Improve multimodal hasher performance for re-used Image prompts (#22825)

Signed-off-by: Staszek Pasko <staszek@gmail.com>

tests/multimodal/test_hasher.py

 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import uuid
 from pathlib import Path
 
 import numpy as np
@@ -72,3 +73,22 @@ def test_hash_non_contiguous_array():
     hasher = MultiModalHasher
     # Both should be hashable and produce the same hashes
     assert hasher.hash_kwargs(data=arr) == hasher.hash_kwargs(data=arr_c)
+
+
+def test_hash_image_exif_id():
+    # Test that EXIF ImageId tag can be used to store UUID
+    # and the hasher will use that instead of the image data.
+    image1 = image2 = Image.new("1", size=(10, 20))
+    id = uuid.uuid4()
+    image1.getexif()[Image.ExifTags.Base.ImageID] = id
+    image2 = Image.open(ASSETS_DIR / "image1.png")
+    image2.getexif()[Image.ExifTags.Base.ImageID] = "Not a UUID"
+    image2a = Image.open(ASSETS_DIR / "image1.png")
+
+    hasher = MultiModalHasher
+    # first image has UUID in ImageID, so it should hash to that UUID
+    assert hasher.hash_kwargs(image=image1) == hasher.hash_kwargs(
+        image=id.bytes)
+    # second image has non-UUID in ImageID, so it should hash to the image data
+    assert hasher.hash_kwargs(image=image2) == hasher.hash_kwargs(
+        image=image2a)

vllm/multimodal/hasher.py

@@ -2,6 +2,7 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import pickle
+import uuid
 from collections.abc import Iterable, Mapping
 from typing import Union
 
@@ -34,6 +35,11 @@ class MultiModalHasher:
             return np.array(obj).tobytes()
 
         if isinstance(obj, Image.Image):
+            exif = obj.getexif()
+            if Image.ExifTags.Base.ImageID in exif and isinstance(
+                    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.item_to_bytes(
                 "image", np.asarray(convert_image_mode(obj, "RGBA")))
         if isinstance(obj, torch.Tensor):