{"id":588,"date":"2024-08-12T16:42:46","date_gmt":"2024-08-12T08:42:46","guid":{"rendered":"http:\/\/mapleleaf666.vip\/?p=588"},"modified":"2024-08-12T16:43:34","modified_gmt":"2024-08-12T08:43:34","slug":"%e4%b9%a6%e7%94%9f%e5%a4%a7%e6%a8%a1%e5%9e%8b%e9%97%af%e5%85%b3%e8%90%a5-8gb%e6%98%be%e5%ad%98%e7%8e%a9%e8%bd%ac%e4%b9%a6%e7%94%9f%e5%a4%a7%e6%a8%a1%e5%9e%8bdemo","status":"publish","type":"post","link":"http:\/\/mapleleaf666.vip\/?p=588","title":{"rendered":"\u4e66\u751f\u5927\u6a21\u578b\u95ef\u5173\u8425\u2014\u20148GB\u663e\u5b58\u73a9\u8f6c\u4e66\u751f\u5927\u6a21\u578bDemo"},"content":{"rendered":"\n
\n

\u672c\u6587\u53c2\u8003\u5b9e\u6218\u8425\u63d0\u4f9b\u7684\u6559\u7a0b\uff1aTutorial\/docs\/L1\/Demo\/readme.md at camp3 \u00b7 InternLM\/Tutorial (github.com)<\/a><\/p>\n<\/blockquote>\n\n\n\n

\u672c\u5173\u5361\u4e3b\u8981\u9700\u8981\u5b8c\u6210\u7684\u4efb\u52a1\u5982\u4e0b\uff1a<\/p>\n\n\n\n

    \n
  1. InternLM2-Chat-1.8B \u6a21\u578b\u7684\u90e8\u7f72\uff08\u57fa\u7840\u4efb\u52a1\uff09<\/li>\n\n\n\n
  2. InternLM-XComposer2-VL-1.8B \u6a21\u578b\u7684\u90e8\u7f72\uff08\u8fdb\u9636\u4efb\u52a1\uff09<\/li>\n\n\n\n
  3. InternVL2-2B \u6a21\u578b\u7684\u90e8\u7f72\uff08\u8fdb\u9636\u4efb\u52a1\uff09<\/li>\n<\/ol>\n\n\n\n

    1.\u73af\u5883\u914d\u7f6e<\/h2>\n\n\n\n

    \u9996\u5148\u6211\u4eec\u9700\u8981\u521b\u5efa\u4e00\u4e2a\u5f00\u53d1\u673a\uff0c\u914d\u7f6e\u5982\u4e0b\uff1a<\/p>\n\n\n\n

    \"\u622a\u5c4f2024-07-28<\/div><\/figure>\n\n\n\n

    \u521b\u5efa\u5b8c\u6210\u540e\uff0c\u6211\u4eec\u9700\u8981\u521b\u5efa\u4e00\u4e2a\u53ef\u7528\u7684\u73af\u5883\u6765\u8fd0\u884c\u6211\u4eec\u7684demo\uff1a<\/p>\n\n\n\n

    # \u521b\u5efa\u73af\u5883\nconda create -n demo python=3.10 -y\n# \u6fc0\u6d3b\u73af\u5883\nconda activate demo\n# \u5b89\u88c5 torch\nconda install pytorch==2.1.2 torchvision==0.16.2 torchaudio==2.1.2 pytorch-cuda=12.1 -c pytorch -c nvidia -y\n# \u5b89\u88c5\u5176\u4ed6\u4f9d\u8d56\npip install transformers==4.38\npip install sentencepiece==0.1.99\npip install einops==0.8.0\npip install protobuf==5.27.2\npip install accelerate==0.33.0\npip install streamlit==1.37.0\n<\/code><\/pre>\n\n\n\n
    2.Cli Demo\u90e8\u7f72InternLM2-Chat-1.8B\u6a21\u578b<\/h2>\n\n\n\n
    \u9996\u5148\u6211\u4eec\u521b\u5efa\u4e00\u4e2a\u4efb\u52a1\u7528\u76ee\u5f55<\/p>\n\n\n\n
    mkdir -p \/root\/work\/demo\ntouch \/root\/work\/demo\/cli_demo.py<\/code><\/pre>\n\n\n\n
    \u63a5\u4e0b\u6765\u6211\u4eec\u901a\u8fc7\u7f16\u5199\u4ee3\u7801\u52a0\u8f7dInternLM\u6a21\u578b\uff0c\u5e76\u5b9e\u73b0\u4e00\u4e2a\u7b80\u5355\u7684\u804a\u5929\u673a\u5668\u4eba\u5e94\u7528\uff1a<\/p>\n\n\n\n
    import torch\nfrom transformers import AutoTokenizer, AutoModelForCausalLM\n\n# \u5b9a\u4e49\u6a21\u578b\u8def\u5f84\nmodel_name_or_path = \"\/root\/share\/new_models\/Shanghai_AI_Laboratory\/internlm2-chat-1_8b\"\n\n# \u52a0\u8f7d\u6a21\u578b\u548c\u5206\u8bcd\u5668\ntokenizer = AutoTokenizer.from_pretrained(model_name_or_path, trust_remote_code=True, device_map='cuda:0')\nmodel = AutoModelForCausalLM.from_pretrained(model_name_or_path, trust_remote_code=True, torch_dtype=torch.bfloat16, device_map='cuda:0')\nmodel = model.eval()\n\nsystem_prompt = \"\"\"You are an AI assistant whose name is InternLM (\u4e66\u751f\u00b7\u6d66\u8bed).\n- InternLM (\u4e66\u751f\u00b7\u6d66\u8bed) is a conversational language model that is developed by Shanghai AI Laboratory (\u4e0a\u6d77\u4eba\u5de5\u667a\u80fd\u5b9e\u9a8c\u5ba4). It is designed to be helpful, honest, and harmless.\n- InternLM (\u4e66\u751f\u00b7\u6d66\u8bed) can understand and communicate fluently in the language chosen by the user such as English and \u4e2d\u6587.\n\"\"\"\n\nmessages = [(system_prompt, '')]\n\nprint(\"=============Welcome to InternLM chatbot, type 'exit' to exit.=============\")\n\n# \u542f\u52a8\u804a\u5929\u5faa\u73af\nwhile True:\n    input_text = input(\"\\nUser  >>> \")\n    input_text = input_text.replace(' ', '')\n    if input_text == \"exit\":\n        break\n\n    length = 0\n    for response, _ in model.stream_chat(tokenizer, input_text, messages):\n        if response is not None:\n            print(response[length:], flush=True, end=\"\")\n            length = len(response)<\/code><\/pre>\n\n\n\n
    \u8fd0\u884c\u8be5\u4ee3\u7801\uff0c\u8fd0\u884c\u6548\u679c\u5982\u4e0b\uff1a<\/p>\n\n\n\n
    \"image-20240728161847402\"\/<\/div><\/figure>\n\n\n\n
    \u6211\u4eec\u5c1d\u8bd5\u7528\u4ed6\u751f\u6210\u4e00\u4e2a\u6709\u5173\u5927\u6d77\u548c\u68a6\u60f3\u7684\u6545\u4e8b\uff0c\u8fd0\u884c\u7ed3\u679c\u5982\u4e0b\uff1a<\/p>\n\n\n\n
    \"image-20240728162641529\"\/<\/div><\/figure>\n\n\n\n
    3.LMDeploy \u90e8\u7f72 InternLM-XComposer2-VL-1.8B \u6a21\u578b<\/h2>\n\n\n\n
    \u5728\u90e8\u7f72\u4e4b\u524d\uff0c\u6211\u4eec\u5148\u4e86\u89e3\u4e00\u4e0bInternLM-XComposer2-VL-1.8B \u6a21\u578b\u548cLMDeploy\u5de5\u5177\u3002<\/p>\n\n\n\n
    InternLM-XComposer2 \u662f\u4e00\u6b3e\u57fa\u4e8e InternLM2 \u7684\u89c6\u89c9\u8bed\u8a00\u5927\u6a21\u578b\uff0c\u5176\u64c5\u957f\u81ea\u7531\u5f62\u5f0f\u7684\u6587\u672c\u56fe\u50cf\u5408\u6210\u548c\u7406\u89e3\u3002\u5176\u4e3b\u8981\u7279\u70b9\u5305\u62ec\uff1a<\/p>\n\n\n\n
      \n
    • \u81ea\u7531\u5f62\u5f0f\u7684\u4ea4\u9519\u6587\u672c\u56fe\u50cf\u5408\u6210\uff1aInternLM-XComposer2 \u53ef\u4ee5\u6839\u636e\u5927\u7eb2\u3001\u8be6\u7ec6\u6587\u672c\u8981\u6c42\u548c\u53c2\u8003\u56fe\u50cf\u7b49\u4e0d\u540c\u8f93\u5165\uff0c\u751f\u6210\u8fde\u8d2f\u4e14\u4e0a\u4e0b\u6587\u76f8\u5173\uff0c\u5177\u6709\u4ea4\u9519\u56fe\u50cf\u548c\u6587\u672c\u7684\u6587\u7ae0\uff0c\u4ece\u800c\u5b9e\u73b0\u9ad8\u5ea6\u53ef\u5b9a\u5236\u7684\u5185\u5bb9\u521b\u5efa\u3002<\/li>\n\n\n\n
    • \u51c6\u786e\u7684\u89c6\u89c9\u8bed\u8a00\u95ee\u9898\u89e3\u51b3\uff1aInternLM-XComposer2 \u57fa\u4e8e\u81ea\u7531\u5f62\u5f0f\u7684\u6307\u4ee4\u51c6\u786e\u5730\u5904\u7406\u591a\u6837\u5316\u548c\u5177\u6709\u6311\u6218\u6027\u7684\u89c6\u89c9\u8bed\u8a00\u95ee\u7b54\u4efb\u52a1\uff0c\u5728\u8bc6\u522b\uff0c\u611f\u77e5\uff0c\u8be6\u7ec6\u6807\u7b7e\uff0c\u89c6\u89c9\u63a8\u7406\u7b49\u65b9\u9762\u8868\u73b0\u51fa\u8272\u3002<\/li>\n\n\n\n
    • \u4ee4\u4eba\u60ca\u53f9\u7684\u6027\u80fd\uff1a\u57fa\u4e8e InternLM2-7B \u7684InternLM-XComposer2 \u5728\u591a\u4e2a\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u4f4d\u4e8e\u5f00\u6e90\u591a\u6a21\u6001\u6a21\u578b\u7b2c\u4e00\u68af\u961f\uff0c\u800c\u4e14\u5728\u90e8\u5206\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u4e0e GPT-4V \u548c Gemini Pro \u76f8\u5f53\u751a\u81f3\u8d85\u8fc7\u5b83\u4eec\u3002<\/li>\n<\/ul>\n\n\n\n
      LMDeploy \u662f\u4e00\u4e2a\u7528\u4e8e\u538b\u7f29\u3001\u90e8\u7f72\u548c\u670d\u52a1 LLM \u7684\u5de5\u5177\u5305\uff0c\u7531 MMRazor \u548c MMDeploy \u56e2\u961f\u5f00\u53d1\u3002\u5b83\u5177\u6709\u4ee5\u4e0b\u6838\u5fc3\u529f\u80fd\uff1a<\/p>\n\n\n\n
        \n
      • \u9ad8\u6548\u7684\u63a8\u7406\uff1aLMDeploy \u901a\u8fc7\u5f15\u5165\u6301\u4e45\u5316\u6279\u5904\u7406\u3001\u5757 KV \u7f13\u5b58\u3001\u52a8\u6001\u5206\u5272\u4e0e\u878d\u5408\u3001\u5f20\u91cf\u5e76\u884c\u3001\u9ad8\u6027\u80fd CUDA \u5185\u6838\u7b49\u5173\u952e\u6280\u672f\uff0c\u63d0\u4f9b\u4e86\u6bd4 vLLM \u9ad8 1.8 \u500d\u7684\u63a8\u7406\u6027\u80fd\u3002<\/li>\n\n\n\n
      • \u6709\u6548\u7684\u91cf\u5316\uff1aLMDeploy \u652f\u6301\u4ec5\u6743\u91cd\u91cf\u5316\u548c k\/v \u91cf\u5316\uff0c4bit \u63a8\u7406\u6027\u80fd\u662f FP16 \u7684 2.4 \u500d\u3002\u91cf\u5316\u540e\u6a21\u578b\u8d28\u91cf\u5df2\u901a\u8fc7 OpenCompass \u8bc4\u4f30\u786e\u8ba4\u3002<\/li>\n\n\n\n
      • \u8f7b\u677e\u7684\u5206\u53d1\uff1a\u5229\u7528\u8bf7\u6c42\u5206\u53d1\u670d\u52a1\uff0cLMDeploy \u53ef\u4ee5\u5728\u591a\u53f0\u673a\u5668\u548c\u8bbe\u5907\u4e0a\u8f7b\u677e\u9ad8\u6548\u5730\u90e8\u7f72\u591a\u6a21\u578b\u670d\u52a1\u3002<\/li>\n\n\n\n
      • \u4ea4\u4e92\u5f0f\u63a8\u7406\u6a21\u5f0f\uff1a\u901a\u8fc7\u7f13\u5b58\u591a\u8f6e\u5bf9\u8bdd\u8fc7\u7a0b\u4e2d\u6ce8\u610f\u529b\u7684 k\/v\uff0c\u63a8\u7406\u5f15\u64ce\u8bb0\u4f4f\u5bf9\u8bdd\u5386\u53f2\uff0c\u4ece\u800c\u907f\u514d\u91cd\u590d\u5904\u7406\u5386\u53f2\u4f1a\u8bdd\u3002<\/li>\n\n\n\n
      • \u4f18\u79c0\u7684\u517c\u5bb9\u6027\uff1aLMDeploy\u652f\u6301 KV Cache Quant\uff0cAWQ \u548c\u81ea\u52a8\u524d\u7f00\u7f13\u5b58\u540c\u65f6\u4f7f\u7528\u3002<\/li>\n<\/ul>\n\n\n\n
        \u4e0b\u9762\u6211\u4eec\u5c06\u5c1d\u8bd5\u914d\u7f6e\u73af\u5883\u4f7f\u7528LMDeploy\uff0c\u5b8c\u6210InternLM-XComposer2-VL-1.8B\u6a21\u578b\u7684\u90e8\u7f72\u3002\u9996\u5148\u9700\u8981\u914d\u7f6e\u76f8\u5173\u7684\u73af\u5883\uff1a<\/p>\n\n\n\n
        conda activate demo\npip install lmdeploy[all]==0.5.1\npip install timm==1.0.7<\/code><\/pre>\n\n\n\n
        \u63a5\u4e0b\u6765\uff0c\u53ef\u4ee5\u901a\u8fc7LMDeploy\u542f\u52a8\u4e00\u4e2a\u4e0eInternLM-XComposer2-VL-1.8B\u6a21\u578b\u4ea4\u4e92\u7684Gradio\u670d\u52a1\u3002<\/p>\n\n\n\n
        \n
        Gradio\u662f\u4e00\u4e2a\u5f00\u6e90\u7684Python\u5e93\uff0c\u5b83\u5141\u8bb8\u7528\u6237\u5feb\u901f\u521b\u5efa\u7528\u6237\u53cb\u597d\u7684Web\u754c\u9762\uff0c\u4ee5\u4fbf\u4e0e\u673a\u5668\u5b66\u4e60\u6a21\u578b\u548c\u6570\u636e\u8fdb\u884c\u4ea4\u4e92\u3002<\/p>\n<\/blockquote>\n\n\n\n
        lmdeploy serve gradio \/share\/new_models\/Shanghai_AI_Laboratory\/internlm-xcomposer2-vl-1_8b --cache-max-entry-count 0.1<\/code><\/pre>\n\n\n\n
        \u8fd0\u884c\u547d\u4ee4\u540e\uff0c\u7ec8\u7aef\u5982\u56fe\u6240\u793a\uff1a<\/p>\n\n\n\n
        \"image-20240728164807857\"\/<\/div><\/figure>\n\n\n\n
        \u540c\u65f6\uff0c\u6211\u4eec\u53ef\u4ee5\u76f4\u63a5\u8bbf\u95ee\u672c\u5730\u76846006\u7aef\u53e3\uff0c\u5373\u53ef\u770b\u5230\u53ef\u89c6\u5316\u754c\u9762\uff1a<\/p>\n\n\n\n
        \"image-20240728164859993\"\/<\/div><\/figure>\n\n\n\n
        \u6211\u4eec\u4e0a\u4f20\u4e00\u5f20\u5c0f\u732b\u7684\u7167\u7247\uff0c\u5e76\u8be2\u95ee\u4ed6\u56fe\u7247\u91cc\u6709\u4ec0\u4e48\uff1a<\/p>\n\n\n\n
        \"linggan\"<\/div><\/figure>\n\n\n\n
        \"image-20240728165902529\"\/<\/div><\/figure>\n\n\n\n
        4.LMDeploy \u90e8\u7f72 InternVL2-2B \u6a21\u578b<\/h2>\n\n\n\n
        InternVL2 \u662f\u4e0a\u6d77\u4eba\u5de5\u667a\u80fd\u5b9e\u9a8c\u5ba4\u63a8\u51fa\u7684\u65b0\u4e00\u4ee3\u89c6\u89c9-\u8bed\u8a00\u591a\u6a21\u6001\u5927\u6a21\u578b\uff0c\u662f\u9996\u4e2a\u7efc\u5408\u6027\u80fd\u5ab2\u7f8e\u56fd\u9645\u95ed\u6e90\u5546\u4e1a\u6a21\u578b\u7684\u5f00\u6e90\u591a\u6a21\u6001\u5927\u6a21\u578b\u3002<\/p>\n\n\n\n
        \u540c\u6837\u7684\uff0c\u6211\u4eec\u53ef\u4ee5\u901a\u8fc7LMDeploy\u542f\u52a8\u4e00\u4e2a\u4e0e InternVL2-2B\u6a21\u578b\u4ea4\u4e92\u7684Gradio\u670d\u52a1<\/p>\n\n\n\n
        lmdeploy serve gradio \/share\/new_models\/OpenGVLab\/InternVL2-2B --cache-max-entry-count 0.1<\/code><\/pre>\n\n\n\n
        \"image-20240728190213213\"\/<\/div><\/figure>\n","protected":false},"excerpt":{"rendered":"
        \u672c\u6587\u53c2\u8003\u5b9e\u6218\u8425\u63d0\u4f9b\u7684\u6559\u7a0b\uff1aTutorial\/docs\/L1\/Demo\/readme.md at camp3 \u00b7 […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[4],"tags":[35],"class_list":["post-588","post","type-post","status-publish","format-standard","hentry","category-note","tag-llm"],"_links":{"self":[{"href":"http:\/\/mapleleaf666.vip\/index.php?rest_route=\/wp\/v2\/posts\/588","targetHints":{"allow":["GET"]}}],"collection":[{"href":"http:\/\/mapleleaf666.vip\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/mapleleaf666.vip\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/mapleleaf666.vip\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/mapleleaf666.vip\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=588"}],"version-history":[{"count":2,"href":"http:\/\/mapleleaf666.vip\/index.php?rest_route=\/wp\/v2\/posts\/588\/revisions"}],"predecessor-version":[{"id":590,"href":"http:\/\/mapleleaf666.vip\/index.php?rest_route=\/wp\/v2\/posts\/588\/revisions\/590"}],"wp:attachment":[{"href":"http:\/\/mapleleaf666.vip\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=588"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/mapleleaf666.vip\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=588"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/mapleleaf666.vip\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=588"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}