model:
  video_backbone:
    encoder:
      name: wan_vae
      model_path: /mnt/data/wangyuran/Wan2.2-TI2V-5B
    name: wan22_ti2v_5b
    model_path: /mnt/data/wangyuran/Wan2.2-TI2V-5B
    from_scratch: false
    shift_video: 5.0
    components:
    - attr: dit
      model_class: openwam.model.video_backbone.wan.models.dit.WanModel
      extra_kwargs:
        has_image_input: false
        patch_size:
        - 1
        - 2
        - 2
        in_dim: 48
        dim: 3072
        ffn_dim: 14336
        freq_dim: 256
        text_dim: 4096
        out_dim: 48
        num_heads: 24
        num_layers: 30
        eps: 1.0e-06
        seperated_timestep: true
        require_clip_embedding: false
        require_vae_embedding: false
        fuse_vae_embedding_in_latents: true
    - attr: vae
      model_class: openwam.model.video_backbone.wan.models.vae.WanVideoVAE38
      extra_kwargs: {}
    - attr: text_encoder
      model_class: openwam.model.video_backbone.wan.models.text_encoder.WanTextEncoder
      extra_kwargs: {}
    tokenizer:
      class: openwam.model.video_backbone.wan.models.text_encoder.HuggingfaceTokenizer
      attr: tokenizer
      subdir: tokenizer/google/umt5-xxl
      path_kwarg: name
      kwargs:
        seq_len: 512
        clean: whitespace
  action_backbone:
    dim: 1024
    ffn_dim: 4096
    shift_action: 5.0
  freeze:
  - video_backbone.text_encoder
  - video_backbone.reason1
  - video_backbone.vae
  - video_backbone.image_encoder
  - video_backbone.video_encoder
  architecture:
    framework: dual_system
    variant: joint_self_attn
    action_dim: 80
    use_proprioception: true
    state_dim: 80
    bridge_layers: null
    bridge_interval: 1
    mot_checkpoint_mixed_attn: true
    attention_mask_mode: action_sees_video
    video_attention_mask_mode: first_frame_causal
    detach_bridge: false
    idm_video_cond_noise_prob: 0.5
dataloader:
  type: wuji_real_task
  dataset_dir: /mnt/data/liangjiaqi/wuji_sft_data_eef
  num_frames: 33
  video_stride: 4
  window_stride: 1
  height: 384
  width: 320
  multiview: true
  target_camera: observation.images.stereo
  camera_layout:
  - observation.images.stereo
  - observation.images.cam_left_wrist
  - observation.images.cam_right_wrist
  normalize_mode: quantile
  split: train
  enable_action_supervision: true
  unify_action: true
  unify_action_map:
  - 0-8
  - 10-29
  - 32-40
  - 42-61
  seed: 42
training:
  debug: false
  learning_rate: 0.0001
  adam_betas:
  - 0.9
  - 0.95
  weight_decay: 0.01
  max_grad_norm: 1.0
  num_epochs: null
  max_steps: 100000
  batch_size: 2
  gradient_accumulation_steps: 1
  lr_scheduler: cosine
  warmup_ratio: 0.05
  lr_min_ratio: 0.01
  action_lr: null
  video_lr: null
  mixed_precision: bf16
  zero_stage: 2
  use_gradient_checkpointing: false
  use_gradient_checkpointing_offload: false
  initialize_model_on_cpu: false
  offload_optimizer_device: none
  lambda_video: 1.0
  lambda_action: 1.0
  max_timestep_boundary: 1.0
  min_timestep_boundary: 0.0
  timestep_sampling: default
  timestep_sampling_lead: video
  timestep_sampling_alpha: 1.0
  output_path: /mnt/data/liangjiaqi/openwam_checkpoints
  save_steps: 2000
  keep_last_k_ckpts: 1
  finetune_ckpt_path: /mnt/data/wangyuran/openwam_checkpoints/pretrained_final_dualsystem_jointselfattention
  resume_ckpt_path: null
  dataset_num_workers: 8
project:
  name: open-wam
  output_dir: outputs/${now:%Y-%m-%d_%H-%M-%S}
  seed: 42
  wandb:
    project: open-wam
    run_name: wuji-eef-sft
    entity: null