Ollama 0.4 版本发布,支持 Llama 3.2 Vision 视觉模型,实测满意个人使用 ...

打印 上一主题 下一主题

主题 866|帖子 866|积分 2598


  更新信息

  最大的更新莫过于对 Llama 3.2 Vision 视觉模型的支持!ollama 的版本也因此从v0.3.14直接升级到v0.4.0。
  升级 ollama 后,运行如下命令体验
  1. # 默认运行 11B 大小模型
  2. ollama run llama3.2-vision
  3. # 运行 90B 大小模型,需要显存支持
  4. ollama run llama3.2-vision:90b
复制代码


  • 支持 Llama 3.2 Vision(即 Llama)架构,即 Llama 最新的视觉模型
  • 向视觉模型发送后续请求将显著加快速率
  • 修复了停止序列无法正确检测的问题
  • 现在可以在运行 ollama create my-model 时从 Safetensors 导入模型,而不需要 Modelfile 文件
  • 办理了在 Windows 下重定向输出到文件会导致写入无效字符的问题
  • 办理了导致 Ollama 堕落的不正当模型数据问题
  图片识别 Demo

  
  1. >>> /Users/geek/Pictures/ollama-hardwriting.jpg what is the content.
  2. Added image '/Users/geek/Pictures/ollama-hardwriting.jpg'
  3. The image contains a handwritten note in Chinese characters, accompanied by an English translation. The note is divided into three sections:
  4. *   **Section 1:** "Ollama" (likely a personal name or title) followed by several Chinese characters.
  5. *   **Section 2:** A numbered list with the following items:
  6.     *   **Item 1:** "Stay hungry, Stay foolish." (an English phrase)
  7.     *   **Item 2:** A series of Chinese characters
  8.     *   **Item 3:** Another series of Chinese characters
  9. The image appears to be a personal note or message written in Chinese and English.
复制代码
从测试来看,手写字体识别中,Llama 3.2 Vision 模型很好的支持英文,无法识别中文,对中文的支持比较差,但是能区分出哪些字符是中文。
  OCR

  
  1. $ ollama run llama3.2-vision
  2. >>> /Users/geek/Pictures/ollama-ocr.png  what is the content.
  3. Added image '/Users/geek/Pictures/ollama-ocr.png'
  4. The image displays a slide from Meta's presentation about Llama 3.2, a large language model developed by Facebook AI
  5. Research (FAIR). The slide provides an overview of the model's capabilities and features.
  6. * **Title**
  7.  + The title of the slide is "Llama 3.2" in large text at the top.
  8. * **Description**
  9.  + Below the title, there is a brief description of Llama 3.2 as a multimodal large language model that can perform
  10. various tasks such as answering questions, generating text, and translating languages.
  11.  + The description also mentions that Llama 3.2 has been trained on a diverse range of datasets, including but not
  12. limited to natural language processing (NLP) tasks like question answering, sentiment analysis, named entity
  13. recognition, part-of-speech tagging, dependency parsing, and word sense induction.
  14. * **Features**
  15.  + The slide highlights some key features of Llama 3.2, including its ability to:
  16.   - Understand and generate human-like language
  17.   - Perform various NLP tasks with high accuracy
  18.   - Learn from large amounts of text data
  19.   - Be fine-tuned for specific applications or domains
  20. * **Benefits**
  21.  + The slide also mentions the benefits of using Llama 3.2, such as:
  22.   - Improving language understanding and generation capabilities
  23.   - Enhancing NLP performance in various tasks
  24.   - Facilitating development of more sophisticated AI systems
  25. Overall, the image provides a concise overview of Llama 3.2's capabilities and features, highlighting its potential
  26. applications and benefits in various fields such as natural language processing, computer vision, and robotics.
  27. >>> which language llama3.2 vision support?
  28. Llama 3.2 Vision supports the following languages:
  29. * English
  30. * German
  31. * French
  32. * Italian
  33. * Portuguese
  34. * Hindi
  35. * Spanish
  36. * Thai
  37. >>> how many size llama3.2 vision has
  38. The Llama 3.2-Vision model is available in two sizes: 11B and 90B.
复制代码
Llama 3.2 Vision 原生支持英文,对英文内容的识别非常准确,输出支持标准的 MarkDown 格式,而且能够正确理解识别的内容,并进行简单的问答
  感觉可以用来做大陆的英文试卷,比如选择题和阅读理解,后续可以测试下看看结果。
  图表

  
  1. >>> /Users/geek/Pictures/ollama-hist.png what is the most popular value?
  2. Added image '/Users/geek/Pictures/ollama-hist.png'
  3. The most popular value in this chart is 4, which appears 49 times. This is represented by the tallest bar on the
  4. graph.
  5. >>> how many bar in the chart.
  6. There are seven bars representing different values from one to seven.
  7. >>> sum the value of bar in the chart
  8. To find the sum of the values represented by each bar, we can add up the numbers on the y-axis for each x-axis value.
  9. *   Bar 1: 7
  10. *   Bar 2: 14
  11. *   Bar 3: 28
  12. *   Bar 4: 49
  13. *   Bar 5: 37
  14. *   Bar 6: 17
  15. *   Bar 7: 8
  16. Summing these values gives us a total of **160**.
复制代码
能够理解问题,并结合图标内容给出答案,缺点是不能精确识别图标中的值,需要提高精确度。
  图片问答

  
  1. >>> /Users/geek/Pictures/ollama-cat.webp what is the animal?
  2. Unknown command '/Users/geek/Pictures/ollama-cat.webp'. Type /? for help
  3. >>> /Users/geek/Pictures/ollama-cat.png what is the animal?
  4. Added image '/Users/geek/Pictures/ollama-cat.png'
  5. The animals in the image are cats. The cat on the left appears to be a long-haired orange and white domestic
  6. shorthair, while the cat on the right seems to be a short-haired brown tabby with a white patch on its chest.
  7. >>> responsd in chinese.
  8. (Animals: Two cats)
  9. This image shows two cats. The left one is an orange and white domestic shorthair, while the right one appears to be
  10. a short-haired brown tabby with a white patch on its chest.
  11. >>> how many cats in this image?
  12. There are 2 cats in this image.
复制代码
从测试看,Llama 3.2 Vision 能够正确识别动物,而且描述图片的内容,能够盘算出有 2 只猫。不足之处是图片格式还不支持 webp。
  性能分析

  

  • 内存占用
  
  Llama 3.2 Vision 11B 在 MacBook Pro 2022, M2x Max 上的内存占用仅有 3.32G,和 Chrome 浏览器占用的内存一个数量级
  

  • ollama 加载与理解图片
  加在图片的时间在 30-60ms 之间。理解图片的内容要长的多,耗时在秒级,具体时间取决于图片内容的复杂度,图片越简单,耗时越少,越复杂耗时越久。像 OCR 和手写字体的例子,耗时在 40s。
  至于 token 的天生速率,基本都在 25 token/s,能够满意日常使用。
  1. # 手写字体识别
  2. /Users/geek/Pictures/ollama-hardwriting.jpg what is the content.
  3. Added image '/Users/geek/Pictures/ollama-hardwriting.jpg'
  4. total duration:       39.276179625s
  5. load duration:        38.170708ms
  6. prompt eval count:    17 token(s)
  7. prompt eval duration: 29.411s
  8. prompt eval rate:     0.58 tokens/s
  9. eval count:           242 token(s)
  10. eval duration:        9.605s
  11. eval rate:            25.20 tokens/s
  12. # 动物识别
  13. >>> /Users/geek/Pictures/ollama-cat.png what is the animal?
  14. Added image '/Users/geek/Pictures/ollama-cat.png'
  15. total duration:       1.628926916s
  16. load duration:        66.047166ms
  17. prompt eval count:    17 token(s)
  18. prompt eval duration: 201ms
  19. prompt eval rate:     84.58 tokens/s
  20. eval count:           40 token(s)
  21. eval duration:        1.15s
  22. eval rate:            34.78 tokens/s
  23. # meta 博客内容 OCR 识别
  24. >>> /Users/geek/Pictures/ollama-ocr.png  what is the content.
  25. Added image '/Users/geek/Pictures/ollama-ocr.png'
  26. total duration:       41.742993542s
  27. load duration:        31.597584ms
  28. prompt eval count:    17 token(s)
  29. prompt eval duration: 29.122s
  30. prompt eval rate:     0.58 tokens/s
  31. eval count:           302 token(s)
  32. eval duration:        12.409s
  33. eval rate:            24.34 tokens/s
复制代码
总结

  Ollama v0.4.0 最大的特点就是对 Meta Llama 3.2 Vision视觉模型的支持了。
  在实际测试过程中,Llama 3.2 Vision能够很好的识别并理解图片内容,理解图片的耗时取决于图片内容的复杂度,简单图片耗时在 5s 内,复杂图片也在 1min 内。
  11B 模型在 Mac 上的内存占用仅 3G 左右,token 的天生速率在 25Token/s,使用结果上个人认为和 ChatGPT 的图片理解在同一水平,能够满意个人的日常使用需求了
  参考资料  [1]
  ollama release: https://github.com/ollama/ollama/releases/tag/v0.4.0
  [2]  Llama 3.2 Vision: https://ollama.com/blog/llama3.2-vision
  [3]  ollama x status: https://x.com/ollama/status/1854269461144174764

免责声明:如果侵犯了您的权益,请联系站长,我们会及时删除侵权内容,谢谢合作!更多信息从访问主页:qidao123.com:ToB企服之家,中国第一个企服评测及商务社交产业平台。

本帖子中包含更多资源

您需要 登录 才可以下载或查看,没有账号?立即注册

x
回复

使用道具 举报

0 个回复

倒序浏览

快速回复

您需要登录后才可以回帖 登录 or 立即注册

本版积分规则

天空闲话

金牌会员
这个人很懒什么都没写!

标签云

快速回复 返回顶部 返回列表