View source Markdown

Install Guide

This document explains how to install and prepare StableSteering on Windows for local development or evaluation.

Published HTML documentation:

• GitHub Pages Docs

Recommended reading before installation:

• Student Tutorial
• Quick Start
• User Guide

Requirements

• Windows with PowerShell
• Python 3.11 or newer
• Node.js and npm
• Git

For the real generation backend you also need:

• a CUDA-capable GPU
• compatible PyTorch/CUDA support

Fastest Setup

From the repository root run:

powershell -ExecutionPolicy Bypass -File scripts/bootstrap.ps1

This will:

• create a local virtual environment in .venv
• install Python dependencies
• install npm dependencies
• optionally prepare Hugging Face model assets if requested

Manual Setup

1. Create and activate a virtual environment

python -m venv .venv
.venv\Scripts\Activate.ps1

2. Install Python dependencies

Development dependencies:

python -m pip install --upgrade pip
python -m pip install -e .[dev]

Real inference dependencies:

python -m pip install -e .[dev,inference]

3. Install browser test dependencies

npm install

Prepare Model Assets

If you want to run the real Diffusers backend:

python scripts/setup_huggingface.py

The prepared model snapshot is stored under models/.

Run the App

python scripts/run_dev.py

Open:

http://127.0.0.1:8000

Useful runtime pages:

• http://127.0.0.1:8000/
• http://127.0.0.1:8000/setup
• http://127.0.0.1:8000/diagnostics/view
• http://127.0.0.1:8000/sessions/{session_id}/trace-report

Run the Tests

Backend tests:

python -m pytest

Default browser suite:

npm run test:e2e:chrome

Headed browser debug run:

npm run test:e2e:debug

Opt-in real-backend browser smoke:

$env:STABLE_STEERING_E2E_REAL="true"
npm run test:e2e:real

Common Notes

• The normal app runtime is GPU-only and uses the real Diffusers backend.
• The mock generator is reserved for tests and explicit test harnesses only.
• The normal user flow starts from the user's text prompt on /setup.
• Round generation and feedback submission run as async jobs with visible progress in the UI.
• Trace logs are written under data/traces/.
• Per-session trace bundles and readable report.html files are written under data/traces/sessions/<session_id>/.
• You can generate a complete real GPU example bundle with python scripts/create_real_e2e_example.py.
• The docs site is generated locally with python scripts/build_pages_site.py.