biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Model] Support Qwen3-VL Model Series (#24727)

Browse Source 
Signed-off-by: Roger Wang <hey@rogerw.io>
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Co-authored-by: Huang Jie <92386084+JJJYmmm@users.noreply.github.com>
Co-authored-by: 松灵 <26085463+wulipc@users.noreply.github.com>
Co-authored-by: Isotr0py <mozf@mail2.sysu.edu.cn>
This commit is contained in:
Roger Wang
2025-09-16 22:01:04 -07:00
committed by GitHub
parent 5801e49776
commit 0f7acdd73c
13 changed files with 2084 additions and 17 deletions
Download Patch File Download Diff File Expand all files Collapse all files

35
tests/models/multimodal/processing/test_common.py
View File

@@ -31,6 +31,7 @@ def glm4_1v_patch_mm_data(mm_data: MultiModalDataDict) -> MultiModalDataDict:
"""
# Ensure video metadata is included
if "video" in mm_data:
# GLM4.1V doesn't support multiple videos
video = mm_data["video"]
num_frames = len(video)
mm_data["video"] = (video, {
@@ -44,6 +45,34 @@ def glm4_1v_patch_mm_data(mm_data: MultiModalDataDict) -> MultiModalDataDict:
return mm_data
def qwen3_vl_patch_mm_data(mm_data: MultiModalDataDict) -> MultiModalDataDict:
"""
Patch the multimodal data for Qwen3-VL model.
"""
def create_metadata(frames: np.ndarray):
num_frames = len(frames)
return {
"total_num_frames": num_frames,
"fps": 2.0,
"duration": num_frames / 2.0,
"video_backend": "opencv",
"frames_indices": list(range(num_frames)),
"do_sample_frames": True,
}
# Ensure video metadata is included
if "video" in mm_data:
video = mm_data["video"]
if isinstance(video, list):
# multiple videos
mm_data["video"] = [(vid, create_metadata(vid)) for vid in video]
else:
# single video
mm_data["video"] = (video, create_metadata(video))
return mm_data
def _test_processing_correctness(
model_id_or_arch: str,
hit_rate: float,
@@ -182,8 +211,10 @@ _IGNORE_MM_KEYS = {
}
MM_DATA_PATCHES = {
# GLM4.1V requires video metadata to be included in the input
# GLM4.1V and Qwen3-VL requires video metadata to be included in the input
"glm4v": glm4_1v_patch_mm_data,
"qwen3_vl": qwen3_vl_patch_mm_data,
"qwen3_vl_moe": qwen3_vl_patch_mm_data,
}
@@ -326,6 +357,8 @@ def _test_processing_correctness_one(
"Qwen/Qwen2.5-VL-3B-Instruct",
"Qwen/Qwen2-Audio-7B-Instruct",
"Qwen/Qwen2.5-Omni-3B",
"Qwen/Qwen3-VL-4B-Instruct",
"Qwen/Qwen3-VL-30B-A3B-Instruct",
"YannQi/R-4B",
"Skywork/Skywork-R1V-38B",
"HuggingFaceTB/SmolVLM2-2.2B-Instruct",