diff --git a/tests/models/multimodal/generation/test_transformers_video.py b/tests/models/multimodal/generation/test_transformers_video.py
index bbb31b7be8..c6c47f9ac0 100644
--- a/tests/models/multimodal/generation/test_transformers_video.py
+++ b/tests/models/multimodal/generation/test_transformers_video.py
@@ -36,6 +36,7 @@ VIDEO_MODEL_SETTINGS: dict[str, dict[str, Any]] = {
             "<|VISION_START|><|VIDEO_PAD|><|VISION_END|>Describe this video."
             "<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>"
         ),
+        "patch_hf_runner": model_utils.cohere_compass_patch_hf_runner,
     },
     "Qwen/Qwen2.5-VL-3B-Instruct": {
         "prompt": (
@@ -88,6 +89,9 @@ def test_transformers_video_generation(
         model_id, dtype="bfloat16", auto_cls=AutoModelForImageTextToText
     ) as hf_model:
         hf_model = model_utils.qwen3_vl_patch_hf_runner(hf_model)
+        patch_hf_runner = VIDEO_MODEL_SETTINGS[model_id].get("patch_hf_runner")
+        if patch_hf_runner is not None:
+            hf_model = patch_hf_runner(hf_model)
         hf_outputs = hf_model.generate_greedy_logprobs_limit(
             [prompt], 128, num_logprobs=10, videos=[video]
         )
diff --git a/tests/models/multimodal/generation/vlm_utils/model_utils.py b/tests/models/multimodal/generation/vlm_utils/model_utils.py
index bc472c400f..9acbe2c0d6 100644
--- a/tests/models/multimodal/generation/vlm_utils/model_utils.py
+++ b/tests/models/multimodal/generation/vlm_utils/model_utils.py
@@ -1111,6 +1111,19 @@ def qwen3_vl_patch_hf_runner(hf_model: HfRunner) -> HfRunner:
     return hf_model
 
 
+def cohere_compass_patch_hf_runner(hf_model: HfRunner) -> HfRunner:
+    """Declare the video modality that CohereCompass omits upstream.
+
+    Without it ``generate`` leaves ``pixel_values_videos`` unconsumed and the
+    forward pass rejects it, see huggingface/transformers#49320. Drop this once
+    the pinned transformers declares the modality itself.
+    """
+    model = hf_model.model
+    if "video" not in model.input_modalities:
+        model.input_modalities = (*model.input_modalities, "video")
+    return hf_model
+
+
 def voxtral_patch_hf_runner(hf_model: "HfRunner") -> "HfRunner":
     """Patch HfRunner for Voxtral's conversation-based processor.