[Model] Reduce MiniMax-H3 Ref2VA RGB frame conversion

tests/diffusion/models/minimax_h3/test_ref2va_frame_conversion.py

@@ -0,0 +1,27 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
+import numpy as np
+import pytest
+import torch
+from PIL import Image
+
+from vllm_omni.model_executor.models.minimax_h3.encoder_processing import _frames_to_tensor
+
+pytestmark = [pytest.mark.core_model, pytest.mark.cpu, pytest.mark.diffusion]
+
+
+@pytest.mark.parametrize("reverse_rows", [False, True])
+def test_rgb_video_array_matches_framewise_conversion_and_owns_data(reverse_rows):
+    frames = np.arange(3 * 5 * 7 * 3, dtype=np.uint8).reshape(3, 5, 7, 3)
+    if reverse_rows:
+        frames = frames[:, ::-1]
+
+    expected = _frames_to_tensor([Image.fromarray(frame) for frame in frames])
+    actual = _frames_to_tensor(frames)
+
+    torch.testing.assert_close(actual, expected, rtol=0, atol=0)
+    assert actual.is_contiguous()
+    original_value = actual[0, 0, 0, 0].item()
+    frames[0, 0, 0, 0] = (original_value + 1) % 256
+    assert actual[0, 0, 0, 0].item() == original_value

vllm_omni/model_executor/models/minimax_h3/encoder_processing.py

@@ -258,6 +258,9 @@ def _image_to_tensor(image: Image.Image) -> torch.Tensor:
 def _frames_to_tensor(frames: Sequence[Any]) -> torch.Tensor:
     if len(frames) == 0:
         raise ValueError("MiniMax H3 reference video must contain frames")
+    if isinstance(frames, np.ndarray) and frames.dtype == np.uint8 and frames.ndim == 4 and frames.shape[-1] == 3:
+        # Keep the existing owned, contiguous output without converting each RGB frame through PIL.
+        return torch.from_numpy(np.array(frames, copy=True, order="C"))
     return torch.stack(
         [
             _image_to_tensor(frame if isinstance(frame, Image.Image) else Image.fromarray(np.asarray(frame)))