# Inference-only configuration. Load model.pt strictly after model construction.
# Supply matching text embeddings, or enable the compatible text encoder.
checkpoint: ./model.pt
mixed_precision: bf16
eval_num_inference_steps: 10
data:
  train:
    shape_meta:
      images:
      - key: robot0_agentview_left
        raw_shape:
        - 3
        - 256
        - 256
        shape:
        - 3
        - 256
        - 256
      - key: robot0_agentview_right
        raw_shape:
        - 3
        - 256
        - 256
        shape:
        - 3
        - 256
        - 256
      - key: robot0_eye_in_hand
        raw_shape:
        - 3
        - 256
        - 256
        shape:
        - 3
        - 256
        - 256
      action:
      - key: default
        raw_shape: 12
        shape: 12
      state:
      - key: default
        raw_shape: 16
        shape: 16
    num_frames: 81
    past_obs_size: 48
    action_chunk: 32
    action_video_freq_ratio: 4
    video_size:
    - 384
    - 320
    concat_multi_camera: robocasa_left_main
    context_len: 128
    processor:
      _target_: fastwam.datasets.lerobot.processors.fastwam_processor.FastWAMProcessor
      shape_meta:
        images:
        - key: robot0_agentview_left
          raw_shape:
          - 3
          - 256
          - 256
          shape:
          - 3
          - 256
          - 256
        - key: robot0_agentview_right
          raw_shape:
          - 3
          - 256
          - 256
          shape:
          - 3
          - 256
          - 256
        - key: robot0_eye_in_hand
          raw_shape:
          - 3
          - 256
          - 256
          shape:
          - 3
          - 256
          - 256
        action:
        - key: default
          raw_shape: 12
          shape: 12
        state:
        - key: default
          raw_shape: 16
          shape: 16
      num_obs_steps: 81
      num_image_steps: 21
      num_output_cameras: 3
      action_output_dim: 12
      proprio_output_dim: 16
      delta_action_dim_mask:
        default:
        - true
        - true
        - true
        - true
        - false
        - true
        - true
        - true
        - true
        - true
        - true
        - false
      action_state_transforms: null
      use_stepwise_action_norm: false
      norm_default_mode: min/max
      norm_exception_mode: null
      action_state_merger:
        _target_: fastwam.datasets.lerobot.transforms.action_state_merger.ConcatLeftAlign
      val_transforms:
      - _target_: fastwam.datasets.lerobot.transforms.image.ToTensor
    pretrained_norm_stats: ./dataset_stats.json
model:
  _target_: fastwam.runtime.create_arwam
  model_id: Wan-AI/Wan2.2-TI2V-5B
  tokenizer_model_id: Wan-AI/Wan2.1-T2V-1.3B
  tokenizer_max_len: 128
  load_text_encoder: false
  proprio_dim: 16
  redirect_common_files: false
  mot_checkpoint_mixed_attn: true
  skip_dit_load_from_pretrain: true
  skip_video_dit_load_from_pretrain: true
  skip_action_dit_load_from_pretrain: true
  longlive_video_weights: null
  action_dit_pretrained_path: null
  arwam_num_frame_per_block: 1
  arwam_local_attn_size: -1
  arwam_sink_size: 0
  arwam_past_obs_size: 48
  arwam_num_imagine_frames: 2
  arwam_imagine_sigma: 0.9
  arwam_imagine_infer_steps: 4
  video_dit_config:
    has_image_input: false
    patch_size:
    - 1
    - 2
    - 2
    in_dim: 48
    hidden_dim: 3072
    ffn_dim: 14336
    freq_dim: 256
    text_dim: 4096
    out_dim: 48
    num_heads: 24
    attn_head_dim: 128
    num_layers: 30
    eps: 1.0e-06
    seperated_timestep: true
    require_clip_embedding: false
    require_vae_embedding: false
    fuse_vae_embedding_in_latents: true
    use_gradient_checkpointing: true
    video_attention_mask_mode: per_frame_causal
    action_conditioned: false
    action_dim: 12
    action_group_causal_mask_mode: group_diagonal
  action_dit_config:
    action_dim: 12
    hidden_dim: 1024
    ffn_dim: 4096
    num_heads: 24
    attn_head_dim: 128
    num_layers: 30
    text_dim: 4096
    freq_dim: 256
    eps: 1.0e-06
    use_gradient_checkpointing: true
  video_scheduler:
    train_shift: 5.0
    infer_shift: 5.0
    num_train_timesteps: 1000
  action_scheduler:
    train_shift: 5.0
    infer_shift: 5.0
    num_train_timesteps: 1000