model:
  model_path: /path/to/pretain_ckpt/hf_ckpt
  tokenizer_path: /path/to/Qwen3-VL-4B-Instruct
  post_training: true
  adanorm_time: true
  config_key: LingbotVLAV2Config
  moe_implementation: fused

data:
  datasets_type: vla
  data_name: multi
  train_path: assets/training_data/robotwin.txt
  robot_config_root: ./configs/robot_configs
  joints:
    - arm.position: 14
    - end.position: 14
    - effector.position: 2
  cameras:
    - camera_top
    - camera_wrist_left
    - camera_wrist_right
  norm_type:
    - arm.position: bounds_99_woclip
    - end.position: bounds_99_woclip
    - effector.position: bounds_99_woclip
  num_workers: 8
  use_future_image: true

train:
  output_dir: /path/to/save_ckpt
  moe_monitor_interval: 1000
  enable_gradient_checkpointing: false # Set to true if GPU memory is insufficient, but this will slow down training.
  precompute_grid_thw: true
  vlm_causal: true
  vlm_fsdp: true
  attention_implementation: flex_cached
  use_moe: true
  token_moe_layers: [0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35]
  token_num_experts: 32
  token_top_k: 4
  token_moe_intermediate_size: 512
  token_shared_intermediate_size: 704
  bias_update_speed: 0
  sequence_wise_mode: "per_sequence"
  sequence_wise_loss_coeff: 1e-3
  router_z_loss_coeff: 1e-4
  router_activation: "sigmoid"
  routed_scaling_factor: 4.0
  use_shared_expert_gate: false
  use_moe_expert_lr: false # Distributed Muon uses one learning rate for matrix parameters.
  loss_type: L1_fm
  data_parallel_mode: fsdp2
  enable_full_shard: false
  module_fsdp_enable: true
  use_compile: true
  use_wandb: false
  rmpad: false
  rmpad_with_pos_ids: false
  ulysses_parallel_size: 1
  freeze_vision_encoder: false
  tokenizer_max_length: 72
  action_dim: 55
  max_action_dim: 55
  max_state_dim: 55
  lr: 1.0e-4
  lr_min: 5.0e-5
  lr_decay_style: cosine
  # Distributed Muon requires FSDP2 and at least two data-parallel ranks.
  optimizer: dist_muon
  muon_momentum: 0.95
  muon_nesterov: true
  muon_ns_steps: 5
  muon_adjust_lr_fn: match_rms_adamw
  dist_muon_attn_per_head: true # Update Q/K/V projections per attention head.
  dist_muon_layers_per_bucket: 2 # Group adjacent layers for communication/computation overlap.
  # Use AdamW for these final-layer matrices; adjust names for other architectures.
  muon_exclude_name_patterns:
    - qwenvl.model.language_model.layers.35.mlp.down_proj.weight
    - qwenvl.model.language_model.layers.35.mlp.gate_proj.weight
    - qwenvl.model.language_model.layers.35.mlp.up_proj.weight
    - qwenvl.model.language_model.layers.35.self_attn.o_proj.weight
  num_train_epochs: 29000
  micro_batch_size: 32
  global_batch_size: 1024 # Training uses 32 GPUs: global_batch_size = micro_batch_size * 32
  max_steps: 50000
  ckpt_manager: dcp
  save_steps: 10000
  save_epochs: 29000
  enable_fp32: true
  enable_resume: true
  align_params:
    mode: 'query'
    num_task_tokens: 8
    depth_loss_weight: 0.004
    future_depth_loss_weight: 0.004
    use_future_video: true
    llm:
      dim_out: 2560
      image_token_size: 8
      image_input_size: 224
    depth:
      model_type: MoRGBD
      moge_path: /path/to/depth/moge2-vitb-normal.pt
      morgbd_path: /path/to/depth/model.pt
      num_layers: 1
      num_heads: 4
      dim_head: 32
      ff_mult: 1
      num_backbone_tokens: 256
      token_size: 16
      dim_out: 1024
      input_size: 224
      use_future_depth: true
      block_future_depth_to_action: false
      detach_future_image_feats: true
    video:
      ckpt_path: /path/to/dino_video/teacher_step_10000.pth
      config_path: /path/to/dino_video/config.yaml
      attention_mode: flex_block_causal
      input_size: 256
      block_suffix_to_future_video: false
      share_future_depth_query: true
      use_shared_future_task_proj: true
      use_current_shared_task_proj: true
      num_future_frames: 1
      # Keeps DINO teacher input as [warmup current, current, future].
      # Current-DINO target uses current_index=1 in code.
      use_warmup_frame: true
      effective_fps: 1.0
      n_blocks: 1
      cls_pool: last
      detach_image_feats: true
      num_layers: 1
      num_heads: 4
      dim_head: 32
      ff_mult: 1
      num_backbone_tokens: 256
      dim_out: 1024
      future_video_loss_weight: 0.004
      use_smooth_l1_loss: false
      use_mse_loss: true
      mse_loss_weight: 1.0
      # Enables current-DINO patch alignment and current depth/DINO shared query projection.
      use_patch_loss: true
      use_current_patch_loss: true
      use_cosine_loss: true
      cosine_loss_weight: 0.2
      use_cls_loss: false
      cls_loss_type: mse
      cls_loss_weight: 0.2
      log_max_samples: 32
      log_scale: 16
    visual_steps: 5000