Quickstart

This guide walks you through starting a Tinker API server using the SkyRL backend, and running a training script against it.

If you run into any issues, please open an issue or message us on Slack.

Prerequisites

CUDA 12.8
uv package manager

1. Clone and Install SkyRL

git clone https://github.com/NovaSky-AI/SkyRL.git
cd SkyRL/skyrl
uv venv --python 3.12
uv sync --extra tinker --extra fsdp

2. Start the Tinker Server

The server hosts the Tinker API and manages a background engine that processes training requests.

uv run --extra tinker --extra fsdp -m skyrl.tinker.api \
    --base-model "Qwen/Qwen3-0.6B" \
    --backend fsdp

This will:

Start the FastAPI server on port 8000
Initialize the training workers
Load the base model and set up inference engines

The server is ready when you see Uvicorn running on http://0.0.0.0:8000.

Server Options

Flag	Default	Description
`--base-model`	(required)	HuggingFace model name or path
`--backend`	(required)	Backend type (`fsdp` or `megatron`)
`--port`	8000	API server port
`--checkpoints-base`	`/tmp/skyrl_checkpoints`	Directory for checkpoint storage
`--database-url`	`sqlite:///...tinker.db`	Database URL for request tracking

3. Run a Training Script

In a separate terminal, run the supervised learning loop from the tinker-cookbook:

git clone https://github.com/thinking-machines-lab/tinker-cookbook.git
cd tinker-cookbook

TINKER_API_KEY=tml-dummy uv run --with tinker --with datasets \
    python -m tinker_cookbook.recipes.sl_loop \
    base_url=http://localhost:8000 \
    model_name="Qwen/Qwen3-0.6B" \
    train_on_what=LAST_ASSISTANT_MESSAGE

You should see training metrics (loss, grad_norm) printed after each batch.

Congratulations! You've run your first Tinker training script on SkyRL.

4. More Recipes

In general, zero code changes are required to execute tinker-cookbook scripts on SkyRL. Simply update the base_url to point to your local Tinker server (e.g., http://localhost:8000 above).

See Cookbook Scripts for a few example commands and recipes that have been validated on SkyRL.