Fix FSMT weight sharing (#26292)

src/transformers/models/fsmt/modeling_fsmt.py

@@ -1056,6 +1056,7 @@ class FSMTModel(PretrainedFSMTModel):
         return self.decoder
 
     def _tie_weights(self):
+        if self.config.tie_word_embeddings:
             self._tie_or_clone_weights(self.decoder.embed_tokens, self.get_input_embeddings())
             self._tie_or_clone_weights(self.decoder.output_projection, self.get_input_embeddings())
 

tests/models/fsmt/test_modeling_fsmt.py

@@ -273,6 +273,8 @@ class FSMTModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin
 
     def test_ensure_weights_are_shared(self):
         config, inputs_dict = self.model_tester.prepare_config_and_inputs()
+
+        config.tie_word_embeddings = True
         model = FSMTForConditionalGeneration(config)
 
         # FSMT shares three weights.
@@ -288,6 +290,22 @@ class FSMTModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin
             1,
         )
 
+        config.tie_word_embeddings = False
+        model = FSMTForConditionalGeneration(config)
+
+        # FSMT shares three weights.
+        # Not an issue to not have these correctly tied for torch.load, but it is an issue for safetensors.
+        self.assertEqual(
+            len(
+                {
+                    model.get_output_embeddings().weight.data_ptr(),
+                    model.get_input_embeddings().weight.data_ptr(),
+                    model.base_model.decoder.output_projection.weight.data_ptr(),
+                }
+            ),
+            2,
+        )
+
     @unittest.skip("can't be implemented for FSMT due to dual vocab.")
     def test_resize_tokens_embeddings(self):
         pass