Always initialize tied output_embeddings if it has a bias term (#28947)

Continue to initialize tied output_embeddings if it has a bias term

The bias term is not tied, and so will need to be initialized accordingly.

src/transformers/modeling_utils.py

@@ -3748,10 +3748,12 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
                 else:
                     _loaded_keys = loaded_keys
                 not_initialized_submodules = set_initialized_submodules(model, _loaded_keys)
-                # if we're about to tie the output embeds to the input embeds we don't need to init them
+                # If we're about to tie the output embeds to the input embeds we don't need to init them
                 if hasattr(model.config, "tie_word_embeddings") and model.config.tie_word_embeddings:
                     output_embeddings = model.get_output_embeddings()
                     if output_embeddings is not None:
+                        # Still need to initialize if there is a bias term since biases are not tied.
+                        if not hasattr(output_embeddings, "bias") or output_embeddings.bias is None:
                             output_embeddings._is_hf_initialized = True
             else:
                 not_initialized_submodules = dict(model.named_modules())