qwen2_vl_multimodal_image_inference_with_qwen_vl_utils.py

1from transformers import Qwen2VLForConditionalGeneration, AutoTokenizer, AutoProcessor
2from qwen_vl_utils import process_vision_info
3
4# default: Load the model on the available device(s)
5model = Qwen2VLForConditionalGeneration.from_pretrained(
6    "Qwen/Qwen2-VL-7B-Instruct", torch_dtype="auto", device_map="auto"
7)
8
9# default processer
10processor = AutoProcessor.from_pretrained("Qwen/Qwen2-VL-7B-Instruct")
11
12# The messages include a local file path and a remote URL for the image
13messages = [
14    {
15        "role": "user",
16        "content": [
17            {"type": "image", "image": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg"},
18            {"type": "text", "text": "Describe this image."},
19        ],
20    }
21]
22
23# Preparation for inference
24text = processor.apply_chat_template(
25    messages, tokenize=False, add_generation_prompt=True
26)
27image_inputs, video_inputs = process_vision_info(messages)
28inputs = processor(
29    text=[text],
30    images=image_inputs,
31    videos=video_inputs,
32    padding=True,
33    return_tensors="pt",
34)
35inputs = inputs.to("cuda")
36
37# Inference: Generation of the output
38generated_ids = model.generate(**inputs, max_new_tokens=128)
39generated_ids_trimmed = [
40    out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
41]
42output_text = processor.batch_decode(
43    generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
44)
45print(output_text)