Add missing None check for hf_quantizer (#28804)

* Add missing None check for hf_quantizer

* Add test, fix logic.

* make style

* Switch test model to Mistral

* Comment

* Update tests/test_modeling_utils.py

---------

Co-authored-by: Younes Belkada <49240599+younesbelkada@users.noreply.github.com>

src/transformers/modeling_utils.py

@@ -3727,10 +3727,12 @@ def _fix_key(key):
 
                 if param.device == torch.device("meta"):
                     value = torch.empty(*param.size(), dtype=target_dtype)
-                    if getattr(
-                        hf_quantizer, "requires_parameters_quantization", False
-                    ) or not hf_quantizer.check_quantized_param(
-                        model, param_value=value, param_name=key, state_dict={}
+                    if (
+                        hf_quantizer is None
+                        or getattr(hf_quantizer, "requires_parameters_quantization", False)
+                        or not hf_quantizer.check_quantized_param(
+                            model, param_value=value, param_name=key, state_dict={}
+                        )
                     ):
                         set_module_tensor_to_device(model, key, "cpu", value)
                     else:

tests/test_modeling_utils.py

@@ -34,6 +34,7 @@
 from transformers import (
     AutoConfig,
     AutoModel,
+    AutoModelForSequenceClassification,
     OwlViTForObjectDetection,
     PretrainedConfig,
     is_torch_available,
@@ -201,6 +202,7 @@ def forward(self, mask, inputs_embeds):
 
 TINY_T5 = "patrickvonplaten/t5-tiny-random"
 TINY_BERT_FOR_TOKEN_CLASSIFICATION = "hf-internal-testing/tiny-bert-for-token-classification"
+TINY_MISTRAL = "hf-internal-testing/tiny-random-MistralForCausalLM"
 
 
 def check_models_equal(model1, model2):
@@ -300,6 +302,15 @@ def test_model_from_pretrained_with_different_pretrained_model_name(self):
                 BertModel.from_pretrained(TINY_T5)
         self.assertTrue("You are using a model of type t5 to instantiate a model of type bert" in cl.out)
 
+    @require_accelerate
+    def test_model_from_pretrained_with_none_quantization_config(self):
+        # Needs a device_map for to enter the low_cpu_mem branch. We also load AutoModelForSequenceClassification
+        # deliberately to enter the missing keys branch.
+        model = AutoModelForSequenceClassification.from_pretrained(
+            TINY_MISTRAL, device_map="auto", quantization_config=None
+        )
+        self.assertIsNotNone(model)
+
     def test_model_from_config_torch_dtype(self):
         # test that the model can be instantiated with dtype of user's choice - as long as it's a
         # float dtype. To make it happen config.torch_dtype needs to be set before instantiating the