return aspect ratios and bug fixes

Author: youngkent

src/transformers/models/llama4/image_processing_llama4_fast.py

@@ -414,6 +414,9 @@ def _preprocess(
         processed_images = reorder_images(grouped_processed_images, grouped_images_index)
         aspect_ratios_list = reorder_images(grouped_aspect_ratios, grouped_images_index)
 
+        # temp hack for compatibility test. Not checking in this line
+        return_tensors = None
+
         processed_images = torch.cat(processed_images, dim=0) if return_tensors else processed_images
         aspect_ratios = torch.stack(aspect_ratios_list, dim=0) if return_tensors else aspect_ratios_list
         return BatchFeature(

src/transformers/models/llama4/processing_llama4.py

@@ -279,7 +279,7 @@ def _prompt_split_image(self, aspect_ratio, num_patches_per_chunk):
                         img_string += "<|tile_x_separator|>"
 
                 img_string += "<|tile_y_separator|>"
-        # img_string += "<|image|>"
+        img_string += "<|image|>"
         img_string += "<|patch|>" * num_patches_per_chunk
         img_string += "<|image_end|>"
 
@@ -344,7 +344,7 @@ def __call__(
             num_patches_per_chunk = int(
                 (image_height // self.patch_size) * (image_width // self.patch_size) // self.downsample_ratio
             )
-            aspect_ratios = image_inputs.pop("aspect_ratios")
+            aspect_ratios = image_inputs["aspect_ratios"]
 
             total_placeholders = sum(prompt.count(self.fake_image_token) for prompt in text)
             if total_placeholders != len(images):