_target_: ardy.model.Ardy
num_base_steps: 10
cfg_type: separated
gen_horizon_len: 8
autoencoder:
  _target_: ardy.model.autoencoder.fsq.FSQVAETransformer
  ckpt_path: ${oc.select:checkpoint_dir}/tokenizer.safetensors
  pose_root_mode: pose
  encode_with_normalization: true
  encode_with_quantization: true
  motion_rep:
    _target_: ardy.motion_rep.ArdyMotionRep
    fps: 20
    stats_path: ${oc.select:checkpoint_dir}/stats/motion/
    skeleton:
      _target_: ardy.skeleton.CoreSkeleton27
  feature_mode:
  - body
  - pose
  - body
  - root
  num_fsq_levels: 128
  fsq_level_list: 64
  latent_embedding_dim: 128
  num_frames_per_token: 4
  latent_dim: 512
  ff_size: 1024
  num_layers: 8
  num_heads: 4
  activation: gelu
  dropout: 0.0
  pe_dropout: 0.0
  norm_first: false
  causal_encoder: true
  causal_decoder: true
denoiser:
  _target_: ardy.model.auto_latent_twostage_denoiser.AutoLatentTwostageDenoiser
  ckpt_path: ${oc.select:checkpoint_dir}/denoiser.safetensors
  motion_mask_mode: concat
  num_frames_per_token: 4
  nframe_root_dim: 20
  latent_embedding_dim: 128
  motion_rep:
    _target_: ardy.motion_rep.ArdyMotionRep
    fps: 20
    stats_path: ${oc.select:checkpoint_dir}/stats/motion/
    skeleton:
      _target_: ardy.skeleton.CoreSkeleton27
  llm_shape:
  - 1
  - 4096
  use_text_mask: false
  latent_dim: 1024
  ff_size: 2048
  num_layers: 8
  num_heads: 8
  activation: gelu
  dropout: 0.0
  pe_dropout: 0.0
  norm_first: false
  input_first_heading_angle: true
  positional_encoding_mode: learned_prefix_zero_at_first_generation