With Qwen 2.5 VL sample code was provided in the readme for processing videos client side to send only the sampled frames encoded in b64 with the video/jpeg type in the request when using Qwen 2.5 VL through vLLM.
import base64
import numpy as np
from PIL import Image
from io import BytesIO
from openai import OpenAI
from qwen_vl_utils import process_vision_info
openai_api_key = "EMPTY"
openai_api_base = "http://localhost:8000/v1"
client = OpenAI(
api_key=openai_api_key,
base_url=openai_api_base,
)
video_messages = [
{"role": "system", "content": "You are a helpful assistant."},
{"role": "user", "content": [
{"type": "text", "text": "请用表格总结一下视频中的商品特点"},
{
"type": "video",
"video": "https://duguang-labelling.oss-cn-shanghai.aliyuncs.com/qiansun/video_ocr/videos/50221078283.mp4",
"total_pixels": 20480 * 28 * 28, "min_pixels": 16 * 28 * 2,
'fps': 3.0 # The default value is 2.0, but for demonstration purposes, we set it to 3.0.
}]
},
]
def prepare_message_for_vllm(content_messages):
"""
The frame extraction logic for videos in `vLLM` differs from that of `qwen_vl_utils`.
Here, we utilize `qwen_vl_utils` to extract video frames, with the `media_typ`e of the video explicitly set to `video/jpeg`.
By doing so, vLLM will no longer attempt to extract frames from the input base64-encoded images.
"""
vllm_messages, fps_list = [], []
for message in content_messages:
message_content_list = message["content"]
if not isinstance(message_content_list, list):
vllm_messages.append(message)
continue
new_content_list = []
for part_message in message_content_list:
if 'video' in part_message:
video_message = [{'content': [part_message]}]
image_inputs, video_inputs, video_kwargs = process_vision_info(video_message, return_video_kwargs=True)
assert video_inputs is not None, "video_inputs should not be None"
video_input = (video_inputs.pop()).permute(0, 2, 3, 1).numpy().astype(np.uint8)
fps_list.extend(video_kwargs.get('fps', []))
# encode image with base64
base64_frames = []
for frame in video_input:
img = Image.fromarray(frame)
output_buffer = BytesIO()
img.save(output_buffer, format="jpeg")
byte_data = output_buffer.getvalue()
base64_str = base64.b64encode(byte_data).decode("utf-8")
base64_frames.append(base64_str)
part_message = {
"type": "video_url",
"video_url": {"url": f"data:video/jpeg;base64,{','.join(base64_frames)}"}
}
new_content_list.append(part_message)
message["content"] = new_content_list
vllm_messages.append(message)
return vllm_messages, {'fps': fps_list}
video_messages, video_kwargs = prepare_message_for_vllm(video_messages)
chat_response = client.chat.completions.create(
model="Qwen/Qwen2.5-VL-7B-Instruct",
messages=video_messages,
extra_body={
"mm_processor_kwargs": video_kwargs
}
)
print("Chat response:", chat_response)
Is this still a recommended or supported way to pass preprocessed videos to Qwen 3 VL when being served through vLLM?
Hello,
With Qwen 2.5 VL sample code was provided in the readme for processing videos client side to send only the sampled frames encoded in b64 with the
video/jpegtype in the request when using Qwen 2.5 VL through vLLM.https://github.com/QwenLM/Qwen3-VL/tree/2f25a646fb0f329647428eb8dacf19293de6f5d4?tab=readme-ov-file#start-an-openai-api-service
Is this still a recommended or supported way to pass preprocessed videos to Qwen 3 VL when being served through vLLM?