up

Author: sayakpaul

src/diffusers/quantizers/nunchaku/nunchaku_quantizer.py

@@ -169,13 +169,16 @@ def _process_model_before_weight_loading(
             model,
             modules_to_not_convert=self.modules_to_not_convert,
             quantization_config=self.quantization_config,
-            pre_quantized=self.pre_quantized,
         )
         model.config.quantization_config = self.quantization_config
 
     def _process_model_after_weight_loading(self, model, **kwargs):
         return model
 
-    # @property
-    # def is_serializable(self):
-    #     return True
+    @property
+    def is_serializable(self):
+        return False
+
+    @property
+    def is_trainable(self):
+        return False

src/diffusers/quantizers/nunchaku/utils.py

@@ -5,16 +5,15 @@
 
 if is_accelerate_available():
     from accelerate import init_empty_weights
-
-if is_nunchaku_available():
-    from nunchaku.models.linear import SVDQW4A4Linear
+    
 
 
 logger = logging.get_logger(__name__)
 
 
 def _replace_with_nunchaku_linear(
     model,
+    svdq_linear_cls,
     modules_to_not_convert=None,
     current_key_name=None,
     quantization_config=None,
@@ -36,7 +35,7 @@ def _replace_with_nunchaku_linear(
                     out_features = module.out_features
 
                     if quantization_config.precision in ["int4", "nvfp4"]:
-                        model._modules[name] = SVDQW4A4Linear(
+                        model._modules[name] = svdq_linear_cls(
                             in_features,
                             out_features,
                             rank=quantization_config.rank,
@@ -62,7 +61,10 @@ def _replace_with_nunchaku_linear(
 
 
 def replace_with_nunchaku_linear(model, modules_to_not_convert=None, current_key_name=None, quantization_config=None):
-    model, _ = _replace_with_nunchaku_linear(model, modules_to_not_convert, current_key_name, quantization_config)
+    if is_nunchaku_available():
+        from nunchaku.models.linear import SVDQW4A4Linear
+
+    model, _ = _replace_with_nunchaku_linear(model, SVDQW4A4Linear, modules_to_not_convert, current_key_name, quantization_config)
 
     has_been_replaced = any(
         isinstance(replaced_module, SVDQW4A4Linear) for _, replaced_module in model.named_modules()

src/diffusers/quantizers/quantization_config.py

@@ -762,3 +762,26 @@ def post_init(self):
         accpeted_precision = ["int4", "nvfp4"]
         if self.precision not in accpeted_precision:
             raise ValueError(f"Only supported precision in {accpeted_precision} but found {self.precision}")
+
+    # Copied from diffusers.quantizers.bitsandbytes.quantization_config.BitsandBytesConfig.to_diff_dict with BitsandBytesConfig->NunchakuConfig
+    def to_diff_dict(self) -> Dict[str, Any]:
+        """
+        Removes all attributes from config which correspond to the default config attributes for better readability and
+        serializes to a Python dictionary.
+
+        Returns:
+            `Dict[str, Any]`: Dictionary of all the attributes that make up this configuration instance,
+        """
+        config_dict = self.to_dict()
+
+        # get the default config dict
+        default_config_dict = NunchakuConfig().to_dict()
+
+        serializable_config_dict = {}
+
+        # only serialize values that differ from the default config
+        for key, value in config_dict.items():
+            if value != default_config_dict[key]:
+                serializable_config_dict[key] = value
+
+        return serializable_config_dict