[Core][MultiModalHasher] Hash images without converting image mode

Signed-off-by: Lukas Geiger <[email protected]>

Author: lgeiger

vllm/multimodal/hasher.py

@@ -12,7 +12,6 @@
 from PIL import Image
 
 from vllm.logger import init_logger
-from vllm.multimodal.image import convert_image_mode
 
 logger = init_logger(__name__)
 
@@ -35,8 +34,10 @@ def serialize_item(cls, obj: object) -> Iterable[Union[bytes, memoryview]]:
                     exif[Image.ExifTags.Base.ImageID], uuid.UUID):
                 # If the image has exif ImageID tag, use that
                 return (exif[Image.ExifTags.Base.ImageID].bytes, )
-            return cls.iter_item_to_bytes(
-                "image", np.asarray(convert_image_mode(obj, "RGBA")))
+            data = {"mode": obj.mode, "data": np.asarray(obj)}
+            if obj.palette is not None:
+                data["palette"] = obj.palette.palette
+            return cls.iter_item_to_bytes("image", data)
         if isinstance(obj, torch.Tensor):
             tensor_obj: torch.Tensor = obj.cpu()
             tensor_dtype = tensor_obj.dtype