from PIL import Image
import requests
import torch
from torchvision import io
from typing import Dict
from transformers import Qwen2VLForConditionalGeneration, AutoTokenizer, AutoProcessor

model_id = "yujiepan/qwen2-vl-tiny-random"

# Load the model in half-precision on the available device(s)
model = Qwen2VLForConditionalGeneration.from_pretrained(
    model_id, torch_dtype="auto", device_map="auto"
)
processor = AutoProcessor.from_pretrained(model_id)

# Image
url = "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg"
image = Image.open(requests.get(url, stream=True).raw)
conversation = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
            },
            {"type": "text", "text": "Describe this image."},
        ],
    }
]
text_prompt = processor.apply_chat_template(conversation, add_generation_prompt=True)
# Excepted output: '<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n<|vision_start|><|image_pad|><|vision_end|>Describe this image.<|im_end|>\n<|im_start|>assistant\n'

inputs = processor(
    text=[text_prompt], images=[image], padding=True, return_tensors="pt"
)
inputs = inputs.to("cuda")

output_ids = model.generate(**inputs, max_new_tokens=128)
generated_ids = [
    output_ids[len(input_ids) :]
    for input_ids, output_ids in zip(inputs.input_ids, output_ids)
]
output_text = processor.batch_decode(
    generated_ids, skip_special_tokens=True, clean_up_tokenization_spaces=True
)
print(output_text)

Codes:

python

import os
from typing import Dict

import requests
import torch
import transformers
from PIL import Image
from torchvision import io
from transformers import (AutoConfig, AutoModelForCausalLM, AutoProcessor,
                          AutoTokenizer, GenerationConfig, pipeline, set_seed)
from transformers.models.qwen2_vl import Qwen2VLForConditionalGeneration

model_id = "Qwen/Qwen2-VL-7B-Instruct"
repo_id = "yujiepan/qwen2-vl-tiny-random"
save_path = f"/tmp/{repo_id}"

config = AutoConfig.from_pretrained(model_id, trust_remote_code=True)
config.hidden_size = 16
config.intermediate_size = 32
config.num_attention_heads = 2
config.num_hidden_layers = 2
config.num_key_value_heads = 1
config.vision_config.embed_dim = 16
config.vision_config.num_heads = 2
config.vision_config.hidden_size = 16
config.vision_config.depth = 2
config.rope_scaling['mrope_section'] = [1, 1, 2]  # sum needs to be 4 here

model = Qwen2VLForConditionalGeneration(config=config)
model = model.to(torch.bfloat16).cuda().eval()
model.generation_config = GenerationConfig.from_pretrained(
    model_id, trust_remote_code=True,
)
set_seed(42)
with torch.no_grad():
    for _, p in sorted(model.named_parameters()):
        torch.nn.init.uniform_(p, -0.3, 0.3)

processor = AutoProcessor.from_pretrained(model_id)
model.save_pretrained(save_path)
processor.save_pretrained(save_path)
os.system(f"ls -alh {save_path}")


def try_inference():
    url = "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg"
    image = Image.open(requests.get(url, stream=True).raw)
    conversation = [
        {
            "role": "user",
            "content": [
                {
                    "type": "image",
                },
                {"type": "text", "text": "Describe this image."},
            ],
        }
    ]
    processor = AutoProcessor.from_pretrained(save_path)
    model = Qwen2VLForConditionalGeneration.from_pretrained(
        save_path, torch_dtype=torch.bfloat16, device_map='cuda')
    text_prompt = processor.apply_chat_template(
        conversation, add_generation_prompt=True)
    inputs = processor(
        text=[text_prompt], images=[image], padding=True, return_tensors="pt"
    )
    inputs = inputs.to("cuda")
    output_ids = model.generate(**inputs, max_new_tokens=16)
    generated_ids = [
        output_ids[len(input_ids):]
        for input_ids, output_ids in zip(inputs.input_ids, output_ids)
    ]
    output_text = processor.batch_decode(
        generated_ids, skip_special_tokens=True, clean_up_tokenization_spaces=True
    )
    print(output_text)


try_inference()