English | 简体中文

State-Tuning Studio

A graphical tool for RWKV State Tuning. Configure models, prepare data, launch training, monitor progress, export weights, and test inference locally—all from one interface.

This repository includes two clients with broadly the same feature set. Choose the one that best fits your platform and workflow:

Client	Location	Tech Stack	Best For
Flutter client	Project root (lib/, etc.)	Flutter / Dart	Desktop (Windows / Linux)
PySide desktop client	pyside_desktop/	Python / PySide6	Native desktop apps for Windows / Linux

The bundled training scripts are located in each client's statetuning_repo/ directory and support RWKV7 State Tuning in bf16, fp16, and fp32 precision.

Preview

Features

• Models: Select pretrained .pth weights and a tokenizer
• Data: Manage the training repository and JSONL datasets
• Training: Configure hyperparameters and launch State Tuning
• Monitoring: View loss curves and training logs
• Export: Export trained state weights
• Settings: Detect or install the Python environment with one click, including PyTorch and other dependencies
• Testing: Load a model and run conversational inference tests

The interface supports English / Simplified Chinese / Traditional Chinese.

Project Structure

statetuning/
├── lib/                    # Flutter application source
├── assets/                 # Flutter assets, including statetuning_repo.zip
├── android/ ios/ windows/ linux/   # Flutter platform projects
├── pyside_desktop/         # PySide desktop client
│   ├── main.py             # Desktop application entry point
│   ├── main_window.py      # Main window UI
│   ├── controller.py       # Application logic
│   ├── locale/             # Localized strings
│   └── statetuning_repo/   # Bundled training repository used by PySide
├── previewimg/             # Interface screenshots
└── README.md

---

Run the Flutter Client

Requirements

• Flutter SDK (Dart SDK ^3.8.1)
• Build tools for your target platform, such as Android Studio, Xcode, or Visual Studio
• The rwkv_mobile_flutter plugin for model loading and inference on mobile devices

Prepare Dependencies

By default, pubspec.yaml references rwkv_mobile_flutter through a local path:

rwkv_mobile_flutter:
  path: ../../rwkv_mobile_flutter

Clone that repository next to statetuning, or update the path / git source in pubspec.yaml.

Example directory layout:

Documents/
├── rwkv_mobile_flutter/
└── statetuning/statetuning/   # Root of this Flutter project

Launch

Run the following commands from the project root containing pubspec.yaml:

flutter pub get
flutter run

To target a specific device:

# Windows desktop
flutter run -d windows

# Linux desktop
flutter run -d linux

# Android device or emulator
flutter run -d android

On first launch, the app extracts and initializes the bundled training repository. Use the Settings page to detect or install the Python training environment.

---

Run the PySide Desktop Client

Requirements

• Python 3.10+ (3.11 or 3.12 recommended)
• Optional: an NVIDIA GPU with CUDA for training and GPU inference
• On Windows, compiling CUDA operators may require Visual Studio Build Tools; the app can detect this requirement and guide you through installation

Launch

Option 1: Launch from the project root (recommended)

cd /path/to/statetuning
python -m pyside_desktop.main

Option 2: Launch from the pyside_desktop directory

cd pyside_desktop
python main.py
# or
python -m main

If PySide6 is not installed, the launcher automatically creates a virtual environment at pyside_desktop/.venv and installs the dependencies without modifying the system Python installation. It retries the default package index and common mirrors automatically. You can also set PIP_INDEX_URL to use a custom mirror.

First-Time Setup

1. Open Settings, then select Detect Environment or One-Click Install to create python_venv and install PyTorch and other dependencies.
2. Alternatively, select Choose Existing Environment and provide a Python virtual environment directory. If the directory is empty, the dependencies are installed there automatically.
3. Once the environment is ready, follow the workflow through the Models, Data, and Training pages.

The PySide client's Testing page loads models and generates text through a separate Python subprocess (_pyside_rwkv_test_worker.py), so it does not require the Flutter plugin.

---

Training Repository

Each client includes a statetuning_repo/ directory containing the RWKV State Tuning scripts. See the following files for detailed parameter documentation:

• Flutter client after extraction: assets/statetuning_repo/README.md
• PySide client: pyside_desktop/statetuning_repo/README.md

You can also run training directly from the command line after configuring the Python environment and the parameters in train.py:

cd pyside_desktop/statetuning_repo
python train.py

Choosing a Client

	Flutter Client	PySide Desktop Client
Mobile	✅ Android / iOS	❌
Desktop	✅ Windows / Linux, etc.	✅ Windows / Linux
Model testing backend	rwkv_mobile_flutter plugin	Python + PyTorch subprocess
Environment setup	Guided setup in the app	One-click setup in the app with uv / pip support
Dependencies	Flutter SDK + rwkv_mobile_flutter	Python 3 + PySide6 (automatic installation available)

License

See the LICENSE file in this repository.