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

extensions_built_in/diffusion_models/hidream/hidream_o1_model.py

@@ -489,6 +489,7 @@ class HidreamO1Model(BaseModel):
         return False
 
     def save_model(self, output_path, meta, save_dtype):
+        from toolkit.util.quantize import dequantize_if_quantized
         transformer: Qwen3VLForConditionalGeneration = unwrap_model(self.model)
         if self.is_comfy_weight:
             sd = transformer.state_dict()
@@ -496,7 +497,8 @@ class HidreamO1Model(BaseModel):
             for key, value in sd.items():
                 if "lm_head.weight" in key:
                     continue  # comfy checkpoint doesnt have the lm head, so skip it
-                save_dict[key] = value.clone().to("cpu", dtype=save_dtype)
+                # dequantize any quantized (e.g. torchao) weights so we save plain full precision tensors
+                save_dict[key] = dequantize_if_quantized(value).clone().to("cpu", dtype=save_dtype)
             
             if not output_path.endswith(".safetensors"):
                 output_path += ".safetensors"