[Bugfix][GLM4V] Fix video dummy profiling and memory usage (#48729)

Signed-off-by: xiaoming <1259730330@qq.com>
This commit is contained in:
labAxiaoming
2026-07-18 01:44:45 +08:00
committed by GitHub
parent efed8a1e83
commit bf578e1abd
2 changed files with 72 additions and 49 deletions
@@ -1,9 +1,15 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from unittest.mock import Mock
import pytest
from vllm.assets.video import VideoAsset
from vllm.model_executor.models.glm4_1v import (
Glm4vForConditionalGeneration,
Glm4vProcessingInfo,
)
from vllm.multimodal import MULTIMODAL_REGISTRY
from vllm.multimodal.inputs import batched_tensors_equal
from vllm.multimodal.video import DynamicVideoBackend, VideoBackend
@@ -11,6 +17,52 @@ from vllm.multimodal.video import DynamicVideoBackend, VideoBackend
from ...utils import build_model_context
@pytest.mark.parametrize(
(
"max_video_pixels",
"max_tokens",
"expected_num_frames",
),
[
(47_040_000, 124_988, 11),
(47_040_000, 30_000, 24),
(100_352_000, 124_988, 21),
(100_352_000, 30_000, 7),
(100_352_000, 0, 1),
],
)
def test_get_max_video_frames_matches_glm_resize(
max_video_pixels: int,
max_tokens: int,
expected_num_frames: int,
):
info = Mock(spec=Glm4vProcessingInfo)
info.get_image_size_with_most_features.return_value = (2184, 2184)
info._get_video_max_pixels.return_value = max_video_pixels
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 = 2
info._get_vision_info.side_effect = lambda **kwargs: (
Glm4vProcessingInfo._get_vision_info(info, **kwargs)
)
num_frames = Glm4vProcessingInfo._get_max_video_frames(
info,
max_tokens=max_tokens,
)
assert num_frames == expected_num_frames
assert info._get_video_max_pixels.call_count == 1
assert info._get_vision_info.call_count == 600
def test_encoder_cudagraph_uses_model_video_frame_limit():
model = Mock()
assert Glm4vForConditionalGeneration.get_max_frames_per_video(model) == 600
@pytest.mark.parametrize("model_id", ["zai-org/GLM-4.1V-9B-Thinking"])
@pytest.mark.parametrize("expected_toks_per_frame", [299])
@pytest.mark.parametrize(