diff --git a/tests/models/multimodal/processing/test_glm4_1v.py b/tests/models/multimodal/processing/test_glm4_1v.py index 8a77783282..e45e741c5b 100644 --- a/tests/models/multimodal/processing/test_glm4_1v.py +++ b/tests/models/multimodal/processing/test_glm4_1v.py @@ -63,6 +63,31 @@ def test_encoder_cudagraph_uses_model_video_frame_limit(): assert Glm4vForConditionalGeneration.get_max_frames_per_video(model) == 600 +@pytest.mark.parametrize( + ("temporal_patch_size", "expected_grid_t"), + [(2, 9), (4, 5), (8, 3)], +) +def test_vision_info_rounds_up_temporal_frames( + temporal_patch_size: int, + expected_grid_t: int, +): + info = Mock(spec=Glm4vProcessingInfo) + vision_config = info.get_hf_config.return_value.vision_config + vision_config.patch_size = 14 + vision_config.spatial_merge_size = 2 + vision_config.temporal_patch_size = temporal_patch_size + + _, num_vision_tokens = Glm4vProcessingInfo._get_vision_info( + info, + image_width=28, + image_height=28, + num_frames=17, + do_resize=False, + ) + + assert num_vision_tokens == expected_grid_t + + @pytest.mark.parametrize("model_id", ["zai-org/GLM-4.1V-9B-Thinking"]) @pytest.mark.parametrize("expected_toks_per_frame", [299]) @pytest.mark.parametrize( diff --git a/vllm/model_executor/models/glm4_1v.py b/vllm/model_executor/models/glm4_1v.py index 787e53e8df..19389d8446 100644 --- a/vllm/model_executor/models/glm4_1v.py +++ b/vllm/model_executor/models/glm4_1v.py @@ -1082,8 +1082,8 @@ class Glm4vProcessingInfo(BaseProcessingInfo): preprocessed_size = ImageSize(width=image_width, height=image_height) # NOTE: Frames are padded to be divisible by `temporal_patch_size` - # https://github.com/huggingface/transformers/blob/v4.48.3/src/transformers/models/qwen2_vl/image_processing_qwen2_vl.py#L294 - padded_num_frames = num_frames + num_frames % temporal_patch_size + # https://github.com/huggingface/transformers/blob/v5.13.0/src/transformers/models/qwen2_vl/video_processing_qwen2_vl.py#L249-L252 + padded_num_frames = num_frames + (-num_frames % temporal_patch_size) grid_t = max(padded_num_frames // temporal_patch_size, 1) grid_h = preprocessed_size.height // patch_size diff --git a/vllm/model_executor/models/kanana_v.py b/vllm/model_executor/models/kanana_v.py index 125d7e71c7..b1a5f78b1b 100644 --- a/vllm/model_executor/models/kanana_v.py +++ b/vllm/model_executor/models/kanana_v.py @@ -409,8 +409,8 @@ class KananaVProcessingInfo(BaseProcessingInfo): preprocessed_size = ImageSize(width=image_width, height=image_height) # NOTE: Frames are padded to be divisible by `temporal_patch_size` - # https://github.com/huggingface/transformers/blob/v4.48.3/src/transformers/models/qwen2_vl/image_processing_qwen2_vl.py#L294 - padded_num_frames = num_frames + num_frames % temporal_patch_size + # https://github.com/huggingface/transformers/blob/v5.13.0/src/transformers/models/qwen2_vl/video_processing_qwen2_vl.py#L249-L252 + padded_num_frames = num_frames + (-num_frames % temporal_patch_size) grid_t = max(padded_num_frames // temporal_patch_size, 1) grid_h = preprocessed_size.height // patch_size diff --git a/vllm/model_executor/models/keye.py b/vllm/model_executor/models/keye.py index dd1fb892ad..c3d69836a7 100644 --- a/vllm/model_executor/models/keye.py +++ b/vllm/model_executor/models/keye.py @@ -983,7 +983,7 @@ class KeyeProcessingInfo(BaseProcessingInfo): else: preprocessed_size = ImageSize(width=image_width, height=image_height) - padded_num_frames = num_frames + num_frames % temporal_patch_size + padded_num_frames = num_frames + (-num_frames % temporal_patch_size) grid_t = max(padded_num_frames // temporal_patch_size, 1) grid_h = preprocessed_size.height // patch_size diff --git a/vllm/model_executor/models/llava_onevision2.py b/vllm/model_executor/models/llava_onevision2.py index 58179ec00d..9dc11d493a 100644 --- a/vllm/model_executor/models/llava_onevision2.py +++ b/vllm/model_executor/models/llava_onevision2.py @@ -1345,7 +1345,7 @@ class LlavaOnevision2ProcessingInfo(BaseProcessingInfo): preprocessed = ImageSize(width=rw, height=rh) else: preprocessed = ImageSize(width=image_width, height=image_height) - padded_frames = num_frames + num_frames % temporal_patch_size + padded_frames = num_frames + (-num_frames % temporal_patch_size) grid_t = max(padded_frames // temporal_patch_size, 1) grid_h = preprocessed.height // patch_size grid_w = preprocessed.width // patch_size diff --git a/vllm/model_executor/models/mimo_v2_omni.py b/vllm/model_executor/models/mimo_v2_omni.py index d0d9589ae1..747cb0e88b 100644 --- a/vllm/model_executor/models/mimo_v2_omni.py +++ b/vllm/model_executor/models/mimo_v2_omni.py @@ -715,7 +715,7 @@ class MiMoV2OmniProcessingInfo(BaseProcessingInfo): effective_frames = num_frames * tokens_per_second else: effective_frames = num_frames - padded_num_frames = effective_frames + effective_frames % temporal_patch_size + padded_num_frames = effective_frames + (-effective_frames % temporal_patch_size) grid_t = max(padded_num_frames // temporal_patch_size, 1) grid_h = preprocessed_size.height // patch_size grid_w = preprocessed_size.width // patch_size diff --git a/vllm/model_executor/models/qwen2_vl.py b/vllm/model_executor/models/qwen2_vl.py index 539f141cba..e2e9f24524 100644 --- a/vllm/model_executor/models/qwen2_vl.py +++ b/vllm/model_executor/models/qwen2_vl.py @@ -898,8 +898,8 @@ class Qwen2VLProcessingInfo(BaseProcessingInfo): preprocessed_size = ImageSize(width=image_width, height=image_height) # NOTE: Frames are padded to be divisible by `temporal_patch_size` - # https://github.com/huggingface/transformers/blob/v4.48.3/src/transformers/models/qwen2_vl/image_processing_qwen2_vl.py#L294 - padded_num_frames = num_frames + num_frames % temporal_patch_size + # https://github.com/huggingface/transformers/blob/v5.13.0/src/transformers/models/qwen2_vl/video_processing_qwen2_vl.py#L249-L252 + padded_num_frames = num_frames + (-num_frames % temporal_patch_size) grid_t = max(padded_num_frames // temporal_patch_size, 1) grid_h = preprocessed_size.height // patch_size