hydra:
  searchpath:
    - pkg://verl/trainer/config

defaults:
  - ppo_trainer
  - _self_

agentlightning:
  port: 9999
  trace_aggregator:
    level: transition  # transition or trajectory, docs refer to https://agent-lightning.github.io/posts/trajectory_level_aggregation/
    trajectory_max_prompt_length: 2048  # supported in trajectory level aggregation, suggest to set as maximum length for the prompt in first turn
    trajectory_max_response_length: 8192  # supported in trajectory level aggregation, suggest to set as maximum length for the cumulative agent responses in the full trajectory, i.e., n_turns * (max_response_length + max_prompt_length)
    debug: False  # supported in trajectory level aggregation, enable to diagnose trace merging failures
    mismatch_log_dir: ./mismatch_cases  # supported in trajectory level aggregation with debug=True, directory to store logs of mismatch cases

data:
  filter_overlong_prompts: false

actor_rollout_ref:
  rollout:
    mode: async
    agent:
      custom_async_server:
        path: pkg://agentlightning.verl.async_server
        name: PatchedvLLMServer