Views
No views yet
1from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor
2from qwen_vl_utils import process_vision_info
3
4model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
5 "/path/to/model",
6 torch_dtype="auto",
7 device_map="auto"
8)
9processor = AutoProcessor.from_pretrained("/path/to/model")
10
11messages = [
12 {
13 "role": "user",
14 "content": [
15 {
16 "type": "video",
17 "video": "/path/to/video.mp4",
18 },
19 {"type": "text", "text": "Describe this video in detail."},
20 ],
21 }
22]
23
24text = processor.apply_chat_template(
25 messages, tokenize=False, add_generation_prompt=True
26)
27image_inputs, video_inputs, video_kwargs = process_vision_info(messages, return_video_kwargs=True)
28inputs = processor(
29 text=[text],
30 images=image_inputs,
31 videos=video_inputs,
32 padding=True,
33 return_tensors="pt",
34 **video_kwargs,
35)
36inputs = inputs.to("cuda")
37
38generated_ids = model.generate(**inputs, max_new_tokens=512)
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[0])