Fine Tuning as a Service (FTAAS)

register → orchestrate → train (HF / TRL / LLaMA-Factory / Unsloth / Axolotl · LoRA / QLoRA / DoRA) → MLflow → vLLM / Ray Serve → API / Console

ftaas.Client / notebook           Console
              \                    /
               ▼                  ▼
              FTAAS gateway  (:8080)
         ┌─────────┼──────────┐
         ▼         ▼          ▼
      registry   control    deploy
                   │
                workflow
                   │
            runner · Ray · MLflow · vLLM

Components

Quick start

python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
./scripts/start_all.sh          # → http://127.0.0.1:8080
./scripts/e2e_smoke.sh
./scripts/stop_all.sh

SDK

from ftaas import Client, Framework, Technique, HyperParameters

with Client() as c:
    ds = c.register_dataset("examples/data/alpaca_sample.jsonl", name="alpaca")
    job = c.create_finetune_job(
        model_name="sshleifer/tiny-gpt2",
        dataset=ds,
        framework=Framework.TRANSFORMERS,
        technique=Technique.LORA,
        parameters=HyperParameters(max_steps=10),
    )
    job = c.wait_for_job(job.job_id)
    model = c.get_model(job.registered_model_name)
    ep = c.create_endpoint(model.model_name, inference_framework="vllm")
    print(c.prompt(ep.endpoint_id, "What is LoRA?").completion)

Layout

FTAAS/
├── packages/ftaas_sdk/ftaas/   # Client SDK
├── services/
│   ├── ftaas_app/              # gateway
│   ├── registry/
│   ├── control/
│   ├── workflow/
│   └── deploy/
├── runner/                     # local + Airflow DAG
├── training/
├── ui/console/
├── examples/
└── scripts/

License

Apache-2.0