model:
  _recursive_: false
  _target_: greenwam.callbacks.evaluation.build_validation_model
  config:
    action_expert: true
    action_gen: true
    action_io_projector_type: domain_aware
    activation_checkpointing:
      determinism_check: default
      mode: selective
      preserve_rng_state: true
      save_ops_regex:
      - fmha
    attention_io_layout: sequence_sharded
    causal_training_strategy: none
    compile:
      ar_post_saturation_mode: default
      compile_dynamic: true
      compiled_region: language
      coordinate_descent_tuning: false
      enabled: true
      max_autotune_pointwise: false
      use_cuda_graphs: false
    diffusion_expert_config:
      base_fps: 24
      enable_action_modality_embedding: true
      enable_fps_modulation: true
      enable_media_modality_embedding: false
      enable_sound_modality_embedding: true
      enable_vision_modality_embeddings: false
      load_weights_from_pretrained: false
      max_vae_latent_side_after_patchify: 20
      patch_spatial: 2
      sound_base_temporal_compression_factor: null
      timestep_range: 1.0
      unified_3d_mrope_reset_spatial_ids: true
      unified_3d_mrope_temporal_modality_margin: 15000
      vision_temporal_position_mode: latent_index
    ema:
      enabled: true
      iteration_shift: 0
      rate: 0.1
    enable_input_bias: true
    exclude_reasoner_weights_from_checkpoint: false
    fixed_step_sampler_config: null
    flex_attention:
      backend: auto
      enabled: false
      mask:
        attention_scope: all_views
        decomposed_temporal_window_seconds: null
    input_caption_key: ai_caption
    input_image_key: images
    input_video_key: video
    joint_attn_implementation: two_way
    latent_downsample_factor: 16
    lbl:
      coeff_gen: null
      coeff_und: null
      method: local
    lidar_fps: null
    lidar_state_ch: null
    lidar_tokenizer: null
    load_vision_tokenizer: true
    lora_alpha: 32
    lora_enabled: false
    lora_rank: 16
    lora_target_modules: q_proj_moe_gen,k_proj_moe_gen,v_proj_moe_gen,o_proj_moe_gen
    max_action_dim: 128
    max_num_tokens_after_packing: 45056
    native_action_dim: 64
    natten_parameter_list: null
    net: null
    num_action_contracts: 90
    num_embodiment_domains: 32
    parallelism:
      cfg_parallel_shard_degree: 1
      context_parallel_shard_degree: 1
      data_parallel_replicate_degree: 2
      data_parallel_shard_degree: 8
      enable_inference_mode: false
      fsdp_master_dtype: float32
      fsdp_reduce_dtype: bfloat16
      vae_load_balance_group_size: 1
    precision: bfloat16
    quantization:
      exclude_regex: []
      fp8_granularity: per_row
      include_regex: []
      method: null
      mixed_precision_first_steps: 0
      mixed_precision_last_steps: 0
      mixed_precision_reasoner_policy: high_precision
      mixed_precision_w8a16_cache: none
      modelopt_fp8_checkpoint_path: null
      modelopt_fp8_target_fqns: []
    rectified_flow_inference_config:
      num_train_timesteps: 1000
      scheduler_type: unipc
      shift: 1
      use_dynamic_shifting: false
    rectified_flow_training_config:
      action_loss_weight: 10.0
      image_loss_scale: null
      independent_action_schedule: false
      independent_sound_schedule: false
      lidar_loss_scale: null
      loss_scale: 10.0
      normalize_loss_by_active: true
      sample_level_loss_averaging: false
      shift:
        '256': 3
        '480': 5
        '720': 10
      shift_action: null
      shift_image: null
      shift_sound: null
      sound_loss_scale: 2.0
      train_time_action_distribution: logitnormal
      train_time_image_distribution: logitnormal
      train_time_sound_distribution: logitnormal
      train_time_video_distribution: waver
      train_time_weight: uniform
      use_discrete_rf: false
      use_dynamic_shift: false
    resolution: '480'
    sound_dim: null
    sound_gen: false
    sound_latent_fps: 25
    sound_tokenizer: null
    state_ch: 48
    state_t: 300
    tokenizer:
      _target_: cosmos_framework.model.generator.tokenizers.wan2pt2_vae_4x16x16.Wan2pt2VAEInterface
      bucket_name: ''
      causal: true
      chunk_duration: 93
      encode_bucket_multiple: null
      encode_chunk_frames:
        '256': 68
        '480': 24
        '720': 8
        '768': 8
      encode_exact_durations:
      - 1
      - 5
      - 9
      - 13
      - 17
      - 21
      - 25
      - 29
      - 33
      keep_decoder_cache: false
      object_store_credential_path_pretrained: ''
      spatial_compression_factor: 16
      temporal_compression_factor: 4
      temporal_window: null
      use_streaming_encode: false
      vae_path: ${oc.env:WAN_VAE_PATH}
    video_temporal_causal: false
    vision_gen: true
    vlm_config:
      layer_module: null
      model_instance:
        _target_: cosmos_framework.model.generator.mot.unified_mot.Nemotron3DenseVLTextForCausalLM
        config:
          _target_: cosmos_framework.configs.base.defaults.reasoner.create_vlm_config
          base_config:
            _target_: cosmos_framework.model.generator.mot.unified_mot.Nemotron3DenseVLMoTConfig.from_json_file
            json_file: cosmos_framework/model/generator/reasoner/nemotron_3_dense_vl/configs/Nemotron-2B-Dense-VL.json
          freeze_und: false
          include_visual: null
          layer_module: MoTDecoderLayer
          qk_norm_for_text: false
          tie_word_embeddings: true
          use_und_k_norm_for_gen: true
      model_name: nvidia/Cosmos3-Edge-Reasoner
      pretrained_weights:
        backbone_path: ''
        checkpoint_format: null
        credentials_path: ''
        enable_gcs_patch_in_boto3: true
        enabled: false
      qk_norm: false
      safetensors_path: ''
      tie_word_embeddings: false
      tokenizer:
        _target_: cosmos_framework.data.generator.processors.build_processor_lazy
        repository: null
        revision: null
        subdir: ''
        tokenizer_type: ${oc.env:COSMOS3_EDGE_DIR}
      use_system_prompt: false
    action_loss_reduction: mean_over_valid_coordinates_v1
    semantic_projectors: true
    inference_noise_source: numpy
    inference_reuse_request_cache: true
    time_embedding_per_token: false
    pad_segment_only_when_padded: false
    normalize_video_in_model_dtype: false
sampling:
  atom_window_policy: full_subtask
  wam_window_policy: dense
  aux_window_policy: wam_window
  n_frames: 17
  wam_n_frames: 33
  wam_horizon_s: 1.0666666666666667