THUDM/cogvlm2-llama3-chinese-chat-19B

CogVLM2

👋 Wechat · 💡Online Demo · 🎈Github Page · 📑 Paper

📍Experience the larger-scale CogVLM model on the ZhipuAI Open Platform.

Model introduction

We launch a new generation of CogVLM2 series of models and open source two models built with Meta-Llama-3-8B-Instruct. Compared with the previous generation of CogVLM open source models, the CogVLM2 series of open source models have the following improvements:

1. Significant improvements in many benchmarks such as TextVQA, DocVQA.
2. Support 8K content length.
3. Support image resolution up to 1344 * 1344.
4. Provide an open source model version that supports both Chinese and English.

You can see the details of the CogVLM2 family of open source models in the table below:

Model name	cogvlm2-llama3-chat-19B	cogvlm2-llama3-chinese-chat-19B
Base Model	Meta-Llama-3-8B-Instruct	Meta-Llama-3-8B-Instruct
Language	English	Chinese, English
Model size	19B	19B
Task	Image understanding, dialogue model	Image understanding, dialogue model
Text length	8K	8K
Image resolution	1344 * 1344	1344 * 1344

Benchmark

Our open source models have achieved good results in many lists compared to the previous generation of CogVLM open source models. Its excellent performance can compete with some non-open source models, as shown in the table below:

Model	Open Source	LLM Size	TextVQA	DocVQA	ChartQA	OCRbench	VCR_EASY	VCR_HARD	MMMU	MMVet	MMBench
CogVLM1.1	✅	7B	69.7	-	68.3	590	73.9	34.6	37.3	52.0	65.8
LLaVA-1.5	✅	13B	61.3	-	-	337	-	-	37.0	35.4	67.7
Mini-Gemini	✅	34B	74.1	-	-	-	-	-	48.0	59.3	80.6
LLaVA-NeXT-LLaMA3	✅	8B	-	78.2	69.5	-	-	-	41.7	-	72.1
LLaVA-NeXT-110B	✅	110B	-	85.7	79.7	-	-	-	49.1	-	80.5
InternVL-1.5	✅	20B	80.6	90.9	83.8	720	14.7	2.0	46.8	55.4	82.3
QwenVL-Plus	❌	-	78.9	91.4	78.1	726	-	-	51.4	55.7	67.0
Claude3-Opus	❌	-	-	89.3	80.8	694	63.85	37.8	59.4	51.7	63.3
Gemini Pro 1.5	❌	-	73.5	86.5	81.3	-	62.73	28.1	58.5	-	-
GPT-4V	❌	-	78.0	88.4	78.5	656	52.04	25.8	56.8	67.7	75.0
CogVLM2-LLaMA3	✅	8B	84.2	92.3	81.0	756	83.3	38.0	44.3	60.4	80.5
CogVLM2-LLaMA3-Chinese	✅	8B	85.0	88.4	74.7	780	79.9	25.1	42.8	60.5	78.9

All reviews were obtained without using any external OCR tools ("pixel only").

Quick Start

here is a simple example of how to use the model to chat with the CogVLM2 model. For More use case. Find in our github

import torch
from PIL import Image
from transformers import AutoModelForCausalLM, AutoTokenizer

MODEL_PATH = "THUDM/cogvlm2-llama3-chinese-chat-19B"
DEVICE = 'cuda' if torch.cuda.is_available() else 'cpu'
TORCH_TYPE = torch.bfloat16 if torch.cuda.is_available() and torch.cuda.get_device_capability()[0] >= 8 else torch.float16

tokenizer = AutoTokenizer.from_pretrained(
    MODEL_PATH,
    trust_remote_code=True
)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_PATH,
    torch_dtype=TORCH_TYPE,
    trust_remote_code=True,
).to(DEVICE).eval()

text_only_template = "A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: {} ASSISTANT:"

while True:
    image_path = input("image path >>>>> ")
    if image_path == '':
        print('You did not enter image path, the following will be a plain text conversation.')
        image = None
        text_only_first_query = True
    else:
        image = Image.open(image_path).convert('RGB')

    history = []

    while True:
        query = input("Human:")
        if query == "clear":
            break

        if image is None:
            if text_only_first_query:
                query = text_only_template.format(query)
                text_only_first_query = False
            else:
                old_prompt = ''
                for _, (old_query, response) in enumerate(history):
                    old_prompt += old_query + " " + response + "\n"
                query = old_prompt + "USER: {} ASSISTANT:".format(query)
        if image is None:
            input_by_model = model.build_conversation_input_ids(
                tokenizer,
                query=query,
                history=history,
                template_version='chat'
            )
        else:
            input_by_model = model.build_conversation_input_ids(
                tokenizer,
                query=query,
                history=history,
                images=[image],
                template_version='chat'
            )
        inputs = {
            'input_ids': input_by_model['input_ids'].unsqueeze(0).to(DEVICE),
            'token_type_ids': input_by_model['token_type_ids'].unsqueeze(0).to(DEVICE),
            'attention_mask': input_by_model['attention_mask'].unsqueeze(0).to(DEVICE),
            'images': [[input_by_model['images'][0].to(DEVICE).to(TORCH_TYPE)]] if image is not None else None,
        }
        gen_kwargs = {
            "max_new_tokens": 2048,
            "pad_token_id": 128002,  
        }
        with torch.no_grad():
            outputs = model.generate(**inputs, **gen_kwargs)
            outputs = outputs[:, inputs['input_ids'].shape[1]:]
            response = tokenizer.decode(outputs[0])
            response = response.split("<|end_of_text|>")[0]
            print("\nCogVLM2:", response)
        history.append((query, response))

License

This model is released under the CogVLM2 LICENSE. For models built with Meta Llama 3, please also adhere to the LLAMA3_LICENSE.

Citation

If you find our work helpful, please consider citing the following papers

@misc{hong2024cogvlm2,
  title={CogVLM2: Visual Language Models for Image and Video Understanding},
  author={Hong, Wenyi and Wang, Weihan and Ding, Ming and Yu, Wenmeng and Lv, Qingsong and Wang, Yan and Cheng, Yean and Huang, Shiyu and Ji, Junhui and Xue, Zhao and others},
  year={2024}
  eprint={2408.16500},
  archivePrefix={arXiv},
  primaryClass={cs.CV}
}

@misc{wang2023cogvlm,
      title={CogVLM: Visual Expert for Pretrained Language Models}, 
      author={Weihan Wang and Qingsong Lv and Wenmeng Yu and Wenyi Hong and Ji Qi and Yan Wang and Junhui Ji and Zhuoyi Yang and Lei Zhao and Xixuan Song and Jiazheng Xu and Bin Xu and Juanzi Li and Yuxiao Dong and Ming Ding and Jie Tang},
      year={2023},
      eprint={2311.03079},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}