{"id":415330,"date":"2024-06-18T22:37:44","date_gmt":"2024-06-18T14:37:44","guid":{"rendered":"https:\/\/www.idc.net\/help\/?p=415330"},"modified":"2024-06-18T22:38:56","modified_gmt":"2024-06-18T14:38:56","slug":"%e7%ab%af%e4%be%a7%e6%9c%80%e5%bc%ba%ef%bc%8cmeta%e7%94%b0%e6%b8%8a%e6%a0%8b%e7%ad%89%e4%ba%ba%e5%8d%b710%e4%ba%bf%e4%bb%a5%e4%b8%8b%e5%8f%82%e6%95%b0%e5%b0%8f%e6%a8%a1%e5%9e%8b%ef%bc%8clecun%ef%bc%9a","status":"publish","type":"post","link":"https:\/\/idc.net\/help\/415330\/","title":{"rendered":"\u7aef\u4fa7\u6700\u5f3a\uff0cMeta\u7b49\u4eba\u537710\u4ebf\u4ee5\u4e0b\u53c2\u6570\u5c0f\u6a21\u578b\uff0cLeCun\uff1a\u5c0f\u6280\u5de7\u542f\u52a8"},"content":{"rendered":"

\u300c\u5728\u79fb\u52a8\u8bbe\u5907\u4e0a\u8fd0\u884c LLM\uff1f\u53ef\u80fd\u9700\u8981 Meta \u7684\u4e00\u4e9b\u6280\u5de7\u3002\u300d\u521a\u521a\uff0c\u56fe\u7075\u5956\u5f97\u4e3b Yann LeCun \u5728\u4e2a\u4eba\u793e\u4ea4\u5e73\u53f0\u8868\u793a\u3002<\/p>\n

\u4ed6\u6240\u5ba3\u4f20\u7684\u8fd9\u9879\u7814\u7a76\u6765\u81ea Meta \u6700\u65b0\u8bba\u6587\u300a MobileLLM: Optimizing Sub-billion Parameter Language Models for On-Device Use Cases \u300b\uff0c\u5728\u4f17\u591a\u4f5c\u8005\u4e2d\u4e5f\u6709\u6211\u4eec\u719f\u6089\u7684\u6765\u81ea Meta FAIR \u3002<\/p>\n

\u300c\u6211\u4eec\u7684 MobileLLM \u9884\u8bad\u7ec3\u6a21\u578b\uff08125M\/350M\uff09\uff0c\u6027\u80fd\u8fbe\u5230 SoTA\uff0c\u7279\u522b\u662f\u5728\u804a\u5929 \/ API \u8c03\u7528\u65b9\u9762\u8868\u73b0\u51fa\u8272\u3002\u6b64\u5916\uff0c\u672c\u5de5\u4f5c\u4e2d\u7684\u4e00\u4e2a\u6709\u8da3\u7814\u7a76\u662f\u8de8 Transformer \u5c42\u7684\u6743\u91cd\u5171\u4eab\uff0c\u8fd9\u6837\u4e0d\u4ec5\u8282\u7701\u4e86\u53c2\u6570\uff0c\u8fd8\u51cf\u5c11\u4e86\u63a8\u7406\u8fc7\u7a0b\u4e2d\u7684\u5ef6\u8fdf\u3002\u300d<\/p>\n

\u73b0\u9636\u6bb5\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u5df2\u7ecf\u6e17\u900f\u5230\u4eba\u7c7b\u751f\u6d3b\u7684\u5404\u4e2a\u65b9\u9762\uff0c\u5c24\u5176\u662f\u4ee5 ChatGPT \u7b49\u4e3a\u4ee3\u8868\u7684\u6a21\u578b\uff0c\u8fd9\u7c7b\u7814\u7a76\u4e3b\u8981\u5728\u4e91\u73af\u5883\u4e2d\u8fd0\u884c\u3002<\/p>\n

\u7136\u800c\u9886\u5148\u7684\u6a21\u578b\u5982 ChatGPT4 \u7684\u53c2\u6570\u91cf\u5df2\u7ecf\u8d85\u8fc7\u4e86 1 \u4e07\u4ebf\u3002\u6211\u4eec\u8bbe\u60f3\u8fd9\u6837\u4e00\u4e2a\u573a\u666f\uff0c\u8fd9\u4e2a\u573a\u666f\u5e7f\u6cdb\u4f9d\u8d56 LLM\uff0c\u4e0d\u4ec5\u7528\u4e8e\u524d\u7aef\u7684\u4f1a\u8bdd\u754c\u9762\uff0c\u4e5f\u7528\u4e8e\u540e\u7aef\u64cd\u4f5c\uff0c\u5982\u63a8\u8350\u7cfb\u7edf\uff0c\u8986\u76d6\u4eba\u7c7b\u7ea6 5% \u7684\u65f6\u95f4\u3002\u5728\u8fd9\u4e00\u5047\u8bbe\u573a\u666f\u4e2d\uff0c\u5047\u5982\u4ee5 GPT-4 \u6bcf\u79d2\u5904\u7406 50 \u4e2a token \u7684\u901f\u7387\u6765\u8ba1\u7b97\uff0c\u5219\u9700\u8981\u90e8\u7f72\u5927\u7ea6\u4e00\u4ebf\u4e2a H100 GPU\uff0c\u6bcf\u4e2a GPU \u7684\u8ba1\u7b97\u80fd\u529b\u4e3a 60 TFLOPs\/s\u3002\u8fd9\u79cd\u8ba1\u7b97\u89c4\u6a21\uff0c\u8fd8\u4e0d\u5305\u62ec\u901a\u4fe1\u548c\u6570\u636e\u4f20\u8f93\u7684\u5f00\u9500\uff0c\u5c31\u5df2\u7ecf\u4e0e 160 \u4e2a Meta \u89c4\u6a21\u7684\u516c\u53f8\u76f8\u5f53\u3002\u968f\u4e4b\u800c\u6765\u7684\u80fd\u6e90\u6d88\u8017\u548c\u4e8c\u6c27\u5316\u78b3\u6392\u653e\u5c06\u5e26\u6765\u5de8\u5927\u7684\u73af\u5883\u6311\u6218\u3002<\/p>\n

\u56e0\u6b64\uff0c\u6700\u597d\u7684\u89e3\u51b3\u65b9\u6848\u662f\u7f29\u5c0f LLM \u7684\u89c4\u6a21\u3002<\/p>\n

\u6b64\u5916\uff0c\u5728\u5f53\u524d\u7684\u79fb\u52a8\u6280\u672f\u9886\u57df\uff0c\u7531\u4e8e\u4e3b\u5185\u5b58\uff08DRAM\uff09\u5bb9\u91cf\u7684\u9650\u5236\uff0c\u5c06\u50cf LLaMAv2 7B \u8fd9\u6837\u7684 LLM \u4e0e 8 \u4f4d\u6743\u91cd\u6574\u5408\u8d77\u6765\u4ee3\u4ef7\u8fc7\u9ad8\u3002\u79fb\u52a8\u8bbe\u5907\u4e2d\u666e\u904d\u7684\u5185\u5b58\u5c42\u7ed3\u6784\u5982\u56fe 2 \u6240\u793a\u3002\u968f\u7740 DRAM \u5bb9\u91cf\u4ece iPhone 15 \u7684 6GB \u5230 Google Pixel 8 Pro \u7684 12GB \u4e0d\u7b49\uff0c\u4e00\u4e2a\u79fb\u52a8\u5e94\u7528\u4e0d\u5e94\u8d85\u8fc7 DRAM \u7684 10%\uff0c\u56e0\u4e3a DRAM \u9700\u8981\u4e0e\u64cd\u4f5c\u7cfb\u7edf\u548c\u5176\u4ed6\u5e94\u7528\u7a0b\u5e8f\u5171\u4eab\u3002\u8fd9\u4e00\u8981\u6c42\u4fc3\u8fdb\u4e86\u90e8\u7f72\u5c0f\u4e8e\u5341\u4ebf\u53c2\u6570 LLM \u66f4\u8fdb\u4e00\u6b65\u7684\u7814\u7a76\u3002<\/p>\n

\u57fa\u4e8e\u4e0a\u8ff0\u8003\u91cf\uff0c\u6765\u81ea Meta \u7684\u7814\u7a76\u8005\u4e13\u6ce8\u4e8e\u8bbe\u8ba1\u53c2\u6570\u5c11\u4e8e\u5341\u4ebf\u7684\u9ad8\u8d28\u91cf LLM\uff0c\u8fd9\u662f\u5728\u79fb\u52a8\u7aef\u90e8\u7f72 LLM \u6bd4\u8f83\u597d\u7684\u89e3\u51b3\u65b9\u6848\u3002<\/p>\n

\u4e0e\u5f3a\u8c03\u6570\u636e\u548c\u53c2\u6570\u6570\u91cf\u5728\u51b3\u5b9a\u6a21\u578b\u8d28\u91cf\u65b9\u9762\u7684\u5173\u952e\u4f5c\u7528\u7684\u666e\u904d\u89c2\u70b9\u76f8\u53cd\uff0cMeta \u5f3a\u8c03\u4e86\u6a21\u578b\u67b6\u6784\u5bf9\u5c11\u4e8e\u5341\u4ebf\uff08sub-billion\uff09\u89c4\u6a21 LLM \u7684\u91cd\u8981\u6027\u3002<\/p>\n

\u57fa\u4e8e\u6df1\u800c\u7a84\u7684\u67b6\u6784\uff0c\u52a0\u4e0a\u5d4c\u5165\u5171\u4eab\u548c\u5206\u7ec4\u67e5\u8be2\u6ce8\u610f\u529b\u673a\u5236\uff0cMeta \u5efa\u7acb\u4e86\u4e00\u4e2a\u5f3a\u5927\u7684\u57fa\u7ebf\u7f51\u7edc\uff0c\u79f0\u4e3a MobileLLM\uff0c\u4e0e\u4e4b\u524d\u7684 125M\/350M \u6700\u5148\u8fdb\u6a21\u578b\u76f8\u6bd4\uff0c\u5176\u51c6\u786e\u7387\u663e\u8457\u63d0\u9ad8\u4e86 2.7%\/4.3% \u3002\u8fd9\u4e5f\u8bf4\u660e\u4e86\u4e0e\u7f29\u653e\u5b9a\u5f8b\uff08scaling law\uff09\u76f8\u53cd\uff0c\u8be5\u7814\u7a76\u8bc1\u660e\u5bf9\u4e8e\u5c0f\u578b LLM \u6765\u8bf4\u6df1\u5ea6\u6bd4\u5bbd\u5ea6\u66f4\u91cd\u8981\uff0c\u4e00\u4e2a\u6df1\u800c\u7a84\u7684\u6a21\u578b\u7ed3\u6784\u5728\u6355\u83b7\u62bd\u8c61\u6982\u5ff5\u65b9\u9762\u66f4\u4e3a\u51fa\u8272\u3002<\/p>\n

\u6b64\u5916\uff0cMeta \u8fd8\u63d0\u51fa\u4e86\u4e00\u79cd\u53ca\u65f6\u9010\u5757\u6743\u91cd\u5171\u4eab\uff08 immediate block-wise weight sharing\uff09\u65b9\u6cd5\uff0c\u8be5\u65b9\u6cd5\u4e0d\u4f1a\u589e\u52a0\u6a21\u578b\u5927\u5c0f\uff0c\u6240\u5f97\u6a21\u578b\u8868\u793a\u4e3a MobileLLM-LS\uff0c\u5176\u51c6\u786e\u7387\u6bd4 MobileLLM 125M\/350M \u8fdb\u4e00\u6b65\u63d0\u9ad8\u4e86 0.7%\/0.8%\u3002\u6b64\u5916\uff0c\u5728\u4e0b\u6e38\u4efb\u52a1\u4e2d\uff0c\u4f8b\u5982 Chat \u548c API \u8c03\u7528\uff0cMobileLLM \u6a21\u578b\u5bb6\u65cf\u663e\u8457\u4f18\u4e8e\u540c\u7b49\u89c4\u6a21\u7684\u6a21\u578b\u3002\u5728 API \u8c03\u7528\u4efb\u52a1\u4e2d\uff0c\u4e0e\u89c4\u6a21\u8f83\u5927\u7684 LLaMA-v2 7B \u76f8\u6bd4\uff0cMobileLLM \u751a\u81f3\u5b9e\u73b0\u4e86\u76f8\u5ab2\u7f8e\u7684\u5206\u6570\u3002<\/p>\n

\u770b\u5230\u8fd9\u9879\u7814\u7a76\u540e\uff0c\u7f51\u53cb\u7eb7\u7eb7\u8868\u793a\u300c\u6211\u4eec\u5e94\u8be5\u5411 Meta \u81f4\u656c\uff0c\u5f88\u9ad8\u5174\u770b\u5230\u8fd9\u4e2a\u9886\u57df\u7684\u6d3b\u8dc3\u73a9\u5bb6\u3002\u8be5\u673a\u6784\u901a\u8fc7\u4f7f\u7528\u4f4e\u4e8e 10 \u4ebf\u53c2\u6570\u7684\u6a21\u578b\uff0c\u5e76\u4e14 350M 8 \u4f4d\u6a21\u578b\u7684\u80fd\u6e90\u6d88\u8017\u4ec5\u4e3a 0.035 J\/token \u00a0\uff0c\u8981\u662f\u90e8\u7f72\u5728 iPhone \u4e0a\u7684\u8bdd\uff0c\u53ef\u4ee5\u652f\u6301\u7528\u6237\u4e00\u6574\u5929\u7684\u4f1a\u8bdd\u4f7f\u7528\u3002\u300d<\/p>\n

\u6539\u8fdb\u5341\u4ebf\u4ee5\u4e0b\u53c2\u6570\u89c4\u6a21\u7684 LLM \u8bbe\u8ba1<\/h4>\n

\u7814\u7a76\u8005\u4ecb\u7ecd\u4e86\u4ece\u5341\u4ebf\u4ee5\u4e0b\u53c2\u6570\u89c4\u6a21\u7684\u57fa\u7ebf\u6a21\u578b\u5230\u65b0\u7684 SOTA \u6a21\u578b\u7684\u6f14\u8fdb\u4e4b\u8def\uff08\u5982\u4e0b\u56fe 3 \u6240\u793a\uff09\u3002\u4ed6\u4eec\u5206\u522b\u7814\u7a76\u4e86 125M \u548c 350M \u53c2\u6570\u89c4\u6a21\u7684\u6a21\u578b\uff0c\u5e76\u5728\u8fd9\u4e24\u4e2a\u89c4\u6a21\u4e0b\u5c55\u793a\u4e86\u4e00\u81f4\u7684\u6539\u8fdb\u3002\u5bf9\u4e8e\u6a21\u578b\u5c3a\u5bf8\u6210\u4e3a\u4e3b\u8981\u5236\u7ea6\u56e0\u7d20\u7684\u8bbe\u5907\u7528\u4f8b\u800c\u8a00\uff0c\u5982\u4f55\u6709\u6548\u5730\u5206\u914d\u6709\u9650\u7684\u6743\u91cd\u53c2\u6570\u53d8\u5f97\u6bd4\u4ee5\u5f80\u66f4\u52a0\u91cd\u8981\u3002<\/p>\n

\u7814\u7a76\u8005\u9996\u5148\u901a\u8fc7\u6d4b\u8bd5\u56db\u79cd\u6709\u76ca\u4e8e\u5341\u4ebf\u4ee5\u4e0b\u89c4\u6a21 LLM \u7684\u6a21\u578b\u8bbe\u8ba1\u65b9\u6cd5\uff0c\u63d0\u51fa\u4e86\u4e00\u4e2a\u540d\u4e3aMobileLLM \u7684\u5f3a\u5927\u57fa\u7ebf\u6a21\u578b\u3002\u8fd9\u56db\u79cd\u6a21\u578b\u8bbe\u8ba1\u65b9\u6cd5\u5305\u62ec 1\uff09\u91c7\u7528 SwiGLU FFN\uff0c2\uff09\u5f3a\u5236\u4f7f\u7528\u6df1\u548c\u8584\u7684\u67b6\u6784\uff0c3\uff09\u91cd\u65b0\u5ba1\u89c6\u5d4c\u5165\u5171\u4eab\u65b9\u6cd5\uff0c4\uff09\u5229\u7528\u5206\u7ec4\u67e5\u8be2\u6ce8\u610f\u529b\u3002<\/p>\n

\u63a5\u4e0b\u6765\uff0c\u7814\u7a76\u8005\u5f00\u53d1\u4e86\u4e00\u79cd\u76f4\u63a5\u7684\u9010\u5757\u5c42\u5171\u4eab\u65b9\u6cd5\uff0c\u57fa\u4e8e\u8be5\u65b9\u6cd5\u53ef\u4ee5\u8fdb\u4e00\u6b65\u63d0\u9ad8\u51c6\u786e\u5ea6\uff0c\u800c\u4e0d\u4ea7\u751f\u4efb\u4f55\u989d\u5916\u7684\u5185\u5b58\u5f00\u9500\uff0c\u5e76\u5728\u5185\u5b58\u6709\u9650\u7684 LM \u89e3\u7801\u8fc7\u7a0b\u4e2d\u4ea7\u751f\u5f88\u5c0f\u7684\u5ef6\u8fdf\u5f00\u9500\u3002\u4ed6\u4eec\u5c06\u5177\u6709\u5c42\u5171\u4eab\u7684\u6a21\u578b\u8868\u793a\u4e3a MobileLLM-LS\u3002<\/p>\n

\u8bad\u7ec3\u8bbe\u7f6e<\/strong><\/p>\n

\u7814\u7a76\u8005\u5728 32 \u4e2a A100 GPU \u4e0a\u8fdb\u884c\u5b9e\u9a8c\uff0c\u5176\u4e2d\u6bcf\u4e2a GPU \u7684\u6279\u5927\u5c0f\u4e3a 32\u3002\u4ed6\u4eec\u5728 0.25T \u7684 tokens \u4e0a\u6267\u884c\u4e86 120k \u6b21\u8fed\u4ee3\u7684\u63a2\u7d22\u6027\u5b9e\u9a8c\u3002\u4e0b\u6587\u4e2d\u8868 3 \u548c\u8868 4 \u62a5\u544a\u4e86\u5728 1T \u7684 tokens \u4e0a\u6267\u884c 480k \u6b21\u8fed\u4ee3\u8bad\u7ec3\u7684 top \u6a21\u578b\u3002<\/p>\n

\u5c42\u5171\u4eab<\/strong><\/p>\n

\u5173\u4e8e\u5c42\u6df1\u5ea6\u4e0e\u5bbd\u5ea6\u5f71\u54cd\u7684\u7814\u7a76\u7ed3\u679c\u8868\u660e\uff0c\u66f4\u6df1\u7684\u5c42\u6709\u5229\u4e8e\u5c0f\u578b transformer \u6a21\u578b\u3002\u8fd9\u4fc3\u4f7f\u672c\u6587\u7814\u7a76\u5c42\u5171\u4eab\u4f5c\u4e3a\u589e\u52a0\u9690\u85cf\u5c42\u6570\u91cf\u800c\u4e0d\u589e\u52a0\u5b58\u50a8\u6210\u672c\u7684\u7b56\u7565\u3002\u8fd9\u79cd\u65b9\u6cd5\u5728\u6a21\u578b\u5927\u5c0f\u6210\u4e3a\u4e3b\u8981\u5236\u7ea6\u56e0\u7d20\u7684\u573a\u666f\u4e2d\u5c24\u5176\u6709\u7528\u3002<\/p>\n

\u4ee4\u4eba\u60ca\u8bb6\u7684\u662f\uff0c\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0c\u901a\u8fc7\u7b80\u5355\u5730\u590d\u5236 transformer \u5757\u5c31\u53ef\u4ee5\u63d0\u9ad8\u51c6\u786e\u5ea6\u800c\u65e0\u9700\u4efb\u4f55\u67b6\u6784\u4fee\u6539\u6216\u6269\u5927\u6a21\u578b\u5c3a\u5bf8\u3002\u7814\u7a76\u8005\u8fdb\u4e00\u6b65\u63a2\u7a76\u4e09\u79cd\u4e0d\u540c\u7684\u6743\u91cd\u5171\u4eab\u7b56\u7565\uff0c\u5177\u4f53\u5982\u4e0b\u56fe 6 \u6240\u793a\u3002<\/p>\n

\u4e0b\u8868 2 \u7ed3\u679c\u8868\u660e\uff0c\u91cd\u590d\u5c42\u5171\u4eab\u7b56\u7565\u5728\u7acb\u5373\u5757\u91cd\u590d\u3001\u5168\u9762\u91cd\u590d\uff08repeat all-over\uff09\u548c\u53cd\u5411\u5171\u4eab\u7b56\u7565\u4e2d\u4ea7\u751f\u4e86\u6700\u4f73\u6027\u80fd\u3002<\/p>\n

\u4e0d\u8fc7\uff0c\u8003\u8651\u5230\u786c\u4ef6\u5185\u5b58\u7684\u5c42\u7ea7\u7ed3\u6784\uff08\u5982\u56fe 2\uff09\uff0c\u7528\u4e8e\u8ba1\u7b97\u7684 SRAM \u901a\u5e38\u9650\u5236\u5728\u4e86 20M \u5de6\u53f3\u3002\u8be5\u5bb9\u91cf\u901a\u5e38\u4ec5\u591f\u5bb9\u7eb3\u5355\u4e2a transformer \u5757\u3002\u56e0\u6b64\uff0c\u5c06\u5171\u4eab\u6743\u91cd\u653e\u5165\u7f13\u5b58\u4e2d\u5e76\u7acb\u5373\u8ba1\u7b97\u4e24\u6b21\u5219\u65e0\u9700\u5728 SRAM \u548c DRAM \u4e4b\u95f4\u4f20\u8f93\u6743\u91cd\uff0c\u63d0\u9ad8\u4e86\u81ea\u56de\u5f52\u63a8\u7406\u7684\u6574\u4f53\u6267\u884c\u901f\u5ea6\u3002<\/p>\n

\u7814\u7a76\u8005\u5728\u6a21\u578b\u8bbe\u8ba1\u4e2d\u9009\u62e9\u4e86\u76f4\u63a5\u7684\u5206\u5757\u5171\u4eab\u7b56\u7565\uff0c\u5e76\u5c06\u63d0\u51fa\u7684\u5e26\u6709\u5c42\u5171\u4eab\u7684\u6a21\u578b\u8868\u793a\u4e3a MobileLLM-LS\u3002<\/p>\n

\u5b9e\u9a8c\u7ed3\u679c<\/h4>\n

\u8be5\u7814\u7a76\u8fdb\u884c\u5b9e\u9a8c\u6bd4\u8f83\u4e86\u6a21\u578b\u5728\u96f6\u6837\u672c\uff08zero-shot\uff09\u5e38\u8bc6\u63a8\u7406\u4efb\u52a1\u3001\u95ee\u7b54\u548c\u9605\u8bfb\u7406\u89e3\u4efb\u52a1\u4e0a\u7684\u6027\u80fd\u3002<\/p>\n

\u96f6\u6837\u672c\u5e38\u8bc6\u63a8\u7406\u4efb\u52a1\u7684\u5b9e\u9a8c\u7ed3\u679c\u5982\u4e0b\u8868 3 \u6240\u793a\uff1a<\/p>\n

\u5728\u95ee\u7b54\u548c\u9605\u8bfb\u7406\u89e3\u4efb\u52a1\u4e0a\uff0c\u8be5\u7814\u7a76\u91c7\u7528 TQA \u95ee\u7b54\u57fa\u51c6\u548c RACE \u9605\u8bfb\u7406\u89e3\u57fa\u51c6\u6765\u8bc4\u4f30\u9884\u8bad\u7ec3\u6a21\u578b\uff0c\u5b9e\u9a8c\u7ed3\u679c\u5982\u4e0b\u8868 4 \u6240\u793a\uff1a<\/p>\n

\u4e3a\u4e86\u9a8c\u8bc1\u5c06\u6a21\u578b\u7528\u4e8e\u8bbe\u5907\u4e0a\u5e94\u7528\u7a0b\u5e8f\u7684\u6709\u6548\u6027\uff0c\u8be5\u7814\u7a76\u8bc4\u4f30\u4e86\u6a21\u578b\u5728\u4e24\u4e2a\u5173\u952e\u4efb\u52a1\u4e0a\u7684\u6027\u80fd\uff1a\u804a\u5929\u548c API \u8c03\u7528\u3002<\/p>\n

\u9488\u5bf9\u804a\u5929\u4efb\u52a1\uff0c\u8be5\u7814\u7a76\u5728\u4e24\u4e2a\u57fa\u51c6\u4e0a\u8fdb\u884c\u4e86\u8bc4\u4f30\u5b9e\u9a8c\uff1aAlpacaEval\uff08\u5355\u8f6e\u804a\u5929\u57fa\u51c6\uff09\u548c MT-Bench\uff08\u591a\u8f6e\u804a\u5929\u57fa\u51c6\uff09\uff0c\u5b9e\u9a8c\u7ed3\u679c\u5982\u4e0b\u8868 5 \u6240\u793a\uff1a<\/p>\n

\u5728 API \u8c03\u7528\u65b9\u9762\uff0c\u5982\u4e0b\u8868 6 \u6240\u793a\uff0cMobileLLM-350M \u8868\u73b0\u51fa\u4e0e LLaMA-v2 7B \u76f8\u5f53\u7684 EM_intent \u548c EM_structure\uff0c\u5176\u4e2d EM_intent \u8d8a\u9ad8\uff0c\u8868\u660e\u6a21\u578b\u5bf9\u7528\u6237\u8ba1\u5212\u8c03\u7528 API \u7684\u9884\u6d4b\u5c31\u8d8a\u51c6\u786e\uff0c\u800c EM_structure \u53cd\u6620\u4e86\u9884\u6d4b API \u51fd\u6570\u5185\u5185\u5bb9\u7684\u719f\u7ec3\u7a0b\u5ea6\u3002<\/p>\n

\u8be5\u7814\u7a76\u8fdb\u4e00\u6b65\u5728 MobileLLM \u548c MobileLLM-LS \u6a21\u578b\u4e0a\u9488\u5bf9\u6bcf\u4e2a token \u8fdb\u884c\u6700\u5c0f \/ \u6700\u5927\u8bad\u7ec3\u540e\u91cf\u5316 (PTQ) \u5b9e\u9a8c\uff0c\u6a21\u578b\u5927\u5c0f\u5206\u522b\u4e3a 125M \u548c 350M\uff0c\u5728 0.25T token \u4e0a\u8fdb\u884c\u8bad\u7ec3\uff0c\u5b9e\u9a8c\u7ed3\u679c\u5982\u4e0b\u56fe 7 \u6240\u793a\uff1a<\/p>\n

\u6a21\u578b\u52a0\u8f7d\u3001\u521d\u59cb\u5316\u548c\u6267\u884c\u65f6\u95f4\u5982\u4e0b\u8868 7 \u6240\u793a\uff1a<\/p>\n

\u66f4\u591a\u6280\u672f\u7ec6\u8282\u8bf7\u53c2\u9605\u539f\u8bba\u6587\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

\u300c\u5728\u79fb\u52a8\u8bbe\u5907\u4e0a\u8fd0\u884c LLM\uff1f\u53ef\u80fd\u9700\u8981 Meta \u7684\u4e00\u4e9b\u6280\u5de7\u3002\u300d\u521a\u521a\uff0c\u56fe\u7075\u5956\u5f97\u4e3b Yann LeCun \u5728\u4e2a\u4eba\u793e […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[7023],"tags":[],"class_list":["post-415330","post","type-post","status-publish","format-standard","hentry","category-chatgpt"],"_links":{"self":[{"href":"https:\/\/idc.net\/help\/wp-json\/wp\/v2\/posts\/415330","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/idc.net\/help\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/idc.net\/help\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/idc.net\/help\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/idc.net\/help\/wp-json\/wp\/v2\/comments?post=415330"}],"version-history":[{"count":0,"href":"https:\/\/idc.net\/help\/wp-json\/wp\/v2\/posts\/415330\/revisions"}],"wp:attachment":[{"href":"https:\/\/idc.net\/help\/wp-json\/wp\/v2\/media?parent=415330"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/idc.net\/help\/wp-json\/wp\/v2\/categories?post=415330"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/idc.net\/help\/wp-json\/wp\/v2\/tags?post=415330"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}