Rework merge_network_on_save to handle dequantization on merging and saving for much more efficient full finetuning.

toolkit/models/base_model.py

@@ -1594,6 +1594,12 @@ class BaseModel:
     def get_transformer_block_names(self) -> Optional[List[str]]:
         # override in child classes to get transformer block names for lora targeting
         return None
+
+    def get_quantization_exclude_modules(self) -> Optional[List[str]]:
+        # override in child classes to keep sensitive modules in full precision when
+        # quantizing. Returns fnmatch patterns matched against the transformer's module
+        # names (e.g. "model.x_embedder*").
+        return None
     
     def get_base_model_version(self) -> str:
         # override in child classes to get the base model version