[Bugfix][Multimodal] Fix PyNvVideoCodec video backend returning NCHW instead of NHWC (#51076)

Signed-off-by: Junpu Yu <davidyu@nvidia.com>

tests/multimodal/media/test_video.py

@@ -20,6 +20,7 @@ from vllm.multimodal.video import (
     PYNVVIDEOCODEC_VIDEO_BACKEND,
     VIDEO_LOADER_REGISTRY,
     VideoLoader,
+    _pynvvc_frames_to_nhwc,
 )
 
 from ..utils import cosine_similarity, create_video_from_image, normalize_image
@@ -498,3 +499,22 @@ class TestMergeKwargsGpuBackendPolicy:
 
     def test_unknown_backend_not_treated_as_gpu(self):
         assert not VIDEO_LOADER_REGISTRY.backend_requires_gpu("totally_unknown")
+
+
+@pytest.mark.parametrize("layout", ["nhwc", "nchw"])
+def test_pynvvc_frames_normalized_to_nhwc(layout: str):
+    """PyNvVideoCodec frame batches are normalized to NHWC regardless of the
+    per-frame layout the decoder emits (it has varied across versions), so the
+    HF video processors (which materialize a PIL image per frame) receive the
+    same NHWC shape as every other video backend."""
+    torch = pytest.importorskip("torch")
+
+    n, h, w, c = 4, 5, 6, 3
+    nhwc = torch.arange(n * h * w * c, dtype=torch.uint8).reshape(n, h, w, c)
+    frames = nhwc if layout == "nhwc" else nhwc.permute(0, 3, 1, 2).contiguous()
+
+    out = _pynvvc_frames_to_nhwc(frames)
+
+    assert out.shape == (n, h, w, c)
+    assert out.is_contiguous()
+    assert torch.equal(out, nhwc)  # content preserved / correctly transposed

vllm/multimodal/video.py

@@ -637,6 +637,24 @@ class TorchCodecVideoBackendMixin:
         return batch.data.numpy(), list(frame_indices)
 
 
+def _pynvvc_frames_to_nhwc(frames):
+    """Return a stacked PyNvVideoCodec frame batch as contiguous NHWC.
+
+    PyNvVideoCodec's per-frame layout has varied across versions (HWC vs CHW),
+    so detect the channel axis rather than assuming a fixed order. NHWC is the
+    layout the other video backends return and the HF video processors expect.
+
+    Args:
+        frames: A ``(N, ?, ?, ?)`` uint8 tensor in either NHWC or NCHW order.
+
+    Returns:
+        The same frames as a contiguous ``(N, H, W, C)`` tensor.
+    """
+    if frames.shape[-1] != 3 and frames.shape[-3] == 3:
+        frames = frames.permute(0, 2, 3, 1)  # NCHW -> NHWC
+    return frames.contiguous()
+
+
 class PyNvVideoCodecVideoBackendMixin:
     """PyNvVideoCodec utilities for GPU-backed frame decode."""
 
@@ -807,7 +825,7 @@ class PyNvVideoCodecVideoBackendMixin:
                         "PyNvVideoCodec returned frames with unexpected shape "
                         f"{tuple(device_frames.shape)}"
                     )
-                device_frames = device_frames.permute(0, 3, 1, 2).contiguous()
+                device_frames = _pynvvc_frames_to_nhwc(device_frames)
                 host_frames = torch.empty(
                     device_frames.shape,
                     dtype=device_frames.dtype,