xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-10 04:34:57 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Bugfix] Fix qwen3 vl dummy data generation with overrides (#26193)

Browse Source 
Signed-off-by: Roger Wang <hey@rogerw.io>
This commit is contained in:
Roger Wang 2025-10-03 18:40:20 -07:00 committed by GitHub
parent 5a05f26603
commit 67bc0c003e
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 43 additions and 20 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

63
vllm/model_executor/models/qwen3_vl.py
View File

@ -47,7 +47,7 @@ from vllm.attention.backends.registry import _Backend
from vllm.attention.layer import check_upstream_fa_availability
from vllm.compilation.decorators import support_torch_compile
from vllm.config import VllmConfig
from vllm.config.multimodal import BaseDummyOptions
from vllm.config.multimodal import BaseDummyOptions, VideoDummyOptions
from vllm.distributed import get_pp_group
from vllm.logger import init_logger
from vllm.model_executor.layers.activation import _ACTIVATION_REGISTRY
@ -741,20 +741,57 @@ class Qwen3VLDummyInputsBuilder(BaseDummyInputsBuilder[Qwen3VLProcessingInfo]):
) -> MultiModalDataDict:
num_images = mm_counts.get("image", 0)
num_videos = mm_counts.get("video", 0)
image_overrides = mm_options.get("image") if mm_options else None
video_overrides = mm_options.get("video") if mm_options else None
target_width, target_height = (
self.info.get_image_size_with_most_features())
target_num_frames = self.info.get_num_frames_with_most_features(
seq_len, mm_counts)
if video_overrides:
assert isinstance(video_overrides, VideoDummyOptions)
num_frames_override = video_overrides.num_frames
if num_frames_override:
if num_frames_override > target_num_frames:
logger.warning(
"video.num_frames override (%d) exceeds model's "
"maximum number of frames (%d), will be ignored",
num_frames_override, target_num_frames)
if num_frames_override < 2:
logger.warning(
"video.num_frames override (%d) cannot be less "
"than 2, will be ignored", num_frames_override)
target_num_frames = min(target_num_frames, num_frames_override)
target_num_frames = max(target_num_frames, 2)
target_video_size, _ = self.info._get_vision_info(
image_width=target_width,
image_height=target_height,
num_frames=target_num_frames,
image_processor=self.info.get_video_processor(),
)
image_overrides = mm_options.get("image") if mm_options else None
video_overrides = mm_options.get("video") if mm_options else None
# NOTE: we need to do this check here since Qwen3-VL resizes video
# frames depending on how many frames there are.
width, height = target_video_size.width, target_video_size.height
if video_overrides:
assert isinstance(video_overrides, VideoDummyOptions)
width_override = video_overrides.width
if width_override:
if width_override > width:
logger.warning(
"video.width override (%d) exceeds model's "
"maximum width (%d), will be ignored", width_override,
width)
width = min(width, width_override)
height_override = video_overrides.height
if height_override:
if height_override > height:
logger.warning(
"video.height override (%d) exceeds model's "
"maximum height (%d), will be ignored",
height_override, height)
height = min(height, height_override)
return {
"image":
@ -764,11 +801,10 @@ class Qwen3VLDummyInputsBuilder(BaseDummyInputsBuilder[Qwen3VLProcessingInfo]):
overrides=image_overrides),
"video":
self._get_dummy_videos(
width=target_video_size.width,
height=target_video_size.height,
width=width,
height=height,
num_frames=target_num_frames,
num_videos=num_videos,
overrides=video_overrides,
),
}
@ -780,7 +816,6 @@ class Qwen3VLDummyInputsBuilder(BaseDummyInputsBuilder[Qwen3VLProcessingInfo]):
num_frames: int,
num_videos: int,
) -> list[VideoItem]:
num_frames = max(num_frames, 2)
video = np.full((num_frames, width, height, 3), 255, dtype=np.uint8)
video_items = []
for i in range(num_videos):
@ -796,18 +831,6 @@ class Qwen3VLDummyInputsBuilder(BaseDummyInputsBuilder[Qwen3VLProcessingInfo]):
video_items.append(video_item)
return video_items
def get_dummy_processor_inputs(self, seq_len, mm_counts):
processor_inputs = super().get_dummy_processor_inputs(
seq_len, mm_counts)
# HACK(Isotr0py): We set do_resize to False here to reuse Qwen2-VL's
# profiling logic, which will be problematic for configurable mm
# profiling.
# TODO(Isotr0py): Switch to the implementation in
# https://github.com/vllm-project/vllm/pull/25557
# after supporting configurable mm profiling.
processor_inputs.hf_processor_mm_kwargs = {"do_resize": False}
return processor_inputs
class Qwen3VLMultiModalProcessor(BaseMultiModalProcessor[Qwen3VLProcessingInfo]
):