_configfs, andspark-pl1ddoes that whenever PL1 drops below 76 W, pausing at SoC ≥ 95 °C. With it, the GEMM holds 77–78 TFLOP/s and the SoC settles around 88 °C. The dashboard shows it as the GPU full power switch. Tested on BIOS F.05 only; it reads/dev/memand loads ACPI tables as root, so read the two files first. - Qwen3.8-Flash-Next FP4-Local on TensorFold, single stream: prefill 2,040 tok/s at 3.3K, 2,400 at 13K, 2,150 at 52K; decode 63–66 tok/s English (thinking on), 85–88 code, 34–36 Chinese. Over 7 minutes of back-to-back 26K prefills it holds ~2,230 tok/s with GPU full power on, and drops to ~1,580 tok/s (−36 %) with it off. Decode is memory-bound and barely changes.
- MiaAI-Lab/Qwen3.8-Flash-Next-Single-DGX-Spark-TensorFold (MIT) — the TensorFold image, engine settings and serve arguments used by the TensorFold tab. See THIRD_PARTY_NOTICES.md.
- The HP throttling analysis by the reporters of NVIDIA/open-gpu-kernel-modules#1521.
- TensorFold, vLLM / NVIDIA NGC containers, Ollama, Open WebUI, Caddy, ttyd, tmux, Tailscale, opencode, Claude Code, Codex, Omarchy.
Spark Dashboard 是一个自托管的控制面板,把 NVIDIA RTX Spark(N1X)笔记本(或 DGX Spark 这类 GB10 机器)变成 7×24 小时运行的本地 AI 服务器,替代 DGX OS 的监控和一键部署功能。所有服务只走 11000 一个端口,一条 SSH 隧道或 Tailscale 就能访问。
- 概览:GPU / CPU / 内存 / 存储 / 网络实时监控,GPU 进程一键结束,远程服务器的每块 GPU 状态;快捷开关桌面、OLED 真黑屏(桌面开着也有效)、重启、关机。
- 模型:Ollama、vLLM、TensorFold。vLLM 部署时按模型自动给出默认参数:按权重 + KV 缓存 + 开销算好内存预算,从模型说明里挑出适合单机的启动参数并对照当前 vLLM 版本检查,从对话模板判断工具调用解析器,混合架构模型自动限制并发。下载断点续传。
- 模型路由:
127.0.0.1:11020/v1汇总本机和服务器上所有在跑的模型,Open WebUI 配一次就能自动看到新部署的模型。 - 服务器:通过 SSH 连接 GPU 服务器(密码只用一次,不保存),查看每块 GPU 的占用,一键在服务器上部署 vLLM / Ollama(不需要 root 和 Docker),端口自动转发回本机。
- 编程助手 / Spark Agents:本机和服务器上的 Claude Code、Codex。Spark Agents 是全屏对话界面:流式回复、工具调用时间线、文件 diff、权限 / 计划 / 提问卡片、模型与推理强度切换、历史对话恢复、断线自动重连,Codex 对话被占用时可以强制接管。
- 终端:同时开多个网页终端(标签页、二分屏、四宫格),本机或服务器都可以;终端在 tmux 里运行,关掉网页也不会断。
- HP OmniBook X 14 专用:修复固件导致的 GPU 持续降频(
extras/hp-omnibook-x14),面板上的「GPU 满功率」开关。 - Docker、网络(Tailscale 全套管理)、JupyterLab、网页版 VS Code、opencode;中英文界面,深色 / 浅色主题。
安装方法见上方 Install。只在作者自己的机器(HP OmniBook X 14 + Omarchy)上测试过。










