app.py

import torch
from transformers import Qwen2VLForConditionalGeneration, AutoProcessor
from PIL import Image
import requests
from io import BytesIO

# Initialize the model and processor
model_name = "Qwen/Qwen2-VL-2B-Instruct"
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

model = Qwen2VLForConditionalGeneration.from_pretrained(model_name).to(device)
processor = AutoProcessor.from_pretrained(model_name)

# Load the image from URL
image_url = "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg"
response = requests.get(image_url)
img = Image.open(BytesIO(response.content))

# Automatically preprocess the image and text input using the processor
text_input = "Describe this image."

# The processor automatically handles resizing, normalization, and tokenization
inputs = processor(
    images=img,
    text=text_input,
    return_tensors="pt",
    padding=True,  # Automatically pad to match model input size
)

# Check the number of tokens generated by the processor and the shape of inputs
print("Input tokens:", inputs.input_ids.shape)
print("Image features shape:", inputs.pixel_values.shape)

# Ensure image and text are properly tokenized and features align
assert inputs.input_ids.shape[1] > 0, "No tokens generated for text input!"
assert inputs.pixel_values.shape[0] > 0, "No features generated for the image!"

# Move inputs to the device (either GPU or CPU)
inputs = {key: value.to(device) for key, value in inputs.items()}

# Inference
generated_ids = model.generate(**inputs, max_new_tokens=128)

# Decode the output
output_text = processor.batch_decode(generated_ids, skip_special_tokens=True)
print(output_text)