from transformers import Qwen2_5_VLForConditionalGeneration from .._mixin import OstrisTransformersMixin class Qwen25VLTextEncoder(Qwen2_5_VLForConditionalGeneration, OstrisTransformersMixin): """Qwen2.5-VL conditioning stack (qwen_image family). Loads from a checkpoint's text_encoder/ subfolder; the edit variants keep the vision tower, plain t2i drops it.""" aitk_subfolder = "text_encoder" aitk_tokenizer_subfolder = "tokenizer" @classmethod def get_transformer_block_names(cls): return ["model.language_model.layers"] def drop_vision_tower(self): """Text-only conditioning: the vision tower is dead weight.""" if getattr(self.model, "visual", None) is not None: self.model.visual = None return self