from __future__ import annotations from psi.config.model_psi0 import Psi0ModelConfig class MotionDistillPsi0ModelConfig(Psi0ModelConfig): # "teacher": motion -> hidden -> action head. # "student": video/audio/text -> Qwen hidden, supervised by teacher hidden. motion_distill_stage: str = "teacher" # Stage-1 motion encoder. motion_encoder_input_horizon: int = 0 motion_encoder_hidden_dim: int = 2048 motion_encoder_dropout: float = 0.0 motion_latent_num_tokens: int = 64 motion_latent_token: str = "<|motion_latent|>" # Stage-2 distillation. motion_teacher_checkpoint: str | None = None motion_latent_cache_dir: str | None = None motion_latent_cache_strict: bool = False motion_hidden_loss_weight: float = 1.0 motion_hidden_cosine_loss_weight: float = 0.0 motion_student_action_loss_weight: float = 0.0 # Student VLM train scope. "video" trains text backbone + visual encoder and # keeps audio/lm_head frozen, which avoids unused trainable parameters for # video-only batches under DDP. motion_distill_vlm_train_scope: str = "video"