| from __future__ import annotations | |
| from transformers import PretrainedConfig | |
| class MiniCPMVLAConfig(PretrainedConfig): | |
| """Configuration for the complete VLM and action-head VLA model.""" | |
| model_type = "minicpm_vla" | |
| is_composition = True | |
| def __init__( | |
| self, | |
| vlm_config: dict | PretrainedConfig | None = None, | |
| action_dim: int = 80, | |
| state_dim: int = 80, | |
| action_horizon: int = 30, | |
| max_num_embodiments: int = 32, | |
| num_inference_timesteps: int = 4, | |
| vlm_dtype: str = "bfloat16", | |
| action_head_dtype: str = "float32", | |
| **kwargs, | |
| ): | |
| super().__init__(**kwargs) | |
| if isinstance(vlm_config, PretrainedConfig): | |
| vlm_config = vlm_config.to_dict() | |
| self.vlm_config = vlm_config | |
| self.action_dim = action_dim | |
| self.state_dim = state_dim | |
| self.action_horizon = action_horizon | |
| self.max_num_embodiments = max_num_embodiments | |
| self.num_inference_timesteps = num_inference_timesteps | |
| self.vlm_dtype = vlm_dtype | |
| self.action_head_dtype = action_head_dtype | |
| self.architectures = ["MiniCPMV_VLA"] | |