checkpoint:
  local:
    keep_last_n: 1
  save_best_model: false
  save_final_model: true
  save_pt_file: false
  save_safetensors_file: true
  sync:
    enabled: false
    keep_last_n: 0
    repo_id: null
datasets:
  vla_data:
    data_mix: flappy_train__bridge
    dataset_py: lerobot_datasets
    eval_data_mix: flappy_train__bridge__val
    latency_curriculum:
      enabled: false
    obs_image_size: null
    per_device_batch_size: 16
framework:
  action_model:
    action_dim: 7
    action_env_dim: 2
    action_hidden_dim: 2560
    action_horizon: 1
    action_model_type: MLP
    loss_type: discrete_ce
  kv_memory:
    enabled: false
    packed_train: false
    rebased_sink: true
    rollout_len: 8
    window: 4
  name: QwenOFT
  qwenvl:
    attn_implementation: flash_attention_2
    base_vlm: /workspace/latency-sensitive-bench/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct
    enable_gradient_checkpointing: true
output_dir: /workspace/latency-sensitive-bench/starVLA/results/Checkpoints/flappy_fix_latency_2_200ep_7k2steps_plain_multi
rl_games:
  env_eval:
    distributed_mode: rank_sharded
    enabled: true
    eval_backend: latency_bench
    eval_parallel_envs: 5
    fixed_episode_seeds: true
    image_size: 224
    image_transform: raw_rgb
    latency:
      prompt_map_path: /workspace/latency-sensitive-bench/starVLA/data/flappy_fix_latency_2_200ep_7k2steps/flappy_train__bridge/latency_prompt_map.json
    mid_train:
      enabled: true
      interval_steps: 250
      latencies:
      - 2
      max_steps_per_episode: 3600
      num_episodes: 20
    prompt_mode: raw
    seed: 42
  model_alias: openvla
  task: flappy
run_id: flappy_fix_latency_2_200ep_7k2steps_plain_multi
run_root_dir: /workspace/latency-sensitive-bench/starVLA/results/Checkpoints
seed: 42
trainer:
  distributed_backend: deepspeed
  eval_action_classification: false
  eval_action_classification_interval: null
  eval_interval: 250
  eval_num_batches: 50
  freeze_llm_layers: []
  freeze_modules: ''
  freeze_tied_embedding: false
  freeze_vit: false
  gradient_accumulation_steps: 4
  is_resume: false
  learning_rate:
    action_model: 0.0001
    base: 2.0e-05
    qwen_vl_interface: 1.0e-05
  logging_frequency: 1
  lr_scheduler_type: cosine_with_min_lr
  max_train_steps: 4000
  num_warmup_steps: 100
  optimizer:
    betas:
    - 0.9
    - 0.95
    eps: 1.0e-08
    fused: true
    weight_decay: 1.0e-08
  per_latency_eval_num_batches: null
  pretrained_checkpoint: null
  profile_timing:
    enabled: false
  reload_modules: null
  save_interval: 500
  scheduler_specific_kwargs:
    min_lr: 1.0e-06
wandb_entity: saberrr-zju
wandb_project: starVLA_rl_games