[Bugfix] Handle HWC images in ImageProcessorItems.get_image_size (#45057)

Signed-off-by: YellowFoxH4XOR <[email protected]>
Co-authored-by: Claude <[email protected]>

tests/multimodal/test_parse.py

@@ -0,0 +1,51 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import numpy as np
+import pytest
+import torch
+from PIL import Image
+
+from vllm.multimodal.parse import ImageProcessorItems, VideoProcessorItems
+
+H, W = 480, 640
+
+
+@pytest.mark.parametrize(
+    "image",
+    [
+        Image.new("RGB", (W, H)),
+        # HWC, e.g. from np.array(PIL.Image)
+        np.zeros((H, W, 3), dtype=np.uint8),
+        torch.zeros((H, W, 3), dtype=torch.uint8),
+        # CHW, standard PyTorch / numpy convention
+        np.zeros((3, H, W), dtype=np.uint8),
+        torch.zeros((3, H, W), dtype=torch.uint8),
+    ],
+)
+def test_image_size_hwc_chw(image):
+    """Image sizes must be channel-layout agnostic.
+
+    `get_image_size` determines the multimodal placeholder count; reading an
+    HWC array (the layout `np.array(PIL.Image)` produces) as CHW yields a
+    bogus size and a placeholder/embedding count mismatch at inference time.
+    """
+    items = ImageProcessorItems([image])
+
+    assert items.get_image_size(0) == (W, H)
+
+
+@pytest.mark.parametrize(
+    "frame",
+    [
+        Image.new("RGB", (W, H)),
+        np.zeros((H, W, 3), dtype=np.uint8),
+        torch.zeros((H, W, 3), dtype=torch.uint8),
+        np.zeros((3, H, W), dtype=np.uint8),
+        torch.zeros((3, H, W), dtype=torch.uint8),
+    ],
+)
+def test_frame_size_hwc_chw(frame):
+    """`get_frame_size` must stay consistent with `get_image_size`."""
+    items = VideoProcessorItems([[frame]])
+
+    assert items.get_frame_size(0) == (W, H)

vllm/multimodal/parse.py

@@ -334,7 +334,13 @@ class ImageProcessorItems(ProcessorBatchItems[HfImageItem | None]):
         if isinstance(image, PILImage.Image):
             return ImageSize(*image.size)
         if isinstance(image, (np.ndarray, torch.Tensor)):
-            _, h, w = image.shape
+            if image.ndim == 3 and image.shape[-1] in (1, 3, 4):
+                # HWC format (e.g. from np.array(PIL.Image)).
+                # PIL images are always channels-last.
+                h, w = image.shape[0], image.shape[1]
+            else:
+                # CHW format (standard PyTorch / numpy convention).
+                _, h, w = image.shape
             return ImageSize(w, h)
 
         assert_never(image)