checkpoint:
  local:
    keep_last_n: 1
  save_best_model: false
  save_final_model: true
  save_pt_file: false
  save_safetensors_file: true
  save_training_state: 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
    per_device_batch_size: 8
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: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct
    enable_gradient_checkpointing: true
output_dir: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/flappy_fix_latency_3_200ep_7k2steps_plain_multi
rl_games:
  env_eval:
    enabled: true
    eval_backend: latency_bench
    prompt_mode: raw
run_id: flappy_fix_latency_3_200ep_7k2steps_plain_multi
run_root_dir: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints
seed: 42
trainer:
  distributed_backend: deepspeed
  freeze_llm_layers: []
  freeze_modules: ''
  freeze_tied_embedding: false
  freeze_vit: false
  gradient_accumulation_steps: 16
  is_resume: true
  learning_rate:
    action_model: 0.0001
    base: 2.0e-05
    qwen_vl_interface: 1.0e-05
  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
  pretrained_checkpoint: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/flappy_fix_latency_3_200ep_7k2steps_plain_multi/checkpoints/steps_4000_state
  scheduler_specific_kwargs:
    min_lr: 1.0e-06