Ovis1.6-9B视觉大模型环境搭建&推理

引子 前阵子，阿里Qwen2-VL刚刚闪亮登场，感兴趣的小伙伴可以移步https://blog.csdn.net/zzq1989_/article/details/142332651?spm=1001.2014.3001.5501。这第一的宝座还没坐多久，自家兄弟Ovis1.6版本就来了，20240919阿里国际AI团队开源多模态大模型Ovis1.6。在多模态权威综合评测基准OpenCompass上，Ovis1.6-Gemma2-9B版本综合得分超越Qwen2VL-7B、InternVL2-26B和MiniCPM-V-2.6等主流开源模型，在300亿以下参数开源模型中位居第一。 一、模型介绍 根据OpenCompass评测基准，Ovis1.6-Gemma2-9B超过了Qwen2-VL-7B、MiniCPM-V-2.6等一众相同参数量级的知名多模态模型。在数学等推理任务中，甚至有媲美70B参数模型的表现。Ovis1.6的幻觉现象和错误率也低于同级别模型，展现了更高的文本质量和准确率。阿里国际AI团队的核心思路是：从结构上对齐视觉和文本嵌入。当前，多数开源多模态大语言模型（MLLM）并非从头训练整个模型，而是通过像多层感知机（MLP）这样的连接器，将预训练的大语言模型（LLM）和视觉Transformer集成起来，给LLM装上“眼睛”。这样一来，就导致了一个问题：MLLM的文本和视觉模块采用不同的嵌入策略，使得视觉和文本信息没办法无缝融合，限制了模型性能的进一步提升。针对这个问题，Ovis采用了视觉tokenizer+视觉嵌入表+大语言模型的架构。 二、环境搭建 1、模型下载 https://huggingface.co/AIDC-AI/Ovis1.6-Gemma2-9B/tree/main 2、环境安装 docker run -it --rm --gpus=all -v /datas/work/zzq:/workspace pytorch/pytorch:2.2.2-cuda12.1-cudnn8-devel bash git clone https://github.com/AIDC-AI/Ovis.git cd /workspace/Ovis/Ovis-main pip install -r requirements.txt -i https://pypi.tuna.tsinghua.edu.cn/simple pip install -e . 三、推理测试 1、修改代码

from dataclasses import field, dataclass
from typing import Optional, Union, List

import torch
from PIL import Image

from ovis.model.modeling_ovis import Ovis
from ovis.util.constants import IMAGE_TOKEN


@dataclass
class RunnerArguments:
    model_path: str
    max_new_tokens: int = field(default=512)
    do_sample: bool = field(default=False)
    top_p: Optional[float] = field(default=None)
    top_k: Optional[int] = field(default=None)
    temperature: Optional[float] = field(default=None)
    max_partition: int = field(default=9)


class OvisRunner:
    def __init__(self, args: RunnerArguments):
        self.model_path = args.model_path
        # self.dtype = torch.bfloat16
        self.device = torch.cuda.current_device()
        # self.dtype = torch.bfloat16
        self.dtype = torch.float16
        self.model = Ovis.from_pretrained(self.model_path, torch_dtype=self.dtype, multimodal_max_length=8192)
        self.model = self.model.eval().to(device=self.device)
        self.eos_token_id = self.model.generation_config.eos_token_id
        self.text_tokenizer = self.model.get_text_tokenizer()
        self.pad_token_id = self.text_tokenizer.pad_token_id
        self.visual_tokenizer = self.model.get_visual_tokenizer()
        self.conversation_formatter = self.model.get_conversation_formatter()
        self.image_placeholder = IMAGE_TOKEN
        self.max_partition = args.max_partition
        self.gen_kwargs = dict(
            max_new_tokens=args.max_new_tokens,
            do_sample=args.do_sample,
            top_p=args.top_p,
            top_k=args.top_k,
            temperature=args.temperature,
            repetition_penalty=None,
            eos_token_id=self.eos_token_id,
            pad_token_id=self.pad_token_id,
            use_cache=True
        )

    def preprocess(self, inputs: List[Union[Image.Image, str]]):
        # for single image and single text inputs, ensure image ahead
        if len(inputs) == 2 and isinstance(inputs[0], str) and isinstance(inputs[1], Image.Image):
            inputs = reversed(inputs)

        # build query
        query = ''
        images = []
        for data in inputs:
            if isinstance(data, Image.Image):
                query += self.image_placeholder + '\n'
                images.append(data)
            elif isinstance(data, str):
                query += data.replace(self.image_placeholder, '')
            elif data is not None:
                raise RuntimeError(f'Invalid input type, expected `PIL.Image.Image` or `str`, but got {type(data)}')

        # format conversation
        prompt, input_ids, pixel_values = self.model.preprocess_inputs(
            query, images, max_partition=self.max_partition)
        attention_mask = torch.ne(input_ids, self.text_tokenizer.pad_token_id)
        input_ids = input_ids.unsqueeze(0).to(device=self.device)
        attention_mask = attention_mask.unsqueeze(0).to(device=self.device)
        if pixel_values is not None:
            pixel_values = [pixel_values.to(device=self.device, dtype=self.dtype)]
        else:
            pixel_values = [None]

        return prompt, input_ids, attention_mask, pixel_values

    def run(self, inputs: List[Union[Image.Image, str]]):
        prompt, input_ids, attention_mask, pixel_values = self.preprocess(inputs)
        output_ids = self.model.generate(
            input_ids,
            pixel_values=pixel_values,
            attention_mask=attention_mask,
            **self.gen_kwargs
        )
        output = self.text_tokenizer.decode(output_ids[0], skip_special_tokens=True)
        input_token_len = input_ids.shape[1]
        output_token_len = output_ids.shape[1]
        response = dict(
            prompt=prompt,
            output=output,
            prompt_tokens=input_token_len,
            total_tokens=input_token_len + output_token_len
        )
        return response


if __name__ == '__main__':
    # runner_args = RunnerArguments(model_path='<model_path>')
    runner_args = RunnerArguments(model_path='/workspace/Ovis/Ovis-main/models')
    runner = OvisRunner(runner_args)
    # image = Image.open('<image_path>')
    image = Image.open('/workspace/Ovis/Ovis-main/test.png')
    # text = '<prompt>'
    text = 'solve the question in this image'
    response = runner.run([image, text])
    print(response['output'])