[Core][Multimodal] Convert PIL Image to array without data copy when hashing (#18682)

Signed-off-by: Lukas Geiger <lukas.geiger94@gmail.com>

Author: lgeiger

vllm/multimodal/hasher.py

@@ -36,8 +36,8 @@ def serialize_item(cls, obj: object) -> bytes:
             return np.array(obj).tobytes()
 
         if isinstance(obj, Image.Image):
-            return cls.item_to_bytes("image",
-                                     np.array(convert_image_mode(obj, "RGBA")))
+            return cls.item_to_bytes(
+                "image", np.asarray(convert_image_mode(obj, "RGBA")))
         if isinstance(obj, torch.Tensor):
             return cls.item_to_bytes("tensor", obj.numpy())
         if isinstance(obj, np.ndarray):

vllm/multimodal/video.py

@@ -164,7 +164,7 @@ def load_base64(self, media_type: str, data: str) -> npt.NDArray:
             )
 
             return np.stack([
-                np.array(load_frame(frame_data))
+                np.asarray(load_frame(frame_data))
                 for frame_data in data.split(",")
             ])