remove checking for base models as it seems ok for them to go through…

…, and the checking might fail anyway

src/transformers/utils/bitsandbytes.py

@@ -264,20 +264,18 @@ def get_keys_to_not_convert(model):
     else:
         tied_keys = sum(tied_params, [])
     has_tied_params = len(tied_keys) > 0
-
-    # Check if it is a base model
-    is_base_model = not hasattr(model, model.base_model_prefix)
-
-    # Ignore this for base models (BertModel, GPT2Model, etc.)
-    if (not has_tied_params) and is_base_model:
-        return []
-
+    if not has_tied_params:
+        # if the model has no tied parameters, we check if the output embeddings are the same as the input embeddings
+        # if they are the same, it just means that the model is using the same weights for the input and output
+        if id(model.get_output_embeddings()) == id(model.get_input_embeddings()):
+            has_tied_params = True
+    # In addition to tied-parameters, we will also keep the "lm-head" in full precision for numerical stability reasons
     output_embeddings = model.get_output_embeddings()
-    # If no output-embeddings are found, and no tied parameters, we quantize the whole model
+    # If no output-embeddings are found, and no tied parameters, we have to quantize the whole model
     if output_embeddings is None and not has_tied_params:
         logger.warning(
             "No output embeddings and tied parameters found, will quantize the whole model. "
-            "This could lead to numerical instability if word/output embeddings are quantized. "
+            "This might lead to numerical instability if word/output embeddings are quantized. "
             "If this is not intended, please check tied parameters definition in your model architecture "
             "or override `get_output_embeddings()` to return correct output embeddings in your model class."
         )