SGFleet - Lightweight Web UI & Orchestrator for SGLang

[joesew]

Hey everyone,

I built a small self-hosted tool called SGFleet to manage SGLang on a dedicated server running a single RTX 5000 Blackwell.

SGLang’s inference speed is great, but manually swapping containers via CLI to manage VRAM on one GPU was getting tedious—especially alongside key management and setting up configs for different coding tools.

SGFleet adds a lightweight Web UI, API gateway, and container manager on top of SGLang (Docker Compose + FastAPI + SQLite).

What it does:

• Model Orchestration: Start/stop model containers from the UI to quickly swap models on a single GPU.
• HF Model Management: Pull Hugging Face weights and launch instances directly from the dashboard.
• API Gateway & Quotas: Create token-restricted keys with per-user metric tracking and dedicated login views.
• Client Config Generators: Auto-generates ready-to-copy configs for Continue.dev, Aider, Roo Code/Cline, Opencode, Cursor, and Claude Code

It bootstraps automatically with an init.sh script that checks the host for NVIDIA runtime requirements and generates environment keys.

Repo: https://github.com/sewingsolutions/sgfleet

Note: It's still an early work-in-progress, so expect some rough edges and bugs! If you give it a spin, bug reports, feedback, and PRs are super welcome.