auto_ft

A Python CLI that wraps ostris/ai-toolkit to launch and manage LoRA fine-tuning jobs from a plain shell.

auto_ft does not train anything itself — it composes none of the YAML, runs none of the GPU code. It is a thin orchestration layer: spawn a detached Ostris subprocess, watch the on-disk artifacts, and surface JSON-shaped status the way other shell tools and scripts can consume.

What you get

A nine-subcommand CLI over an existing Ostris install:

Command	Purpose
auto_ft init	Create .autoft/state.json and register a job.
auto_ft prepare	Validate a dataset (count, format, resolution, PIL integrity) and write a hash manifest.
auto_ft train	Spawn detached Ostris training; returns in ~2s.
auto_ft status	Filesystem-derived job status (running / completed / stopped / stale / failed).
auto_ft logs	Tail N lines of train.log (platform-independent).
auto_ft stop	Terminate a running job (Windows: TerminateProcess).
auto_ft checkpoints	List *.safetensors checkpoints.
auto_ft samples	List sample image paths.
auto_ft export	Copy a checkpoint to a deployable path.

Every successful invocation emits a single JSON object on stdout. Errors emit {"error_code": "...", "message": "...", "details": {...}} and exit 1.

Install

pipx install auto-ft

auto_ft requires Python 3.11+ and a working ostris/ai-toolkit install on the same machine. The CLI launches Ostris as a subprocess; it does not vendor or install Ostris itself.

Configure Ostris

Tell auto_ft how to find your Ostris install. Two options:

Environment variables:

$env:AUTO_FT_OSTRIS_PYTHON = "C:\path\to\ai-toolkit\venv\Scripts\python.exe"
$env:AUTO_FT_OSTRIS_RUN_PY = "C:\path\to\ai-toolkit\run.py"

Config file at ~/.auto_ft/config.toml:

[ostris]
python = "C:\\path\\to\\ai-toolkit\\venv\\Scripts\\python.exe"
run_py = "C:\\path\\to\\ai-toolkit\\run.py"

Env vars take precedence. If both are absent, auto_ft raises E_OSTRIS_CONFIG_MISSING.

Quickstart

# 1. Register a job in the current project.
auto_ft init --name mydog --trigger zyx --dataset .\images

# 2. Validate the dataset (writes <images>\.auto_ft_prep.json).
auto_ft prepare .\images --trigger zyx

# 3. Hand auto_ft an Ostris-shaped YAML and launch.
auto_ft train .\config.yaml

# 4. Check progress (read-only; no side effects).
auto_ft status mydog
auto_ft logs mydog --tail 50

# 5. When the job is done, export the LoRA.
auto_ft checkpoints mydog
auto_ft export .\output\mydog\mydog.safetensors

auto_ft does not generate the Ostris YAML for you — bring your own, hand-authored or composed by any tool you prefer. The CLI's only YAML requirement is that config.process[0].training_folder is an absolute path; Ostris owns the <training_folder>/<config.name>/ join itself.

Trust boundary

auto_ft is a thin wrapper. When you run auto_ft train cfg.yaml, you are trusting:

• the Python CLI you installed (this package), and
• the Ostris installation pointed at by AUTO_FT_OSTRIS_PYTHON / AUTO_FT_OSTRIS_RUN_PY.

The CLI never imports ai-toolkit; it only spawns the Ostris Python interpreter as a subprocess. Resume semantics, model loading, GPU ownership — everything compute-bearing — happens inside Ostris.

Status

Alpha (0.1.x). Tested on Windows 10/11 with NVIDIA 8GB VRAM running SDXL LoRA training. The CLI itself is arch-agnostic — any model Ostris can train, auto_ft can launch — but only the SDXL path has been exercised end-to-end so far.

License

MIT — see LICENSE.