{"id":2791,"date":"2024-04-01T16:28:34","date_gmt":"2024-04-01T08:28:34","guid":{"rendered":"https:\/\/www.aqwu.net\/wp\/?p=2791"},"modified":"2024-04-28T20:00:27","modified_gmt":"2024-04-28T12:00:27","slug":"llm-%e9%87%8f%e5%8c%96llm-quantization-gptq-qat-awq-gguf-ggml-ptq","status":"publish","type":"post","link":"https:\/\/www.aqwu.net\/wp\/?p=2791","title":{"rendered":"LLM \u91cf\u5316(LLM Quantization)| GPTQ | QAT | AWQ | GGUF | GGML | PTQ"},"content":{"rendered":"\n

\u8fd9\u7bc7\u6587\u7ae0\u662f\u5173\u4e8e\u4f60\u5728\u8ba8\u8bba\u4e2dLLM\u4e0d\u65ad\u542c\u5230\u7684\u5404\u79cd\u91cf\u5316\u6280\u672f\u3002\u6b64\u5904\u7684\u76ee\u7684\u662f\u63d0\u4f9b\u5206\u6b65\u8bf4\u660e\u4ee5\u53ca\u4ee3\u7801\uff0c\u60a8\u53ef\u4ee5\u4f7f\u7528\u8fd9\u4e9b\u4ee3\u7801\u81ea\u884c\u6267\u884c\u8fd9\u4e9b\u6280\u672f\u8fdb\u884c\u6a21\u578b\u538b\u7f29\u3002<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

1. Quantization \u91cf\u5316<\/strong><\/h2>\n\n\n\n

\u91cf\u5316\u662f\u6307\u5c06\u9ad8\u7cbe\u5ea6\u6570\u503c\u8f6c\u6362\u4e3a\u4f4e\u7cbe\u5ea6\u6570\u503c\u3002\u8f83\u4f4e\u7cbe\u5ea6\u7684\u5b9e\u4f53\u53ef\u4ee5\u5b58\u50a8\u5728\u78c1\u76d8\u4e0a\u7684\u72ed\u5c0f\u7a7a\u95f4\u4e2d\uff0c\u4ece\u800c\u51cf\u5c11\u5185\u5b58\u9700\u6c42\u3002\u8ba9\u6211\u4eec\u4ece\u4e00\u4e2a\u7b80\u5355\u7684\u91cf\u5316\u793a\u4f8b\u5f00\u59cb\uff0c\u4ee5\u660e\u786e\u6982\u5ff5\u3002<\/p>\n\n\n\n

1.1 \u91cf\u5316\u7684\u7b80\u5355\u793a\u4f8b<\/strong><\/h3>\n\n\n\n

\u5047\u8bbe\u60a8\u6709 FP16 \u683c\u5f0f\u7684 25 \u4e2a\u6743\u91cd\u503c\uff0c\u5982\u4e0b\u56fe\u6240\u793a\u3002<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n
    \n
  1. \u65e7 Range = \u5728 fp16 \u683c\u5f0f\u4e0b\u662f\u6700\u5927\u6743\u91cd\u503c\u51cf\u53bb\u6700\u5c0f\u6743\u91cd\u503c = 0.932\u20130.0609 = 0.871<\/li>\n\n\n\n
  2. \u65b0 Range = \u5728 int8 \u683c\u5f0f\u4e0b\u662f -128 to 127. \u56e0\u6b64, Range = 127-(-128) = 255<\/li>\n\n\n\n
  3. Scale <\/strong>= \u65b0 Range \u6700\u5927\u503c \/ \u65e7 Range \u6700\u5927\u503c = 127 \/ 0.932 = 136.24724986904138<\/strong><\/li>\n\n\n\n
  4. Quantized Value(\u91cf\u5316\u503c<\/strong>) <\/strong>= Round(Scale * Original Value)<\/li>\n<\/ol>\n\n\n\n
    \"\"<\/figure>\n\n\n\n
      \n
    • 5. Dequantized Value(\u53bb\u91cf\u5316\u503c<\/strong>) <\/strong>= Quantized Value \/ Scale<\/li>\n<\/ul>\n\n\n\n
      \"\"<\/figure>\n\n\n\n

      6. Rounding Error(\u820d\u5165\u8bef\u5dee)<\/strong> \u2014 \u8fd9\u91cc\u9700\u8981\u6ce8\u610f\u7684\u91cd\u8981\u4e00\u70b9\u662f\uff0c\u5f53\u6211\u4eec\u5c06 fp16 \u683c\u5f0f\u53bb\u91cf\u5316\u65f6\uff0c\u6211\u4eec\u6ce8\u610f\u5230\u8fd9\u4e9b\u6570\u5b57\u4f3c\u4e4e\u5e76\u4e0d\u5b8c\u5168\u76f8\u540c\u3002\u7b2c\u4e00\u4e2a\u5143\u7d20 0.5415 \u53d8\u4e3a 0.543\u3002\u5728\u5927\u591a\u6570\u5143\u7d20\u4e2d\u90fd\u53ef\u4ee5\u6ce8\u610f\u5230\u76f8\u540c\u7684\u95ee\u9898\u3002\u8fd9\u662f\u91cf\u5316\u7ed3\u679c\u7684\u9519\u8bef – \u53bb\u91cf\u5316\u8fc7\u7a0b\u3002<\/p>\n\n\n\n

      \u73b0\u5728\u6211\u4eec\u5df2\u7ecf\u4e86\u89e3\u4e86\u91cf\u5316\u7684\u6838\u5fc3\uff0c\u8ba9\u6211\u4eec\u7ee7\u7eed\u8ba8\u8bba\u91cf\u5316\u7684LLM\u7c7b\u578b\u3002<\/p>\n\n\n\n

      2. GPTQ<\/strong><\/h2>\n\n\n\n
        \n
      • GPTQ \u662f\u8bad\u7ec3\u540e\u91cf\u5316\u65b9\u6cd5\u3002\u8fd9\u610f\u5473\u7740\u4e00\u65e6\u4f60\u8fdb\u884c\u4e86\u9884\u8bad\u7ec3LLM\uff0c\u4f60\u53ea\u9700\u5c06\u6a21\u578b\u53c2\u6570\u8f6c\u6362\u4e3a\u8f83\u4f4e\u7684\u7cbe\u5ea6\u3002<\/li>\n\n\n\n
      • GPTQ \u662f GPU \u800c\u4e0d\u662f CPU \u7684\u9996\u9009\u3002<\/li>\n\n\n\n
      • \u4e0b\u9762\u5217\u51fa\u4e86\u5404\u79cd GPTQ\u3002\n
          \n
        • Static Range(\u9759\u6001\u8303\u56f4) GPTQ<\/em> \u2014 \u60a8\u53ef\u4ee5\u4ee5\u8f83\u4f4e\u7684\u7cbe\u5ea6\u8f6c\u6362\u6743\u91cd\u548c\u6fc0\u6d3b\u3002<\/li>\n\n\n\n
        • Dynamic Range(\u52a8\u6001\u8303\u56f4<\/em>) GPTQ<\/em> \u2014 \u60a8\u53ef\u4ee5\u4ee5\u8f83\u4f4e\u7684\u7cbe\u5ea6\u8f6c\u6362\u6743\u91cd\uff0c\u5e76\u5f00\u53d1\u4e00\u4e2a\u7528\u4e8e\u5c06\u6fc0\u6d3b\u8f6c\u6362\u4e3a\u8f83\u4f4e\u7cbe\u5ea6\u7684\u51fd\u6570\u3002\u6b64\u51fd\u6570\u6700\u7ec8\u5c06\u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d\u7528\u4e8e\u91cf\u5316\u6fc0\u6d3b\u3002<\/li>\n\n\n\n
        • Weight Quantization<\/em>(\u6743\u91cd\u91cf\u5316) \u2014 \u91cf\u5316\u53ef\u4ee5\u8282\u7701\u7a7a\u95f4\uff0c\u56e0\u4e3a\u5b83\u4f1a\u964d\u4f4e\u6743\u91cd\u7684\u7cbe\u5ea6\u548c\/\u6216\u6a21\u578b\u7684\u6fc0\u6d3b\u3002\u5728\u8fd9\u91cc\uff0c\u5728\u63a8\u7406\u671f\u95f4\uff0c\u8f93\u5165\u4ecd\u7136\u662f float32 \u683c\u5f0f\uff0c\u5e76\u4e14\u8981\u4f7f\u7528\u8f93\u5165\u8fdb\u884c\u6743\u91cd\u8ba1\u7b97\uff0c\u6211\u4eec\u9700\u8981\u4ee5\u4e0e\u8f93\u5165\u76f8\u540c\u7684\u7cbe\u5ea6\u8fd4\u56de\u6743\u91cd\u3002\u7531\u4e8e\u56db\u820d\u4e94\u5165\u95ee\u9898\uff0c\u6b64\u8fc7\u7a0b\u4f1a\u5bfc\u81f4\u51c6\u786e\u6027\u4e0b\u964d\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n

          2.1 Static Range Quantization \u9759\u6001\u8303\u56f4\u91cf\u5316<\/strong><\/h3>\n\n\n\n
            \n
          • \u5982\u679c\u4f60\u6253\u7b97\u91cf\u5316\u6743\u91cd\u548c\u6fc0\u6d3b\uff0c\u4f60\u9700\u8981\u4e00\u4e2a\u6837\u672c\u6821\u51c6\u6570\u636e\u96c6\u6765\u505aGPTQ\u3002<\/li>\n\n\n\n
          • \u6821\u51c6\u6570\u636e\u96c6 \u2014 \u6b64\u6570\u636e\u96c6\u53ef\u4ee5\u4ece\u539f\u59cb\u6570\u636e\u96c6\u4e2d\u91c7\u6837\u3002\u4f8b\u5982\uff0c\u4ece\u539f\u59cb\u9884\u8bad\u7ec3\u6570\u636e\u96c6\u4e2d\u91c7\u6837\u7684 1000 \u4e2a\u6570\u636e\u70b9\u5145\u5f53\u6574\u4e2a\u6570\u636e\u96c6\u7684\u4ee3\u8868\u6027\u6837\u672c\u3002<\/li>\n\n\n\n
          • \u5728\u6821\u51c6\u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u63a8\u7406 \u2014 \u60a8\u5c06\u4f7f\u7528\u6b64\u6821\u51c6\u6570\u636e\u96c6\u8fdb\u884c\u63a8\u7406\uff0c\u4ee5\u67e5\u627e\u91c7\u6837\u6743\u91cd\u548c\u76f8\u5e94\u6fc0\u6d3b\u7684\u5206\u5e03\u3002\u8be5\u5206\u5e03\u5c06\u4f5c\u4e3a\u91cf\u5316\u7684\u57fa\u7840\u3002\n
              \n
            • \u4f8b\u5982\uff0c\u7279\u5b9a\u5c42\u4e2d\u7684\u6fc0\u6d3b\u8303\u56f4\u4e3a 0.2 \u5230 0.9\uff0c\u6743\u91cd\u8303\u56f4\u4e3a 0.1 \u5230 0.3\u3002\u4e00\u65e6\u4f60\u6709\u4e86\u8303\u56f4\uff0c\u5373\u6700\u5c0f\u503c-\u6700\u5927\u503c\uff0c\u4f60\u5c31\u53ef\u4ee5\u4f7f\u7528\u672c\u6587\u524d\u9762\u89e3\u91ca\u7684\u6570\u5b66\u65b9\u6cd5\u5bf9\u8be5\u5c42\u8fdb\u884c\u91cf\u5316\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n
            • \u9759\u6001\u8ddd\u79bb\u91cf\u5316\u7b97\u6cd5\u6982\u8ff0<\/em>\n
                \n
              • \u6211\u4eec\u5728 GPTQ \u7b97\u6cd5\u4e2d\u9010\u5c42\u91cf\u5316\u795e\u7ecf\u7f51\u7edc\u3002<\/li>\n\n\n\n
              • \u6211\u4eec\u5c06\u6bcf\u5c42\u7684\u6743\u91cd\u6307\u6807\u5212\u5206\u4e3a\u51e0\u7ec4\u5217\u3002<\/li>\n\n\n\n
              • \u8fd9\u4e9b\u5217\u7ec4\u4ee5\u8fed\u4ee3\u65b9\u5f0f\u8fdb\u884c\u5904\u7406\u3002\u8ba9\u6211\u4eec\u901a\u8fc7\u4e00\u4e2a\u4f8b\u5b50\u6765\u4e86\u89e3\u8fd9\u4e00\u70b9\u3002<\/li>\n\n\n\n
              • \u5982\u679c\u6211\u4eec\u5c06 GPTQ \u7684 group_size \u53c2\u6570\u8bbe\u7f6e\u4e3a 128;\u6743\u91cd\u6307\u6807\u5206\u4e3a\u4e00\u7ec4\uff0c\u6bcf\u7ec4 128 \u5217\u3002<\/li>\n\n\n\n
              • \u5728\u5305\u542b 128 \u5217\u7684\u6bcf\u4e2a\u7ec4\u4e2d\uff0c\u5bf9\u4e00\u5217\u7684\u6570\u636e\u8fdb\u884c\u91cf\u5316;\u4e4b\u540e\uff0c\u8be5\u7ec4\u4e2d\u7684\u5176\u4f59\u6743\u91cd\u5c06\u66f4\u65b0\uff0c\u4ee5\u8865\u507f\u91cf\u5316\u5f15\u5165\u7684\u8bef\u5dee\u3002<\/li>\n\n\n\n
              • \u4e00\u6b21\uff0c\u5904\u7406\u4e00\u7ec4\u5217\u5bf9\u5e94\u7684\u6570\u636e;\u6574\u4e2a\u77e9\u9635\uff08\u5176\u4ed6\u7ec4\uff09\u4e0a\u7684\u5176\u4f59\u5217\u5c06\u66f4\u65b0\u4ee5\u8865\u507f\u8bef\u5dee\u3002<\/li>\n\n\n\n
              • \u8fd9\u4e2a\u5b8c\u6574\u7684\u8fc7\u7a0b\u4e5f\u79f0\u4e3a\u201c\u5ef6\u8fdf\u6279\u91cf\u66f4\u65b0\u201d<\/li>\n\n\n\n
              • \u73b0\u5728\u8ba9\u6211\u4eec\u770b\u4e00\u4e9b\u4ee3\u7801\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n

                2.2 GPTQ\u4ee3\u7801<\/strong><\/h3>\n\n\n\n
                  \n
                • \u8fd9\u79cd\u91cf\u5316\u9700\u8981 GPU\u3002<\/li>\n\n\n\n
                • \u8d77\u521d\uff0c\u6211\u5c1d\u8bd5\u91cf\u5316 7B \u5206\u7247\u7c73\u65af\u7279\u62c9\u5c14\u6a21\u578b\uff0c\u4f46\u5931\u8d25\u4e86\u3002\u8fd9\u662f\u56e0\u4e3a\u8be5\u6a21\u578b\u5728\u4e0b\u8f7d\u65f6\u9996\u5148\u52a0\u8f7d\u5230 CPU\uff0c\u800c T4 \u6ca1\u6709\u8db3\u591f\u7684 CPU RAM \u6765\u652f\u6301\u3002<\/li>\n\n\n\n
                • \u6211\u6700\u7ec8\u9009\u62e9\u4e86\u4e00\u4e2a\u5c0f\u5c3a\u5bf8\u7684\u6a21\u578b\uff0c\u53ef\u4ee5\u5bb9\u7eb3\u5728 Google Collab \u4e0a\u7684\u514d\u8d39 T4 \u5b9e\u4f8b\u4e2d\u3002\u8be5\u6a21\u578b\u662f\u6765\u81ea HF repo \u7684 bigscience\/bloom-3b\u3002<\/li>\n<\/ul>\n\n\n\n
                  pip install auto_gptq<\/pre><\/div>\n\n\n\n
                  import torch\nfrom auto_gptq import AutoGPTQForCausalLM, BaseQuantizeConfig\nfrom transformers import TextGenerationPipeline\nfrom transformers import AutoTokenizer\n\npretrained_model_name = \"bigscience\/bloom-3b\" \nquantize_config = BaseQuantizeConfig(bits=4, group_size=128)\n\n# Tensors of bloom are of float16. Hence, torch_dtype=torch.float16. Do not leave torch_dtype as \"auto\" as this leads to a warning of implicit dtype conversion\nmodel = AutoGPTQForCausalLM.from_pretrained(pretrained_model_name, quantize_config, trust_remote_code=False, device_map=\"auto\", torch_dtype=torch.float16)  # changing device map to \"cuda\" does not have any impact on T4 GPU mem usage.\ntokenizer = AutoTokenizer.from_pretrained(pretrained_model_name)\n\n# Calibration\nexamples = [\n    tokenizer(\n        \"Automated machine learning is the process of automating the tasks of applying machine learning to real-world problems. AutoML potentially includes every stage from beginning with a raw dataset to building a machine learning model ready for deployment.\"\n    )\n]  # giving only 1 example here for testing. In an real world scenario, you might want to give 500-1000 samples.\nmodel.quantize(examples)\n\nquantized_model_dir = \"bloom3b_q4b_gs128\"\nmodel.save_quantized(quantized_model_dir)<\/pre><\/div>\n\n\n\n
                  \u539f\u6587\u94fe\u63a5\uff1ahttps:\/\/medium.com\/@siddharth.vij10\/llm-quantization-gptq-qat-awq-gguf-ggml-ptq-2e172cd1b3b5<\/p>\n","protected":false},"excerpt":{"rendered":"
                  \u8fd9\u7bc7\u6587\u7ae0\u662f\u5173\u4e8e\u4f60\u5728\u8ba8\u8bba\u4e2dLLM\u4e0d\u65ad\u542c\u5230\u7684\u5404\u79cd\u91cf\u5316\u6280\u672f\u3002\u6b64\u5904\u7684\u76ee\u7684\u662f\u63d0\u4f9b\u5206\u6b65\u8bf4\u660e\u4ee5\u53ca\u4ee3\u7801\uff0c\u60a8\u53ef\u4ee5\u4f7f\u7528\u8fd9\u4e9b\u4ee3\u7801\u81ea\u884c […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"default","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"set","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"var(--ast-global-color-4)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"_jetpack_memberships_contains_paid_content":false,"footnotes":""},"categories":[313,289,443,442,312],"tags":[408,406,404,409,407,405,403],"class_list":["post-2791","post","type-post","status-publish","format-standard","hentry","category-chatgpt","category-gpt","category-llm","category-llms","category-openai","tag-awq","tag-gguf","tag-llm","tag-ptq","tag-qat","tag-qptq","tag-quantization"],"views":2694,"jetpack_sharing_enabled":true,"jetpack_featured_media_url":"","_links":{"self":[{"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/posts\/2791","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=2791"}],"version-history":[{"count":4,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/posts\/2791\/revisions"}],"predecessor-version":[{"id":2803,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=\/wp\/v2\/posts\/2791\/revisions\/2803"}],"wp:attachment":[{"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=2791"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=2791"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.aqwu.net\/wp\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=2791"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}