{"id":2317,"date":"2024-02-28T15:35:47","date_gmt":"2024-02-28T07:35:47","guid":{"rendered":"https:\/\/www.aqwu.net\/wp\/?p=2317"},"modified":"2024-04-28T20:04:04","modified_gmt":"2024-04-28T12:04:04","slug":"1-%e4%bd%8d-llm-%e6%97%b6%e4%bb%a3%ef%bc%9a%e6%89%80%e6%9c%89%e5%a4%a7%e5%9e%8b%e8%af%ad%e8%a8%80%e6%a8%a1%e5%9e%8b%e9%83%bd%e6%98%af1-58%e4%bd%8d","status":"publish","type":"post","link":"https:\/\/www.aqwu.net\/wp\/?p=2317","title":{"rendered":"1 \u4f4d LLM \u65f6\u4ee3\uff1a\u6240\u6709\u5927\u578b\u8bed\u8a00\u6a21\u578b\u90fd\u662f1.58\u4f4d"},"content":{"rendered":"\n

\u9a6c\u8212\u660e \u738b\u5b8f\u5b871<\/sup>\u9a6c\u73b2\u73b2 \u738b\u78ca \u738b\u6587\u8f89
\u9ec4\u5c11\u6db5 \u674e\u4e1c\u745e\u5e73 \u738b\u7ee7\u9f99 \u859b\u5bcc\u8339 \u9b4f\u5f66\u22c4
https:\/\/aka.ms\/GeneralAI<\/a>
<\/strong>\u540c\u7b49\u8d21\u732e\u3002\u22c4\u901a\u8baf\u4f5c\u8005\u3002S. \u9a6c\u3001L. \u9a6c\u3001L. Wang\u3001W. Wang\u3001S. Huang\u3001L. Dong\u3001J. Xue\u3001F. Wei \u5c31\u804c\u4e8e Microsoft Research\u3002H. Wang \u548c R. Wang \u5c31\u804c\u4e8e\u4e2d\u56fd\u79d1\u5b66\u9662\u5927\u5b66\u3002<\/p>\n\n\n\n

0. \u6458\u8981<\/strong><\/h2>\n\n\n\n

\u6700\u8fd1\u7684\u7814\u7a76\uff0c\u5982BitNet[23<\/a>]\uff0c\u6b63\u5728\u4e3a 1 \u4f4d\u5927\u578b\u8bed\u8a00\u6a21\u578b \uff08LLM\uff09 \u7684\u65b0\u65f6\u4ee3\u94fa\u5e73\u9053\u8def\u3002\u5728\u8fd9\u9879\u5de5\u4f5c\u4e2d\uff0c\u6211\u4eec\u5f15\u5165\u4e86\u4e00\u4e2a 1 \u4f4d LLM \u53d8\u4f53\uff0c\u5373 BitNet b1.58<\/strong>\uff0c\u5176\u4e2d LLM \u7684\u6bcf\u4e2a\u53c2\u6570\uff08\u6216\u6743\u91cd\uff09\u90fd\u662f\u4e09\u5143 {-1\uff0c 0\uff0c 1}\u3002\u5b83\u4e0e\u5168\u7cbe\u5ea6\uff08\u5373 FP16 \u6216 BF16\uff09Transformer LLM \u5728\u56f0\u60d1\u5ea6\u548c\u6700\u7ec8\u4efb\u52a1\u6027\u80fd\u65b9\u9762\u5177\u6709\u76f8\u540c\u7684\u6a21\u578b\u5927\u5c0f\u548c\u8bad\u7ec3\u4ee4\u724c\u76f8\u5339\u914d\uff0c\u540c\u65f6\u5728\u5ef6\u8fdf\u3001\u5185\u5b58\u3001\u541e\u5410\u91cf\u548c\u80fd\u8017\u65b9\u9762\u66f4\u5177\u6210\u672c\u6548\u76ca\u3002 \u66f4\u6df1\u523b\u7684\u662f\uff0c1.58 \u4f4d LLM \u5b9a\u4e49\u4e86\u4e00\u79cd\u65b0\u7684\u6269\u5c55\u5b9a\u5f8b\u548c\u65b9\u6cd5\uff0c\u7528\u4e8e\u8bad\u7ec3\u9ad8\u6027\u80fd\u548c\u5177\u6709\u6210\u672c\u6548\u76ca\u7684\u65b0\u4e00\u4ee3 LLM\u3002\u6b64\u5916\uff0c\u5b83\u8fd8\u5b9e\u73b0\u4e86\u4e00\u79cd\u65b0\u7684\u8ba1\u7b97\u8303\u5f0f\uff0c\u5e76\u4e3a\u8bbe\u8ba1\u9488\u5bf9 1 \u4f4d LLM \u4f18\u5316\u7684\u7279\u5b9a\u786c\u4ef6\u6253\u5f00\u4e86\u5927\u95e8\u3002<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n
    \n
  1. <\/li>\n<\/ol>\n\n\n\n

    1. 1 \u4f4d LLM \u65f6\u4ee3(The Era of 1-bit LLMs)<\/strong><\/h2>\n\n\n\n

    \u8fd1\u5e74\u6765\uff0c\u4eba\u5de5\u667a\u80fd\u9886\u57df\u7684\u5927\u578b\u8bed\u8a00\u6a21\u578b \uff08LLM\uff09 \u7684\u89c4\u6a21\u548c\u529f\u80fd\u8fc5\u901f\u589e\u957f\u3002\u8fd9\u4e9b\u6a21\u578b\u5728\u5e7f\u6cdb\u7684\u81ea\u7136\u8bed\u8a00\u5904\u7406\u4efb\u52a1\u4e2d\u8868\u73b0\u51fa\u4e86\u5353\u8d8a\u7684\u6027\u80fd\uff0c\u4f46\u5b83\u4eec\u4e0d\u65ad\u6269\u5927\u7684\u89c4\u6a21\u7ed9\u90e8\u7f72\u5e26\u6765\u4e86\u6311\u6218\uff0c\u5e76\u5f15\u53d1\u4e86\u4eba\u4eec\u5bf9\u9ad8\u80fd\u8017\u5bf9\u73af\u5883\u548c\u7ecf\u6d4e\u5f71\u54cd\u7684\u62c5\u5fe7\u3002 \u89e3\u51b3\u8fd9\u4e9b\u6311\u6218\u7684\u4e00\u79cd\u65b9\u6cd5\u662f\u4f7f\u7528\u8bad\u7ec3\u540e\u91cf\u5316\u6765\u521b\u5efa\u7528\u4e8e\u63a8\u7406\u7684\u4f4e\u4f4d\u6a21\u578b[24<\/a>,5<\/a>,2<\/a>,18<\/a>].\u8fd9\u79cd\u6280\u672f\u964d\u4f4e\u4e86\u6743\u91cd\u548c\u6fc0\u6d3b\u7684\u7cbe\u5ea6\uff0c\u5927\u5927\u964d\u4f4e\u4e86 LLM \u7684\u5185\u5b58\u548c\u8ba1\u7b97\u8981\u6c42\u3002\u8d8b\u52bf\u662f\u4ece 16 \u4f4d\u8f6c\u5411\u66f4\u4f4e\u7684\u4f4d\uff0c\u4f8b\u5982 4 \u4f4d\u53d8\u4f53[5<\/a>,9<\/a>].\u7136\u800c\uff0c\u8bad\u7ec3\u540e\u91cf\u5316\u662f\u6b21\u4f18\u7684\uff0c\u5c3d\u7ba1\u5b83\u88ab\u5e7f\u6cdb\u7528\u4e8e\u5de5\u4e1a LLM\u3002<\/p>\n\n\n\n

    \u6700\u8fd1\u5173\u4e8e 1 \u4f4d\u6a21\u578b\u67b6\u6784\uff08\u5982 BitNet\uff09\u7684\u5de5\u4f5c[23<\/a>]\uff0c\u4e3a\u964d\u4f4e LLM \u7684\u6210\u672c\u540c\u65f6\u4fdd\u6301\u5176\u6027\u80fd\u63d0\u51fa\u4e86\u4e00\u4e2a\u6709\u524d\u9014\u7684\u65b9\u5411\u3002Vanilla LLM \u662f 16 \u4f4d\u6d6e\u70b9\u503c\uff08\u5373 FP16 \u6216 BF16\uff09\uff0c\u4efb\u4f55 LLM \u7684\u5927\u90e8\u5206\u90fd\u662f\u77e9\u9635\u4e58\u6cd5\u3002\u56e0\u6b64\uff0c\u4e3b\u8981\u7684\u8ba1\u7b97\u6210\u672c\u6765\u81ea\u6d6e\u70b9\u52a0\u6cd5\u548c\u4e58\u6cd5\u8fd0\u7b97\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0cBitNet \u7684\u77e9\u9635\u4e58\u6cd5\u4ec5\u6d89\u53ca\u6574\u6570\u52a0\u6cd5\uff0c\u8fd9\u4e3a LLM \u8282\u7701\u4e86\u6570\u91cf\u7ea7\u7684\u80fd\u6e90\u6210\u672c\u3002\u7531\u4e8e\u8bb8\u591a\u82af\u7247\u8ba1\u7b97\u6027\u80fd\u7684\u6839\u672c\u9650\u5236\u662f\u529f\u8017\uff0c\u56e0\u6b64\u8282\u80fd\u4e5f\u53ef\u4ee5\u8f6c\u5316\u4e3a\u66f4\u5feb\u7684\u8ba1\u7b97\u3002<\/p>\n\n\n\n

    \u9664\u4e86\u8ba1\u7b97\u4e4b\u5916\uff0c\u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d\uff0c\u5c06\u6a21\u578b\u53c2\u6570\u4eceDRAM\u4f20\u8f93\u5230\u7247\u4e0a\u52a0\u901f\u5668\uff08\u4f8b\u5982SRAM\uff09\u7684\u5b58\u50a8\u5668\u7684\u8fc7\u7a0b\u53ef\u80fd\u975e\u5e38\u6602\u8d35\u3002\u6709\u4eba\u8bd5\u56fe\u6269\u5927SRAM\u4ee5\u63d0\u9ad8\u541e\u5410\u91cf\uff0c\u4f46\u8fd9\u4f1a\u5e26\u6765\u6bd4DRAM\u9ad8\u5f97\u591a\u7684\u6210\u672c\u3002\u4e0e\u5168\u7cbe\u5ea6\u6a21\u578b\u76f8\u6bd4\uff0c\u4ece\u5bb9\u91cf\u548c\u5e26\u5bbd\u7684\u89d2\u5ea6\u6765\u770b\uff0c1 \u4f4d LLM \u7684\u5185\u5b58\u5360\u7528\u8981\u4f4e\u5f97\u591a\u3002\u8fd9\u53ef\u4ee5\u663e\u8457\u964d\u4f4e\u4ece DRAM \u52a0\u8f7d\u6743\u91cd\u7684\u6210\u672c\u548c\u65f6\u95f4\uff0c\u4ece\u800c\u5b9e\u73b0\u66f4\u5feb\u3001\u66f4\u9ad8\u6548\u7684\u63a8\u7406\u3002<\/p>\n\n\n\n

    \u5728\u8fd9\u9879\u5de5\u4f5c\u4e2d\uff0c\u6211\u4eec\u5f15\u5165\u4e86\u4e00\u4e2a\u91cd\u8981\u7684 1 \u4f4d LLM \u53d8\u4f53\uff0c\u79f0\u4e3a BitNet b1.58<\/strong>\uff0c\u5176\u4e2d\u6bcf\u4e2a\u53c2\u6570\u90fd\u662f\u4e09\u5143\u7684\uff0c\u53d6\u503c\u4e3a {-1\uff0c 0\uff0c 1}\u3002\u6211\u4eec\u5728\u539f\u6765\u7684 1 \u4f4d BitNet \u4e2d\u6dfb\u52a0\u4e86\u4e00\u4e2a\u989d\u5916\u7684\u503c 0\uff0c\u4ece\u800c\u5728\u4e8c\u8fdb\u5236\u7cfb\u7edf\u4e2d\u4ea7\u751f 1.58 \u4f4d\u3002BitNet b1.58 \u4fdd\u7559\u4e86\u539f\u59cb 1 \u4f4d BitNet \u7684\u6240\u6709\u4f18\u70b9\uff0c\u5305\u62ec\u5176\u65b0\u7684\u8ba1\u7b97\u8303\u5f0f\uff0c\u8be5\u8303\u5f0f\u51e0\u4e4e\u4e0d\u9700\u8981\u77e9\u9635\u4e58\u6cd5\u8fd0\u7b97\uff0c\u5e76\u4e14\u53ef\u4ee5\u9ad8\u5ea6\u4f18\u5316\u3002\u6b64\u5916\uff0c\u4e0eFP16 LLM\u57fa\u7ebf\u76f8\u6bd4\uff0c\u5b83\u5177\u6709\u4e0e\u539f\u59cb1\u4f4dBitNet\u76f8\u540c\u7684\u80fd\u8017\uff0c\u5e76\u4e14\u5728\u5185\u5b58\u6d88\u8017\uff0c\u541e\u5410\u91cf\u548c\u5ef6\u8fdf\u65b9\u9762\u6548\u7387\u66f4\u9ad8\u3002\u6b64\u5916\uff0cBitNet b1.58 \u8fd8\u5177\u6709\u4e24\u4e2a\u989d\u5916\u7684\u4f18\u52bf\u3002\u9996\u5148\uff0c\u7531\u4e8e\u5176\u5bf9\u7279\u5f81\u8fc7\u6ee4\u7684\u660e\u786e\u652f\u6301\uff0c\u5176\u5efa\u6a21\u80fd\u529b\u66f4\u5f3a\uff0c\u901a\u8fc7\u5728\u6a21\u578b\u6743\u91cd\u4e2d\u5305\u542b 0 \u6765\u5b9e\u73b0\uff0c\u8fd9\u53ef\u4ee5\u663e\u7740\u63d0\u9ad8 1 \u4f4d LLM \u7684\u6027\u80fd\u3002 \u5176\u6b21\uff0c\u6211\u4eec\u7684\u5b9e\u9a8c\u8868\u660e\uff0cBitNet b1.58 \u5728\u56f0\u60d1\u5ea6\u548c\u6700\u7ec8\u4efb\u52a1\u6027\u80fd\u65b9\u9762\u90fd\u53ef\u4ee5\u5339\u914d\u5168\u7cbe\u5ea6\uff08\u5373 FP16\uff09\u57fa\u7ebf\uff0c \u4ece 3B \u5927\u5c0f\u5f00\u59cb\uff0c\u5f53\u4f7f\u7528\u76f8\u540c\u7684\u914d\u7f6e\uff08\u4f8b\u5982\uff0c\u6a21\u578b\u5927\u5c0f\u3001\u8bad\u7ec3\u4ee4\u724c\u7b49\uff09\u65f6\u3002<\/p>\n\n\n\n

    2. BitNet b1.58<\/strong><\/h2>\n\n\n\n

    BitNet b1.58 \u57fa\u4e8e BitNet \u67b6\u6784\uff0c\u5b83\u662f\u4e00\u4e2a\u53d6\u4ee3 nn \u7684 Transformer\u3002\u7ebf\u6027\u4e0e<\/em> BitLinear<\/em>\u3002\u5b83\u662f\u4ece\u5934\u5f00\u59cb\u8bad\u7ec3\u7684\uff0c\u5177\u6709 1.58 \u4f4d\u6743\u91cd\u548c 8 \u4f4d\u6fc0\u6d3b\u3002\u4e0e\u539f\u6765\u7684BitNet\u76f8\u6bd4\uff0c\u5b83\u5f15\u5165\u4e86\u4e00\u4e9b\u4fee\u6539\uff0c\u6211\u4eec\u603b\u7ed3\u5982\u4e0b\u3002<\/p>\n\n\n\n

    2.1 \u91cf\u5316\u51fd\u6570(Quantization Function)<\/strong><\/h3>\n\n\n\n

    \u4e3a\u4e86\u5c06\u6743\u91cd\u9650\u5236\u4e3a -1\u30010 \u6216 +1\uff0c\u6211\u4eec\u91c7\u7528\u4e86 absmean<\/em> \u91cf\u5316\u51fd\u6570\u3002\u5b83\u9996\u5148\u6309\u5176\u5e73\u5747\u7edd\u5bf9\u503c\u7f29\u653e\u6743\u91cd\u77e9\u9635\uff0c\u7136\u540e\u5c06\u6bcf\u4e2a\u503c\u56db\u820d\u4e94\u5165\u5230 {-1\uff0c 0\uff0c +1} \u4e4b\u95f4\u6700\u63a5\u8fd1\u7684\u6574\u6570\uff1a<\/p>\n\n\n\n

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

    \u6fc0\u6d3b\u7684\u91cf\u5316\u51fd\u6570\u9075\u5faa BitNet \u4e2d\u7684\u76f8\u540c\u5b9e\u73b0\uff0c\u53ea\u662f\u6211\u4eec\u4e0d\u4f1a\u5c06\u975e\u7ebf\u6027\u51fd\u6570\u4e4b\u524d\u7684\u6fc0\u6d3b\u7f29\u653e\u5230\u8303\u56f4[0,Qb<\/em><\/sub>].\u76f8\u53cd\uff0c\u6fc0\u6d3b\u5c06\u5168\u90e8\u7f29\u653e\u5230[\u2212Qb<\/em><\/sub>,Qb<\/em><\/sub>]\u6bcf\u4e2a\u4ee4\u724c\u6765\u6446\u8131\u96f6\u70b9\u91cf\u5316\u3002\u8fd9\u5bf9\u4e8e\u5b9e\u73b0\u548c\u7cfb\u7edf\u7ea7\u4f18\u5316\u6765\u8bf4\u90fd\u66f4\u52a0\u65b9\u4fbf\u548c\u7b80\u5355\uff0c\u540c\u65f6\u5728\u6211\u4eec\u7684\u5b9e\u9a8c\u4e2d\u5bf9\u6027\u80fd\u7684\u5f71\u54cd\u53ef\u4ee5\u5ffd\u7565\u4e0d\u8ba1\u3002<\/p>\n\n\n\n

    2.2 \u7c7b\u4f3c LLaMA \u7684\u7ec4\u4ef6(LLaMA-alike Components)<\/strong><\/h3>\n\n\n\n

    LLaMA\u7684\u67b6\u6784[19<\/a>,20<\/a>]\u4e00\u76f4\u662f\u5f00\u6e90 LLM \u4e8b\u5b9e\u4e0a\u7684\u652f\u67f1\u3002\u4e3a\u4e86\u62e5\u62b1\u5f00\u6e90\u793e\u533a\uff0c\u6211\u4eec\u7684 BitNet b1.58 \u8bbe\u8ba1\u91c7\u7528\u4e86\u7c7b\u4f3c LLaMA \u7684\u7ec4\u4ef6\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u5b83\u4f7f\u7528 RMSNorm[27<\/a>]\u3001\u65af\u5a01\u683c\u5362(SwiGLU)[16<\/a>]\u3001\u65cb\u8f6c\u5d4c\u5165(rotary embedding)[14<\/a>]\uff0c\u5e76\u6d88\u9664\u6240\u6709\u504f\u89c1\u3002\u901a\u8fc7\u8fd9\u79cd\u65b9\u5f0f\uff0cBitNet b1.58 \u53ef\u4ee5\u96c6\u6210\u5230\u6d41\u884c\u7684\u5f00\u6e90\u8f6f\u4ef6\uff08\u4f8b\u5982\uff0cHuggingface\u3001vLLM[8<\/a>]\u548c llama.cpp1<\/sup>\uff09\u4ee5\u6700\u5c0f\u7684\u52aa\u529b\u3002<\/p>\n\n\n\n

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

    \u8868 1\uff1aBitNet b1.58 \u548c LLaMA LLM \u7684\u56f0\u60d1\u5ea6\u4ee5\u53ca\u6210\u672c\u3002<\/p>\n\n\n\n

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

    \u88682\uff1aBitNet b1.58 \u548c LLaMA LLM \u5728\u6700\u7ec8\u4efb\u52a1\u4e0a\u7684\u96f6\u70b9\u7cbe\u5ea6\u3002<\/p>\n\n\n\n

    3. \u7ed3\u679c(Results)<\/strong><\/h2>\n\n\n\n

    \u6211\u4eec\u5c06 BitNet b1.58 \u4e0e\u6211\u4eec\u590d\u5236\u7684\u5404\u79cd\u5c3a\u5bf8\u7684 FP16 LLaMA LLM \u8fdb\u884c\u4e86\u6bd4\u8f83\u3002\u4e3a\u4e86\u786e\u4fdd\u516c\u5e73\u7684\u6bd4\u8f83\uff0c\u6211\u4eec\u5728 RedPajama \u6570\u636e\u96c6\u4e0a\u9884\u8bad\u7ec3\u4e86\u6a21\u578b[4<\/a>]1000\u4ebf\u4e2a\u4ee3\u5e01\u3002\u6211\u4eec\u8bc4\u4f30\u4e86\u5305\u62ecARC-Easy\u5728\u5185\u7684\u4e00\u7cfb\u5217\u8bed\u8a00\u4efb\u52a1\u7684\u96f6\u6837\u672c\u6027\u80fd[25<\/a>]\u3001ARC-Challenge[25<\/a>]\uff0c Hellaswag[26<\/a>]\uff0c Winogrande [15<\/a>], PIQA [1<\/a>], OpenbookQA [10<\/a>], and BoolQ [3<\/a>].\u6211\u4eec\u8fd8\u62a5\u544a\u4e86WikiText2\u4e0a\u7684\u9a8c\u8bc1\u56f0\u60d1[11<\/a>]\u548c C4<\/sub>[13<\/a>]\u6570\u636e\u3002<\/p>\n\n\n\n

    \u6211\u4eec\u6bd4\u8f83\u4e86 LLaMA LLM \u548c BitNet b1.58 \u7684\u8fd0\u884c\u65f6 GPU \u5185\u5b58\u548c\u5ef6\u8fdf\u3002\u4f7f\u7528FasterTransformer\u6d4b\u91cf\u7ed3\u679c2<\/sup>\u4ee3\u7801\u5e93\uff0c\u9488\u5bf9 GPU \u8bbe\u5907\u4e0a\u7684 LLM \u63a8\u7406\u5ef6\u8fdf\u8fdb\u884c\u4e86\u4f18\u5316\u3002Ladder \u7684 2 \u4f4d\u5185\u6838[22<\/a>]\u8fd8\u96c6\u6210\u4e86 BitNet b1.58\u3002\u6211\u4eec\u62a5\u544a\u4e86\u6bcf\u4e2a\u8f93\u51fa\u4ee4\u724c\u7684\u65f6\u95f4\uff0c\u56e0\u4e3a\u5b83\u662f\u63a8\u7406\u7684\u4e3b\u8981\u6210\u672c\u3002<\/p>\n\n\n\n

    \u8868 1<\/a> \u603b\u7ed3\u4e86 BitNet b1.58 \u548c LLaMA LLM \u7684\u56f0\u60d1\u548c\u6210\u672c\u3002\u5b83\u8868\u660e\uff0cBitNet b1.58 \u5728 3B \u6a21\u578b\u5927\u5c0f\u4e0b\u5f00\u59cb\u4e0e\u5168\u7cbe\u5ea6 LLaMA LLM \u76f8\u5339\u914d\uff0c\u540c\u65f6\u901f\u5ea6\u63d0\u9ad8\u4e86 2.71 \u500d\uff0c\u4f7f\u7528\u7684 GPU \u5185\u5b58\u51cf\u5c11\u4e86 3.55 \u500d\u3002\u7279\u522b\u662f\uff0c\u5177\u6709 3.9B \u6a21\u578b\u5927\u5c0f\u7684 BitNet b1.58 \u901f\u5ea6\u5feb 2.4 \u500d\uff0c\u5185\u5b58\u6d88\u8017\u51cf\u5c11 3.32 \u500d\uff0c\u4f46\u6027\u80fd\u660e\u663e\u4f18\u4e8e LLaMA LLM 3B\u3002<\/p>\n\n\n\n

    \u8868 2<\/a> \u62a5\u544a\u4e86\u6700\u7ec8\u4efb\u52a1\u7684\u96f6\u70b9\u7cbe\u5ea6\u7684\u8be6\u7ec6\u7ed3\u679c\u3002\u6211\u4eec\u9075\u5faa\u4e86 lm-evaluation-harness<\/em> \u7684\u7ba1\u90533<\/sup>\u6267\u884c\u8bc4\u4f30\u3002\u7ed3\u679c\u8868\u660e\uff0c\u968f\u7740\u6a21\u578b\u89c4\u6a21\u7684\u589e\u52a0\uff0cBitNet b1.58\u548cLLaMA LLM\u4e4b\u95f4\u7684\u6027\u80fd\u5dee\u8ddd\u9010\u6e10\u7f29\u5c0f\u3002\u66f4\u91cd\u8981\u7684\u662f\uff0cBitNet b1.58 \u53ef\u4ee5\u5339\u914d\u4ece 3B \u5927\u5c0f\u5f00\u59cb\u7684\u5168\u7cbe\u5ea6\u57fa\u7ebf\u7684\u6027\u80fd\u3002\u4e0e\u5bf9\u56f0\u60d1\u7684\u89c2\u5bdf\u7c7b\u4f3c\uff0c\u6700\u7ec8\u4efb\u52a1\u7ed3\u679c\u663e\u793a\uff0cBitNet b1.58 3.9B \u4ee5\u66f4\u4f4e\u7684\u5185\u5b58\u548c\u5ef6\u8fdf\u6210\u672c\u4f18\u4e8e LLaMA LLM 3B\u3002\u8fd9\u8868\u660e BitNet b1.58 \u662f\u5bf9\u6700\u5148\u8fdb\u7684 LLM \u6a21\u578b\u7684\u5e15\u7d2f\u6258\u6539\u8fdb\u3002<\/p>\n\n\n\n

    3.1 \u5185\u5b58\u548c\u5ef6\u8fdf(Memory and Latency)<\/strong><\/h3>\n\n\n\n
    \"\"<\/figure>\n\n\n\n

    \u56fe 2\uff1aBitNet b1.58 \u89e3\u7801\u5ef6\u8fdf\uff08\u5de6\uff09\u548c\u5185\u5b58\u6d88\u8017\uff08\u53f3\uff09\u968f\u6a21\u578b\u5927\u5c0f\u7684\u53d8\u5316<\/p>\n\n\n\n

    \u6211\u4eec\u8fdb\u4e00\u6b65\u5c06\u6a21\u578b\u5c3a\u5bf8\u6269\u5927\u5230 7B\u300113B \u548c 70B\uff0c\u5e76\u8bc4\u4f30\u4e86\u6210\u672c\u3002\u56fe 2<\/a> \u8bf4\u660e\u4e86\u5ef6\u8fdf\u548c\u5185\u5b58\u7684\u8d8b\u52bf\uff0c\u663e\u793a\u968f\u7740\u6a21\u578b\u5927\u5c0f\u7684\u6269\u5c55\uff0c\u52a0\u901f\u4e5f\u4f1a\u589e\u52a0\u3002\u7279\u522b\u662f\uff0cBitNet b1.58 70B \u6bd4 LLaMA LLM \u57fa\u7ebf\u5feb 4.1 \u500d\u3002\u8fd9\u662f\u56e0\u4e3a nn.\u7ebf\u6027<\/em>\u968f\u7740\u6a21\u578b\u5927\u5c0f\u7684\u589e\u957f\u800c\u589e\u957f\u3002\u5185\u5b58\u6d88\u8017\u9075\u5faa\u7c7b\u4f3c\u7684\u8d8b\u52bf\uff0c\u56e0\u4e3a\u5d4c\u5165\u4fdd\u6301\u5168\u7cbe\u5ea6\uff0c\u5e76\u4e14\u5bf9\u4e8e\u8f83\u5927\u7684\u578b\u53f7\uff0c\u5176\u5185\u5b58\u6bd4\u4f8b\u8f83\u5c0f\u3002\u5ef6\u8fdf\u548c\u5185\u5b58\u90fd\u662f\u7528 2 \u4f4d\u5185\u6838\u6d4b\u91cf\u7684\uff0c\u56e0\u6b64\u4ecd\u6709\u4f18\u5316\u7684\u7a7a\u95f4\u4ee5\u8fdb\u4e00\u6b65\u964d\u4f4e\u6210\u672c\u3002<\/p>\n\n\n\n

    3.2 \u80fd\u6e90(Energy)<\/strong><\/h3>\n\n\n\n

    \u6211\u4eec\u8fd8\u4f30\u8ba1\u4e86 BitNet b1.58 \u548c LLaMA LLM \u7684\u7b97\u672f\u8fd0\u7b97\u80fd\u8017\u3002 \u6211\u4eec\u4e3b\u8981\u5173\u6ce8\u77e9\u9635\u4e58\u6cd5\u7684\u8ba1\u7b97\uff0c\u56e0\u4e3a\u5b83 \u5bf9 LLM \u6210\u672c\u7684\u8d21\u732e\u6700\u5927\u3002 \u56fe 3<\/a> \u8bf4\u660e\u4e86\u80fd\u6e90\u6210\u672c\u7684\u6784\u6210\u3002BitNet b1.58 \u7684\u5927\u90e8\u5206\u662f INT8 \u52a0\u6cd5\u8ba1\u7b97\uff0c\u800c LLaMA LLM \u7531 FP16 \u52a0\u6cd5\u548c FP16 \u4e58\u6cd5\u7ec4\u6210\u3002\u6839\u636e\u80fd\u91cf\u6a21\u578b[7<\/a>,28<\/a>]\uff0cBitNet b1.58 \u4e3a 7nm \u82af\u7247\u4e0a\u7684\u77e9\u9635\u4e58\u6cd5\u8282\u7701\u4e86 71.4 \u500d\u7684\u7b97\u672f\u8fd0\u7b97\u80fd\u8017\u3002\u6211\u4eec\u8fdb\u4e00\u6b65\u62a5\u544a\u4e86\u5177\u6709 512 \u4e2a\u4ee3\u5e01\u7684\u6a21\u578b\u7684\u7aef\u5230\u7aef\u80fd\u6e90\u6210\u672c\u3002\u6211\u4eec\u7684\u7ed3\u679c\u8868\u660e\uff0c\u968f\u7740\u6a21\u578b\u89c4\u6a21\u7684\u6269\u5927\uff0c\u4e0eFP16 LLaMA LLM\u57fa\u7ebf\u76f8\u6bd4\uff0cBitNet b1.58\u5728\u80fd\u8017\u65b9\u9762\u7684\u6548\u7387\u8d8a\u6765\u8d8a\u9ad8\u3002\u8fd9\u662f\u56e0\u4e3a nn.\u7ebf\u6027<\/em>\u968f\u7740\u6a21\u578b\u5927\u5c0f\u7684\u589e\u957f\u800c\u589e\u957f\uff0c\u800c\u5bf9\u4e8e\u8f83\u5927\u7684\u6a21\u578b\uff0c\u5176\u4ed6\u7ec4\u4ef6\u7684\u6210\u672c\u66f4\u5c0f\u3002<\/p>\n\n\n\n

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

    \u56fe 3\uff1aBitNet b1.58 \u4e0e LLaMA LLM \u5728 7nm \u5de5\u827a\u8282\u70b9\u4e0a\u7684\u80fd\u8017\u76f8\u6bd4\u3002\u5de6\u8fb9\u662f\u7b97\u672f\u8fd0\u7b97\u80fd\u91cf\u7684\u5206\u91cf\u3002\u53f3\u8fb9\u662f\u4e0d\u540c\u578b\u53f7\u5c3a\u5bf8\u7684\u7aef\u5230\u7aef\u80fd\u6e90\u6210\u672c<\/p>\n\n\n\n

    3.3 \u541e\u5410\u91cf(Throughput)<\/strong><\/h3>\n\n\n\n

    \"\"<\/p>\n\n\n\n

    \u88683\uff1aBitNet b1.58 70B \u548c LLaMA LLM 70B \u4e4b\u95f4\u7684\u541e\u5410\u91cf\u6bd4\u8f83<\/p>\n\n\n\n

    \u6211\u4eec\u4f7f\u7528\u6d41\u6c34\u7ebf\u5e76\u884c\u6027\u6bd4\u8f83\u4e86 BitNet b1.58 \u548c LLaMA LLM \u4e0e\u4e24\u4e2a 80GB A100 \u5361\u4e0a 70B \u53c2\u6570\u7684\u541e\u5410\u91cf[6<\/a>]\u4ee5\u4fbf LLaMA LLM 70B \u53ef\u4ee5\u5728\u8bbe\u5907\u4e0a\u8fd0\u884c\u3002\u6211\u4eec\u589e\u52a0\u4e86\u6279\u5904\u7406\u5927\u5c0f\uff0c\u76f4\u5230\u8fbe\u5230 GPU \u5185\u5b58\u9650\u5236\uff0c\u5e8f\u5217\u957f\u5ea6\u4e3a 512\u3002\u8868 3<\/a> \u663e\u793a\uff0cBitNet b1.58 70B \u53ef\u4ee5\u652f\u6301\u9ad8\u8fbe 11 \u500d\u7684 LLaMA LLM \u6279\u91cf\u5927\u5c0f\uff0c\u4ece\u800c\u5c06\u541e\u5410\u91cf\u63d0\u9ad8 8.9 \u500d\u3002<\/p>\n\n\n\n

    BitNet b1.58<\/strong> \u5728\u6a21\u578b\u6027\u80fd\u548c\u63a8\u7406\u6210\u672c\u65b9\u9762\u5b9e\u73b0\u4e86\u65b0\u7684\u6269\u5c55\u5b9a\u5f8b<\/strong>\u3002\u4f5c\u4e3a\u53c2\u8003\uff0c\u6211\u4eec\u53ef\u4ee5\u6839\u636e\u56fe 2<\/a> \u548c\u56fe 3<\/a> \u4e2d\u7684\u7ed3\u679c\uff0c\u5728 1.58 \u4f4d\u548c 16 \u4f4d\u7684\u4e0d\u540c\u6a21\u578b\u5927\u5c0f\u4e4b\u95f4\u83b7\u5f97\u4ee5\u4e0b\u7b49\u4ef7\u6027\u3002<\/p>\n\n\n\n