Skip to content

Latest commit

 

History

History
33 lines (22 loc) · 1.17 KB

MODELS.md

File metadata and controls

33 lines (22 loc) · 1.17 KB

Model inference

Player-1 policies run on OpenAI-compatible vision endpoints. Official checkpoints are available on Hugging Face: Player-1-9B and Player-1-35B-A3B, or you can use your own SFT/RL export.


Setup

python -m pip install -r requirements/gym.txt
python -m playwright install chromium

With a local HF export and a vLLM version supporting Qwen3.5:

vllm serve /path/to/model --served-model-name Player-1-9B --port 8000 \
  --chat-template player1/agents/chat_template/qwen3_5_player1_full_history.jinja \
  --enable-auto-tool-choice --tool-call-parser qwen3_coder --reasoning-parser qwen3

Run

Set the model, endpoint, and input path at the top of each example, then run from the repository root:

PYTHONPATH=. python examples/inference/single_image.py
PYTHONPATH=. python examples/inference/play_game.py

The first requests an action for one screenshot; the second runs the policy in Gym. For authenticated endpoints, set ModelConfig.api_key_env to the environment variable holding the key.