Operator UI

A web console for the uaii CLI, running locally or self-hosted from dashboard/. Chat with Hugging Face decoder folders and GGUF blk.* models, pull from the Hub, and manage device and memory settings.

What it provides

Chat

Generation through uaii generate or a warm uaii chat --jsonl worker, with streaming output, system prompt, token limit, temperature, top-p, top-k, repetition penalty, cancellation, and a FIFO request queue.

Hugging Face

Hub pulls and folder imports for decoder-only CausalLM checkpoints, which require a current uaii build. See Hugging Face.

GGUF

Dense and MoE blk.* transformers including the MLA expand path, plus a built-in demo model for smoke tests.

Model library

Upload GGUF files, list Hugging Face and GGUF models, convert, and delete.

Runtime health

Run uaii doctor --load-plugins from the browser.

Benchmarks

Run uaii_bench and read the results as a table or raw JSON.

Settings

Bind address, port, binary paths, device selection defaulting to auto, GEMM provider, thread count, and auth token.

API

A REST surface under /api/* plus OpenAI-compatible /v1/models and /v1/chat/completions.

Themes

Light and dark, following your system preference by default.

Running locally

Build the runtime first, then start the dashboard on loopback:

# 1) Build UAII
cmake --build build --target uaii uaii_bench --parallel

# Windows App Control frequently blocks unsigned .exe files. Build under WSL:
#   bash scripts/build_uaii_wsl.sh
#   bash scripts/test_generate_wsl.sh

# 2) Dashboard
cd dashboard
npm run install:all
npm run build && npm start
# → http://127.0.0.1:8787

# Development, with hot reload
npm run dev   # UI :5174 · API :8787

Helper scripts: dashboard/scripts/start-local.ps1 on Windows and dashboard/scripts/start-local.sh on Linux and macOS.

Self-hosting on a LAN or server

export UAII_DASH_TOKEN=$(openssl rand -hex 16)
export UAII_DASH_BIND=0.0.0.0
cd dashboard
./scripts/start-host.sh
# Open http://SERVER:8787 and unlock with the token

Docker:

export UAII_DASH_TOKEN=$(openssl rand -hex 16)
export UAII_HOST_BIN=$PWD/../build/libs/uaii-cli/uaii
export UAII_HOST_BENCH=$PWD/../build/benchmarks/uaii_bench
cd dashboard
docker compose up --build

Configuration

Copy dashboard/uaii-dash.example.json to uaii-dash.json, or set environment variables, which take precedence over the file. The default device is "backend": "auto"; an older config pinned to "cpu" keeps that value until you reset it in Settings. Bind address and port changes take effect after restarting the process.

Common workflows

1. Verify the installOpen Chat, select the uaii-tiny-demo model, and run. Tokens stream immediately with no download.
2. Run a Hugging Face modelOpen Models, choose Pull from Hub or Import HF folder, then switch to Chat, select the folder, and run with paging and device selection applied.
3. Run a GGUF modelOpen Models, import a .gguf file, then switch to Chat, select it, set a token limit, and run.
4. Check runtime healthOpen Runtime and run doctor.
5. Measure performanceOpen Benchmarks and run the microbenchmark suite.
6. Serve a teamSelf-host, then issue a Bearer token for /v1/chat/completions.

OpenAI-compatible clients

curl -s http://SERVER:8787/v1/models \
  -H "Authorization: Bearer $UAII_DASH_TOKEN"

curl -s http://SERVER:8787/v1/chat/completions \
  -H "Authorization: Bearer $UAII_DASH_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"model":"uaii-tiny-demo","messages":[{"role":"user","content":"hi"}]}'

# Reference a local file as "uaii-file-my-model.gguf"

Related

• dashboard/README.md and docs/prd-dashboard.md in the repository
• CLI — the generate and chat commands the console drives
• Features — GGUF architecture coverage
• Configuration — runtime and dashboard environment variables