cogact.yaml

# CogACT model server — base model (microsoft/CogACT)
# Weight: CogACT/CogACT-Base (HuggingFace)
# Output: 16 actions × 7-DoF (future_action_window_size=15 → 16 steps)
#
# Usage: vla-eval serve --config configs/model_servers/cogact.yaml
#
# Available checkpoints:
#   CogACT/CogACT-Small  (action_model_type: DiT-S)
#   CogACT/CogACT-Base   (action_model_type: DiT-B)
#   CogACT/CogACT-Large  (action_model_type: DiT-L)
#
# unnorm_key selects the dataset statistics used for action
# denormalization.  Required for multi-dataset checkpoints like
# CogACT-Base.  Common keys: fractal20220817_data (RT-1),
# bridge_orig (BridgeData V2), kuka, bc_z, taco_play, etc.

script: "src/vla_eval/model_servers/cogact.py"
args:
  model_path: CogACT/CogACT-Base
  action_model_type: DiT-B
  future_action_window_size: 15
  unnorm_key: bridge_orig
  chunk_size: 16
  cfg_scale: 1.5
  use_ddim: true
  num_ddim_steps: 10
  port: 8000