Re-enable colqwen 2.5 Omni; fix resize token embeddings

colpali_engine/__init__.py

@@ -16,7 +16,7 @@
     ColQwen2,
     ColQwen2_5,
     ColQwen2_5_Processor,
-    # ColQwen2_5Omni,
-    # ColQwen2_5OmniProcessor,
+    ColQwen2_5Omni,
+    ColQwen2_5OmniProcessor,
     ColQwen2Processor,
 )

colpali_engine/models/__init__.py

@@ -3,3 +3,4 @@
 from .paligemma import BiPali, BiPaliProcessor, BiPaliProj, ColPali, ColPaliProcessor
 from .qwen2 import BiQwen2, BiQwen2Processor, ColQwen2, ColQwen2Processor
 from .qwen2_5 import BiQwen2_5, BiQwen2_5_Processor, ColQwen2_5, ColQwen2_5_Processor
+from .qwen_omni import ColQwen2_5Omni, ColQwen2_5OmniProcessor

colpali_engine/models/qwen_omni/__init__.py

@@ -0,0 +1 @@
+from .colqwen_omni import ColQwen2_5Omni, ColQwen2_5OmniProcessor

colpali_engine/models/qwen_omni/colqwen_omni/__init__.py

@@ -0,0 +1,2 @@
+from .modeling_colqwen_omni import ColQwen2_5Omni
+from .processing_colqwen_omni import ColQwen2_5OmniProcessor

colpali_engine/models/qwen_omni/colqwen_omni/modeling_colqwen_omni.py

@@ -21,9 +21,17 @@ def __init__(self, config: Qwen2_5OmniThinkerConfig, mask_non_image_embeddings:
         self.lm_head = nn.Identity()  # Disable the original lm_head
         self.padding_side = "left"
         self.mask_non_image_embeddings = mask_non_image_embeddings
-        self.lm_head = nn.Identity()  # Disable the original lm_head
         self.post_init()
 
+    def get_output_embeddings(self) -> None:  # -> None | Any:
+        """
+        Transformers >=4.54.0 fails during resize_token_embeddings() due to a new get_output_embeddings()
+        impl. The latter used to return None unless overridden, but they made it try harder to return
+        *something*. Of course, this was not flagged as a breaking change. Eventually I found the
+        responsible PR, and it endorses this change: https://github.com/huggingface/transformers/pull/39339
+        """
+        return None
+
     def forward(self, *args, **kwargs) -> torch.Tensor:
         # # Handle the custom "pixel_values" input obtained with `ColQwen2Processor` through unpadding
         # if "pixel_values" in kwargs: