Complete Transformers 5.x multimodal and save support (#2)

- Complete Transformers 5.x multimodal and save support (c27010395083d85570830cd469c545e582b33076)
- Fix Transformers 5 tied-weight save mapping (6030c5ce87f00639f2d8e3461954bbaacdb3a0eb)

modeling_colqwen3.py

@@ -139,7 +139,10 @@ class ColQwen3(ColQwen3PreTrainedModel):
         )
         self.padding_side = getattr(config, "padding_side", "left")
         self.mask_non_image_embeddings = mask_non_image_embeddings
-        self._tied_weights_keys = [f"vlm.{k}" for k in (self.vlm._tied_weights_keys or [])]
+        inner_tied = self.vlm._tied_weights_keys or {}
+        self._tied_weights_keys = {
+            f"vlm.{target}": f"vlm.{source}" for target, source in inner_tied.items()
+        }
 
         self.post_init()
 

processing_colqwen3.py

@@ -57,6 +57,7 @@ class ColQwen3ProcessorKwargs(ProcessingKwargs, total=False):
     _defaults = {
         "text_kwargs": {
             "padding": "longest",
+            "return_mm_token_type_ids": True,
         },
         "images_kwargs": {
             "data_format": "channels_first",
@@ -401,10 +402,7 @@ class ColQwen3Processor(ProcessorMixin):
         self._check_special_mm_tokens(text_prompts, text_inputs, modalities=["image", "video"])
 
         if return_mm_token_type_ids:
-            array_ids = np.array(text_inputs["input_ids"])
-            mm_token_type_ids = np.zeros_like(text_inputs["input_ids"])
-            mm_token_type_ids[array_ids == self.image_token_id] = 1
-            text_inputs["mm_token_type_ids"] = mm_token_type_ids.tolist()
+            text_inputs["mm_token_type_ids"] = self.create_mm_token_type_ids(text_inputs["input_ids"])
 
         return BatchFeature(data={**text_inputs, **image_inputs, **videos_inputs}, tensor_type=return_tensors)