Strata Supervisor

A web UI for Strata, the local LLM server for large mixture-of-experts models (Qwen, DeepSeek) on consumer GPUs. It is built FROM the official Strata image: one container, no Docker socket, no Strata code changed. It drops into Unraid or any other Docker host with an NVIDIA or AMD GPU.

Strata's own Docker image loads one model, fixed by environment variables, at container start. This project keeps the container running without a model and lets you manage several model configs and load or unload them by hand in the browser.

Unofficial private project, not affiliated with, endorsed by, or supported by Niko1221 or the Strata project.

Features

• Several model configs, one TOML file each, editable in the UI or on the host. Several configs of the same model are fine (e.g. different contexts).
• Load and unload only by hand. Nothing loads at container start, nothing loads or swaps by itself. Switching models needs no container restart.
• Strata's catalog in the form: models grouped like in Strata's desktop app, with download size, ON DISK and a fit marker for your cards and RAM. Or pick a folder of your own GGUF files.
• Multi-GPU through roles (main, helper, off), passed as --gpu-roles. Card numbers come from nvidia-smi, never from the code.
• RAM mode per model: all experts in RAM, only what no card holds, a RAM budget, or straight from the model files. The budget respects the container's RAM limit.
• Plan and Check download ask Strata what a load would need, without changing anything.
• Models page: Strata's downloads and your own GGUF folders with size and the configs using them; delete with confirmation.
• Logs of the launch, the engine and the supervisor, with the API key hidden.
• While no model runs, the API port answers 503 with "No model loaded in Strata." and GET /v1/models lists nothing.

Requirements

• A graphics card Strata supports, with the matching image (see "Image tags" below):
  • NVIDIA: driver 580 or newer (575 or newer for cuda12); on Unraid the Nvidia-Driver plugin, elsewhere the NVIDIA Container Toolkit (--gpus all).
  • AMD RX 7000, RX 9000 or Strix Halo: the amdgpu kernel driver (built into Unraid), the devices /dev/kfd and /dev/dri in the container, and --group-add video.
• Disk space: 60 to 190 GB per model, on a fast disk.
• RAM: depends on the model and RAM mode; see Strata's docs/MODELS.md.
• amd64 only, like Strata's images.

Quick start

Unraid

Install strata-supervisor from Community Applications, or add unraid/strata-supervisor.xml as a template. Then follow the user guide.

Docker

docker run -d --name strata-supervisor --gpus all \
  -p 8081:8081 -p 8080:8080 \
  --ulimit memlock=-1:-1 --shm-size 2g \
  -e ALLOWED_HOSTS=192.168.1.10,myserver \
  -v /srv/strata/data:/data \
  -v /srv/strata/config:/config \
  ghcr.io/sognix/strata-supervisor:latest

AMD (here an RX 7000):

docker run -d --name strata-supervisor \
  --device /dev/kfd --device /dev/dri --group-add video \
  -p 8081:8081 -p 8080:8080 \
  --ulimit memlock=-1:-1 --shm-size 2g \
  -e ALLOWED_HOSTS=192.168.1.10,myserver \
  -v /srv/strata/data:/data \
  -v /srv/strata/config:/config \
  ghcr.io/sognix/strata-supervisor:latest-rocm-gfx110x

Open http://<server>:8081. Nothing is loaded at start.

Image tags

ghcr.io/sognix/strata-supervisor, one image per Strata flavor:

Each release also has fixed tags such as 1.0.0 / 1.0.0-rocm-gfx110x and 1.0 / 1.0-rocm-gfx110x. dev and dev-<flavor> follow the main branch and are for testing.

Configuration

docker stop must finish within Unraid's "Docker Stop Timeout" (10 s by default). The supervisor stops a loaded model within 8 s.

Model configs

One flat TOML file per config in /config/models/<name>.toml. These files decide: every load passes --save-preset <name>, never --preset, so nothing old from Strata's presets comes back.

model = "Q2_0"
family = "qwen"
context = "64k"
kv = "int8"
vision = "off"
gpu_roles = "0=main,1=helper"
ram_mode = "budget"          # normal | resident-experts | budget | mmap-experts
ram_budget_gib = 28
vram_reserve_mib = 512       # optional, first card in Strata's order (the helper when split)
extra_set = ["prefill=65536"]

Changes to the loaded model apply at its next load.

Security

The web UI has no login. Anyone who can reach port 8081 can load and unload models, delete model files and read the API key. Run it only on a trusted network and don't expose it to the internet. Set ALLOWED_HOSTS to the names you open the UI with: the UI then refuses other host names, which stops web pages on other sites from reaching it through DNS rebinding. Changes are accepted only as JSON POST requests whose Origin matches the host.

The inference API on port 8080 is Strata's own and needs the API key.

To report a vulnerability, see SECURITY.md.

Known limitations

• One image per Strata flavor is published: cuda13, cuda12, rocm-gfx120x, rocm-gfx110x and rocm-gfx1151. Tested on real cards: only cuda13 (RTX 3090 + RTX 5060 Ti) and rocm-gfx110x (one RX 7900 XTX). Intel Arc (sycl) is not built.
• The UI depends on some internals of Strata (catalog, log lines, folder layout). A new Strata version may need an update of this project; the Strata version is pinned in the Dockerfile (STRATA_VERSION).
• Only one model is loaded at a time. Strata's "one model per card" mode is not supported.
• Tuning (--tune) is not offered.

Documentation

• docs/USER-GUIDE.md - installing, getting models, settings, troubleshooting
• CONTRIBUTING.md - building, testing and running it yourself
• AI/IMPLEMENTATION.md - background, decisions and test results (written as working notes for AI coding agents, but readable for humans too)
• CHANGELOG.md - what changed in each release

Support

This is a private hobby project, maintained in spare time. Bug reports with the details the issue template asks for are welcome and get looked at; feature requests too, but without any promise. For problems with Strata itself (model quality, speed, engine crashes), please check the Strata repository first.

License

Copyright (c) 2026 sognix, MIT License, see LICENSE.

The image is built on the official Strata image (MIT) and contains everything in it; see NOTICE.md for Strata's license and the third-party components.