diff --git a/tests/models/multimodal/processing/test_molmo2.py b/tests/models/multimodal/processing/test_molmo2.py index c12f70c276..5633d47a97 100644 --- a/tests/models/multimodal/processing/test_molmo2.py +++ b/tests/models/multimodal/processing/test_molmo2.py @@ -3,9 +3,14 @@ from types import SimpleNamespace +import pytest import torch -from vllm.model_executor.models.molmo2 import build_flat_image_bool_length +from vllm.config.multimodal import VideoDummyOptions +from vllm.model_executor.models.molmo2 import ( + Molmo2DummyInputsBuilder, + build_flat_image_bool_length, +) def test_build_flat_image_bool_length_matches_molmoweb_processor_tokens(): @@ -54,3 +59,32 @@ def test_build_flat_image_bool_length_respects_disabled_col_tokens(): assert len(image_tokens) == 45 assert image_tokens[0].item() == hf_config.low_res_image_start_token_id assert (image_tokens == hf_config.image_col_id).sum().item() == 0 + + +@pytest.mark.parametrize( + ("num_frames_override", "expected_frames"), + [(1, 2), (2, 2), (3, 3)], +) +def test_dummy_video_num_frames_override_honors_min_of_two( + num_frames_override, expected_frames +): + """A ``num_frames`` override below 2 must be ignored (the model needs at + least 2 frames), matching the "cannot be less than 2, will be ignored" + warning.""" + builder = object.__new__(Molmo2DummyInputsBuilder) + builder.info = SimpleNamespace( + get_hf_processor=lambda: SimpleNamespace( + video_processor=SimpleNamespace(size={"width": 64, "height": 64}) + ), + get_num_frames_with_most_features=lambda seq_len, mm_counts: 16, + get_image_size_with_most_features=lambda: (64, 64), + ) + + data = builder.get_dummy_mm_data( + seq_len=128, + mm_counts={"image": 0, "video": 1}, + mm_options={"video": VideoDummyOptions(num_frames=num_frames_override)}, + ) + + video, _metadata = data["video"][0] + assert video.shape[0] == expected_frames diff --git a/vllm/model_executor/models/molmo2.py b/vllm/model_executor/models/molmo2.py index d73160d689..0306529fe3 100644 --- a/vllm/model_executor/models/molmo2.py +++ b/vllm/model_executor/models/molmo2.py @@ -1879,6 +1879,7 @@ class Molmo2DummyInputsBuilder(BaseDummyInputsBuilder[Molmo2ProcessingInfo]): num_frames_override, ) target_num_frames = min(target_num_frames, num_frames_override) + target_num_frames = max(target_num_frames, 2) dummy_videos = self._get_dummy_videos( width=video_size["width"],