PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度
fix gemma vlm gguf regression (#685)
Wenhua Cheng3f36254c380b
diff --git a/auto_round/export/export_to_gguf/convert.py b/auto_round/export/export_to_gguf/convert.pyindex 034b86590..45b99d2f0 100644--- a/auto_round/export/export_to_gguf/convert.py+++ b/auto_round/export/export_to_gguf/convert.py@@ -767,9 +767,9 @@ class ModelBase(OriModel):if "visual." in name and not name.startswith("visual."):name = ".".join(name.split(".")[1:])- # for gemma3+ # # for gemma3if self.model_arch == gguf.MODEL_ARCH.GEMMA3 or isinstance(self, Gemma3VisionModel):- visual_keys = ["multi_modal_projector", "vision_tower", "multimodal_projector", "vision_model"]+ visual_keys = ["multi_modal_projector", "vision_tower", "multimodal_projector"]name = remove_prefix(name, visual_keys)# for LlavaForConditionalGeneration