DashAISoftware/dashAI

dashAI – A graphical toolbox for AI model life‑cycle

What it is

• An open‑source desktop/web application that lets you train, evaluate and deploy modern AI models (especially large language models in GGUF format) through a visual interface.
• The UI runs in a browser but is launched from a local server, so everything stays on your machine.

Key capabilities

• Model handling – Supports PyTorch models and LLMs via llama‑cpp‑python (Llama, Mistral, Qwen, etc.)
• Experiment management – Stores datasets, runs, and results in a local SQLite DB (migrations handled by Alembic).
• GPU acceleration – Optional CUDA (12.8) or ROCm (6.4) builds; otherwise works CPU‑only.
• Cross‑platform – Windows 10/11, macOS 14+ (Apple Silicon only), and modern Linux x64.
• Packaging options – Ready‑to‑run installers (AppImage, .exe, .dmg), PyPI package, or Docker images (CPU or CUDA).

How to get it

1. Desktop installer – Download the appropriate file from the latest GitHub release and run it. No Python or dependencies needed; the installer bundles Python 3.12, CPU‑only PyTorch and llama‑cpp‑python.
2. PyPI – uv tool install dashai (recommended) or pip install dashai. After installation you can run dashai to start the server.
3. Docker – Build the supplied Dockerfile (CPU) or Dockerfile.cuda (CUDA) and run the container, exposing port 8000.

Typical workflow

# start the app (installed via pip/uv or from source)
$ dashai               # or: python -m DashAI
# open a browser to http://localhost:8000

• Use the UI to upload a dataset (see the companion DashAI_Datasets repo), configure training parameters, launch experiments, and monitor metrics.
• Results are saved under the directory you specify with --local‑path (default ~/.DashAI).

Installation nuances

• CPU‑only: the bundled installers already contain a compatible PyTorch wheel. For a slimmer pip install, reinstall PyTorch from the cpu index.
• GPU: after installing the base package, reinstall PyTorch from the CUDA or ROCm index and compile llama‑cpp‑python with the appropriate CMake flag (-DGGML_CUDA=on or -DGGML_HIP=on). You need a recent compiler, the CUDA Toolkit (≥12.8) or ROCm SDK, and matching drivers.
• Apple Silicon: installers work on ARM macOS; pip install works on both Intel and Apple Silicon, but the bundled binaries require AVX2 on x86.

Development

• Front‑end (React/Yarn) lives in DashAI/front. Run yarn install then yarn start for hot‑reloading.
• Back‑end (Python) uses uv for dependency management; uv sync sets up a virtualenv with dev extras (cpu, cuda). Run the server with uv run dashai or uv run python -m DashAI.
• Database schema migrations are managed by Alembic and run automatically on startup.
• Tests are written with pytest and can be executed from DashAI/back.

Who’s behind it

• Developed by students and researchers at the University of Chile, UTFSM, and partner institutions (CENIA, IMFD, Unholster) with funding from ANID.

When to choose dashAI

• You want a local, GUI‑driven environment for experimenting with LLMs without writing code.
• You need a lightweight, cross‑platform setup that can run on CPU or leverage a GPU when available.
• You prefer an all‑in‑one installer or a containerized deployment.

---

All information above is taken directly from the project's README.