import torch
from transformers import AutoModel, AutoTokenizer
from utils import load_image, load_video

if __name__ == "__main__":

    dir, rev = 'morpheushoc/InternVL2_5-2B', 'main'

    # path = 'OpenGVLab/InternVL2_5-2B'
    model = AutoModel.from_pretrained(dir,
                                      torch_dtype=torch.bfloat16,
                                      load_in_8bit=False,
                                      low_cpu_mem_usage=True,
                                      use_flash_attn=True,
                                      trust_remote_code=True,
                                      revision=rev).eval().cuda()
    tokenizer = AutoTokenizer.from_pretrained(dir, trust_remote_code=True, use_fast=False)
    generation_config = dict(max_new_tokens=1024, do_sample=False)

    paths = [
        'image1.jpg',
        'image1.jpg',
        'image2.jpg',
        'red-panda.mp4',
    ]

    questions = [
        'describe this image',
        'describe this image',
        'describe this image',
        'describe this video'
    ]

    pixel_values, num_patches_list = [], []
    for i, fp in enumerate(paths):

        if fp.endswith('mp4'):
            pxl_val, num_patches = load_video(fp, num_segments=8, max_num=1)
            prefix = ''.join([f'Frame{i+1}: <image>\n' for i in range(len(num_patches))])
        else:
            pxl_val = load_image(fp, max_num=12).to(torch.bfloat16)
            num_patches = [len(pxl_val)]
            prefix = '<image>\n'
        
        pixel_values.append(pxl_val)
        num_patches_list.append(num_patches)
        questions[i] = prefix + questions[i]
    
    pixel_values = torch.cat(pixel_values).to(torch.bfloat16).cuda()
    response = model.batch_chat(tokenizer, pixel_values, questions, generation_config,
                                num_patches_list=num_patches_list, history=None, return_history=False)
    
    for q, r in zip(questions, response):
        print(f'User: {q}\nAssistant: {r}')
        print('\n')