{"id":15166,"date":"2024-12-09T20:44:25","date_gmt":"2024-12-09T12:44:25","guid":{"rendered":"https:\/\/www.aisharenet.com\/?p=15166"},"modified":"2024-12-09T20:44:25","modified_gmt":"2024-12-09T12:44:25","slug":"kaiyuan16bxiaomoxingai","status":"publish","type":"post","link":"https:\/\/www.kdjingpai.com\/en\/kaiyuan16bxiaomoxingai\/","title":{"rendered":"\u5f00\u6e901.6B\u5c0f\u6a21\u578b\u300c\u5c0f\u72d0\u72f8\u300d\uff0c\u8868\u73b0\u8d85\u540c\u7c7b\u6a21\u578bQwen\u548cGemma"},"content":{"rendered":"<p><img loading=\"lazy\" decoding=\"async\" class=\"\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/88b17dbc8e55a10.jpg\" width=\"760\" height=\"443\" \/><\/p>\n<p>&nbsp;<\/p>\n<p>\u81ea\u4ece <a href=\"https:\/\/www.kdjingpai.com\/chatgpt-6\/\">Chatgpt<\/a> \u8bde\u751f\u4ee5\u6765\uff0cLLM\uff08\u5927\u8bed\u8a00\u6a21\u578b\uff09\u7684\u53c2\u6570\u91cf\u4f3c\u4e4e\u5c31\u6210\u4e3a\u4e86\u5404\u4e2a\u516c\u53f8\u7684\u7ade\u8d5b\u6307\u6807\u3002GPT-1 \u53c2\u6570\u91cf\u4e3a 1.17 \u4ebf\uff08117M\uff09\uff0c\u800c\u5b83\u7684\u7b2c\u56db\u4ee3 GPT-4 \u53c2\u6570\u91cf\u5df2\u7ecf\u5237\u65b0\u5230\u4e86 1.8 \u4e07\u4ebf\uff081800B\uff09\u3002<\/p>\n<p>\u50cf\u5176\u4ed6 LLM \u6a21\u578b\u4f8b\u5982 Bloom(1760 \u4ebf\uff0c176B)\u548c Chinchilla\uff08700 \u4ebf\uff0c70B\uff09\u7684\u53c2\u6570\u91cf\u4e5f\u5728\u4e0d\u65ad\u98d9\u5347\u3002\u53c2\u6570\u7684\u6570\u91cf\u76f4\u63a5\u5f71\u54cd\u4e86\u6a21\u578b\u7684\u6027\u80fd\u548c\u80fd\u529b\uff0c\u66f4\u591a\u7684\u53c2\u6570\u610f\u5473\u7740\u6a21\u578b\u80fd\u591f\u5904\u7406\u66f4\u590d\u6742\u7684\u8bed\u8a00\u6a21\u5f0f\uff0c\u7406\u89e3\u66f4\u4e30\u5bcc\u7684\u4e0a\u4e0b\u6587\u4fe1\u606f\uff0c\u5e76\u5728\u591a\u79cd\u4efb\u52a1\u4e0a\u8868\u73b0\u51fa\u66f4\u9ad8\u7684\u667a\u80fd\u6c34\u5e73\u3002<\/p>\n<p>\u4f46\u662f\uff0c\u8fd9\u4e9b\u5de8\u91cf\u53c2\u6570\u4e5f\u4f1a\u76f4\u63a5\u5f71\u54cd LLM \u7684\u8bad\u7ec3\u6210\u672c\u548c\u5f00\u53d1\u73af\u5883\uff0c\u4e5f\u9650\u5236\u4e86\u5927\u591a\u6570\u666e\u901a\u7814\u7a76\u516c\u53f8\u5bf9\u4e8e LLM \u7684\u63a2\u7d22\uff0c\u5bfc\u81f4\u5927\u8bed\u8a00\u6a21\u578b\u9010\u6e10\u6210\u4e3a\u4e86\u5927\u516c\u53f8\u4e4b\u95f4\u7684\u519b\u5907\u7ade\u8d5b\u3002<\/p>\n<p>\u8fd1\u65e5\uff0c\u65b0\u5174 AI \u516c\u53f8 TensorOpera \u53d1\u5e03\u4e86<strong>\u5f00\u6e90\u5c0f\u8bed\u8a00\u6a21\u578b FOX<\/strong>\uff0c\u5411\u4e1a\u5185\u8bc1\u660e\u4e86\u5c0f\u8bed\u8a00\u6a21\u578b\uff08SLM\uff09\u4e5f\u53ef\u4ee5\u5728\u667a\u80fd\u4f53\u9886\u57df\u5c55\u73b0\u8db3\u591f\u7684\u5b9e\u529b\u3002<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/3f7f60a9a397828.png\" width=\"620\" height=\"649\" \/><\/p>\n<p>&nbsp;<\/p>\n<p>FOX \u662f\u4e00\u4e2a<strong>\u4e13\u4e3a\u4e91\u8ba1\u7b97\u548c\u8fb9\u7f18\u8ba1\u7b97\u8bbe\u8ba1\u7684\u5c0f\u578b\u8bed\u8a00\u6a21\u578b<\/strong>\u3002\u4e0e\u52a8\u8f84\u4e0a\u767e\u4ebf\u53c2\u6570\u7684\u5927\u8bed\u8a00\u6a21\u578b\u4e0d\u540c\uff0cFOX <strong>\u4ec5\u6709 16 \u4ebf\u53c2\u6570<\/strong>\uff0c\u5374\u80fd\u5728\u591a\u9879\u4efb\u52a1\u4e2d\u5c55\u73b0\u51fa\u60ca\u4eba\u7684\u6027\u80fd\u3002<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/d586ff25f872ec4.png\" width=\"789\" height=\"524\" \/><\/p>\n<p>&nbsp;<\/p>\n<p><strong>\u8bba\u6587\u9898\u76ee\uff1a<\/strong><\/p>\n<p><em>FOX-1 TECHNICAL REPORT<\/em><\/p>\n<p><strong>\u8bba\u6587\u94fe\u63a5\uff1a<\/strong><\/p>\n<p>https:\/\/arxiv.org\/abs\/2411.05281<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/3655e9204946bbd.png\" width=\"758\" height=\"323\" \/><\/p>\n<p>&nbsp;<\/p>\n<h2>TensorOpera \u662f\u8c01<\/h2>\n<p>TensorOpera \u662f\u4e00\u5bb6\u4f4d\u4e8e\u52a0\u5dde\u7845\u8c37\u7684\u521b\u65b0\u4eba\u5de5\u667a\u80fd\u516c\u53f8\u3002\u4ed6\u4eec\u4e4b\u524d\u5f00\u53d1\u4e86 TensorOpera\u00ae AI Platform \u751f\u6210\u578b AI \u751f\u6001\u7cfb\u7edf\u548c TensorOpera\u00ae FedML \u8054\u90a6\u5b66\u4e60\u548c\u5206\u6790\u5e73\u53f0\u3002\u516c\u53f8\u540d\u79f0 TensorOpera, \u00a0\u662f\u6280\u672f\u548c\u827a\u672f\u7684\u7ed3\u5408\uff0c\u8c61\u5f81\u7740 GenAI \u6700\u7ec8\u5b9e\u73b0\u591a\u6a21\u6001\u548c\u591a\u6a21\u578b\u590d\u5408 AI \u7cfb\u7edf\u7684\u53d1\u5c55\u8d8b\u3002<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/49feed792989d33.png\" width=\"753\" height=\"136\" \/><\/p>\n<p>TensorOpera \u7684\u8054\u5408\u521b\u59cb\u4eba\u517c CEO Jared Kaplan \u535a\u58eb\u8868\u793a\uff1a&#8221;FOX \u6a21\u578b\u7684\u8bbe\u8ba1\u521d\u8877\u662f\u4e3a\u4e86\u5728\u4fdd\u6301\u9ad8\u6027\u80fd\u7684\u540c\u65f6\uff0c\u5927\u5e45\u964d\u4f4e\u8ba1\u7b97\u8d44\u6e90\u9700\u6c42\u3002\u8fd9\u4e0d\u4ec5\u8ba9 AI \u6280\u672f\u66f4\u52a0\u4eb2\u6c11\uff0c\u4e5f\u4e3a\u4f01\u4e1a\u964d\u4f4e\u4e86\u4f7f\u7528\u95e8\u69db\u3002&#8221;<\/p>\n<p>&nbsp;<\/p>\n<h2>Fox \u6a21\u578b\u662f\u600e\u4e48\u505a\u7684<\/h2>\n<p>\u4e3a\u4e86\u901a\u8fc7\u8f83\u5c0f\u7684\u53c2\u6570\u91cf\u8fbe\u5230\u4e0e LLM \u76f8\u540c\u7684\u6548\u679c\uff0cFox-1 \u6a21\u578b<strong>\u4ec5\u91c7\u7528\u89e3\u7801\u5668<\/strong>\u67b6\u6784\uff0c\u5e76\u5f15\u5165\u4e86\u5404\u79cd\u6539\u8fdb\u548c\u91cd\u65b0\u8bbe\u8ba1\u4ee5\u83b7\u5f97\u66f4\u597d\u7684\u6027\u80fd\u3002\u5176\u5305\u62ec<\/p>\n<p>\u2460 <strong>\u7f51\u7edc\u5c42\u6570<\/strong>\uff1a\u5728\u6a21\u578b\u67b6\u6784\u8bbe\u8ba1\u4e2d\uff0c\u66f4\u5bbd\u4e14\u66f4\u6d45\u7684\u795e\u7ecf\u7f51\u7edc\u62e5\u6709\u66f4\u597d\u7684\u8bb0\u5fc6\u80fd\u529b\uff0c\u800c\u66f4\u6df1\u4e14\u66f4\u7626\u7684\u7f51\u7edc\u5219\u5448\u73b0\u51fa\u66f4\u5f3a\u7684\u63a8\u7406\u80fd\u529b\u3002\u6839\u636e\u8fd9\u4e00\u539f\u5219\uff0cFox-1 \u4f7f\u7528\u7684\u67b6\u6784\u6bd4\u5927\u591a\u6570\u73b0\u4ee3 SLMs \u66f4\u6df1\u3002\u5177\u4f53\u6765\u8bf4\uff0cFox-1 \u7531 32 \u4e2a\u81ea\u6ce8\u610f\u529b\u5c42\u7ec4\u6210\uff0c\u6bd4 Gemma-2B\uff0818 \u5c42\uff09\u6df1 78%\uff0c\u6bd4 StableLM-2-1.6B\uff0824 \u5c42\uff09\u548c Qwen1.5-1.8B\uff0824 \u5c42\uff09\u6df1 33%\u3002<\/p>\n<p>\u2461 <strong>\u5171\u4eab\u5d4c\u5165<\/strong>\uff1aFox-1 \u91c7\u7528 2,048 \u7684\u9690\u85cf\u7ef4\u5ea6\u53bb\u6784\u5efa\u5171\u8ba1 256,000 \u7684\u8bcd\u6c47\u8868\uff0c\u5176\u6570\u636e\u91cf\u5927\u7ea6\u6709 5 \u4ebf\u53c2\u6570\u3002\u66f4\u5927\u7684\u6a21\u578b\u901a\u5e38\u5bf9\u8f93\u5165\u5c42\uff08\u8bcd\u6c47\u8868\u5230\u5d4c\u5165\u8868\u8fbe\uff09\u548c\u8f93\u51fa\u5c42\uff08\u5d4c\u5165\u8868\u8fbe\u5230\u8bcd\u6c47\u8868\uff09\u4f7f\u7528\u5355\u72ec\u7684\u5d4c\u5165\u5c42\u3002\u5bf9\u4e8e Fox-1\uff0c\u4ec5\u5d4c\u5165\u5c42\u5c31\u9700\u8981 10 \u4ebf\u53c2\u6570\u3002\u4e3a\u4e86\u51cf\u5c11\u603b\u53c2\u6570\u6570\u91cf\uff0c\u5171\u4eab\u8f93\u5165\u548c\u8f93\u51fa\u5d4c\u5165\u5c42\uff0c\u53ef\u4ee5\u6700\u5927\u9650\u5ea6\u5730\u63d0\u9ad8\u6743\u91cd\u5229\u7528\u7387\u3002<\/p>\n<p>\u2462 <strong>\u9884\u5f52\u4e00\u5316<\/strong>\uff1aFox-1 \u5bf9\u6bcf\u4e2a\u53d8\u6362\u5c42\u7684\u8f93\u5165\u4f7f\u7528 RMSNorm \u8fdb\u884c\u5f52\u4e00\u5316\u3002RMSNorm \u662f\u73b0\u4ee3\u5927\u578b\u8bed\u8a00\u6a21\u578b\u4e2d\u9884\u5f52\u4e00\u5316\u7684\u9996\u9009\uff0c\u5b83\u6bd4 LayerNorm \u8868\u73b0\u51fa\u66f4\u597d\u7684\u6548\u7387\u3002<\/p>\n<p>\u2463 <strong>\u65cb\u8f6c\u5f0f\u4f4d\u7f6e\u7f16\u7801\uff08RoPE\uff09<\/strong>\uff1aFox-1 \u9ed8\u8ba4\u63a5\u53d7\u6700\u591a 8K \u957f\u5ea6\u7684 input token\u3002\u4e3a\u4e86\u63d0\u9ad8\u66f4\u957f\u4e0a\u4e0b\u6587\u7a97\u53e3\u7684\u6027\u80fd\uff0cFox-1 \u91c7\u7528\u4e86\u65cb\u8f6c\u5f0f\u4f4d\u7f6e\u7f16\u7801\uff0c\u5176\u4e2d \u03b8 \u8bbe\u7f6e\u4e3a 10,000\uff0c\u4ee5\u4fbf\u4e8e\u7f16\u7801 <a href=\"https:\/\/www.kdjingpai.com\/tokenization\/\">token<\/a> \u4e4b\u95f4\u7684\u76f8\u5bf9\u4f4d\u7f6e\u4f9d\u8d56\u6027\u3002<\/p>\n<p>\u2464 <strong>\u5206\u7ec4\u67e5\u8be2\u6ce8\u610f\u529b\uff08GQA\uff09<\/strong>\uff1a\u5206\u7ec4\u67e5\u8be2\u6ce8\u610f\u529b\u5c06\u591a\u5934\u6ce8\u610f\u529b\u5c42\u7684\u67e5\u8be2\u5934\u5206\u4e3a\u7ec4\uff0c\u6bcf\u4e2a\u7ec4\u5171\u4eab\u76f8\u540c\u7684\u4e00\u7ec4\u952e\u503c\u5934\u3002Fox-1 \u914d\u5907 4 \u4e2a\u952e\u503c\u5934\u548c 16 \u4e2a\u6ce8\u610f\u529b\u5934\uff0c\u4ee5\u63d0\u9ad8\u8bad\u7ec3\u548c\u63a8\u7406\u901f\u5ea6\uff0c\u5e76\u51cf\u5c11\u5185\u5b58\u4f7f\u7528\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/32fa172a62c39b9.png\" \/><\/p>\n<p>\u9664\u4e86\u6a21\u578b\u7ed3\u6784\u6539\u826f\u4e4b\u5916\uff0c<strong>FOX-1 \u4e5f\u5728\u5206\u8bcd\uff08Tokenization\uff09\u548c\u8bad\u7ec3\u4e0a\u8fdb\u884c\u6539\u8fdb<\/strong>\u3002<\/p>\n<p><strong>\u5206\u8bcd\u65b9\u9762<\/strong>\uff0cFox-1 \u91c7\u7528\u57fa\u4e8e SentencePiece \u7684 Gemma \u5206\u8bcd\u5668\uff0c\u5b83\u63d0\u4f9b\u4e86 256K \u7684\u8bcd\u6c47\u8868\u5927\u5c0f\u3002\u589e\u52a0\u8bcd\u6c47\u8868\u5927\u5c0f\u81f3\u5c11\u6709\u4e24\u4e2a\u4e3b\u8981\u597d\u5904\u3002\u9996\u5148\uff0c\u7531\u4e8e\u6bcf\u4e2a token \u7f16\u7801\u4e86\u66f4\u5bc6\u96c6\u7684\u4fe1\u606f\uff0c\u4e0a\u4e0b\u6587\u7684\u9690\u85cf\u4fe1\u606f\u957f\u5ea6\u5f97\u4ee5\u5ef6\u957f\u3002\u4f8b\u5982\uff0c\u5927\u5c0f\u4e3a 26 \u7684\u8bcd\u6c47\u8868\u53ea\u80fd\u5728[a-z]\u4e2d\u7f16\u7801\u4e00\u4e2a\u5b57\u7b26\uff0c\u4f46\u5927\u5c0f\u4e3a 262 \u7684\u8bcd\u6c47\u8868\u53ef\u4ee5\u540c\u65f6\u7f16\u7801\u4e24\u4e2a\u5b57\u6bcd\uff0c\u8fd9\u4f7f\u5f97\u5728\u56fa\u5b9a\u957f\u5ea6\u7684 token \u4e2d\u53ef\u4ee5\u8868\u793a\u66f4\u957f\u7684\u5b57\u7b26\u4e32\u3002\u5176\u6b21\uff0c\u66f4\u5927\u7684\u8bcd\u6c47\u8868\u5927\u5c0f\u51cf\u5c11\u4e86\u672a\u77e5\u5355\u8bcd\u6216\u77ed\u8bed\u7684\u6982\u7387\uff0c\u4ece\u800c\u5728\u5b9e\u8df5\u4e2d\u5b9e\u73b0\u4e86\u66f4\u597d\u7684\u4e0b\u6e38\u4efb\u52a1\u6027\u80fd\u3002Fox-1 \u91c7\u7528\u7684\u5927\u8bcd\u6c47\u8868\u5bf9\u4e8e\u7ed9\u5b9a\u7684\u6587\u672c\u8bed\u6599\u5e93\u4ea7\u751f\u8f83\u5c11\u7684 token\uff0c\u8fd9\u53ef\u4ee5\u4ea7\u751f\u66f4\u597d\u7684\u63a8\u7406\u6027\u80fd\u3002<\/p>\n<p>Fox-1 \u7684<strong>\u9884\u8bad\u7ec3\u6570\u636e<\/strong>\u6765\u6e90\u4e8e Redpajama\u3001SlimPajama\u3001Dolma\u3001Pile \u548c Falcon \u6570\u636e\u96c6\uff0c\u5171\u8ba1 3 \u4e07\u4ebf\u4e2a\u6587\u672c\u6570\u636e\u3002\u4e3a\u4e86\u7f13\u89e3\u56e0\u5176\u6ce8\u610f\u529b\u673a\u5236\u5bfc\u81f4\u7684\u957f\u5e8f\u5217\u7684\u9884\u8bad\u7ec3\u6548\u7387\u4f4e\u4e0b\uff0cFox-1 \u5728\u9884\u8bad\u7ec3\u9636\u6bb5\u5f15\u5165\u4e86\u4e00\u4e2a<strong>\u4e09\u9636\u6bb5\u7684\u8bfe\u7a0b\u5b66\u4e60\u7b56\u7565<\/strong>\uff0c\u5176\u4e2d\u8bad\u7ec3\u6837\u672c\u7684 chunk \u957f\u5ea6\u9010\u6e10\u4ece 2K \u589e\u52a0\u5230 8K\uff0c\u4ee5\u5c0f\u6210\u672c\u786e\u4fdd\u957f\u4e0a\u4e0b\u6587\u80fd\u529b\u3002\u4e3a\u4e86\u4e0e\u4e09\u9636\u6bb5\u8bfe\u7a0b\u9884\u8bad\u7ec3\u7ba1\u9053\u4fdd\u6301\u4e00\u81f4\uff0cFox-1 \u5c06\u539f\u59cb\u6570\u636e\u91cd\u65b0\u7ec4\u7ec7\u6210\u4e09\u4e2a\u4e0d\u540c\u7684\u96c6\u5408\uff0c\u5305\u62ec\u65e0\u76d1\u7763\u548c\u6307\u4ee4\u8c03\u4f18\u6570\u636e\u96c6\uff0c\u4ee5\u53ca\u4ee3\u7801\u3001\u7f51\u7edc\u5185\u5bb9\u3001\u6570\u5b66\u548c\u79d1\u5b66\u6587\u6863\u7b49\u4e0d\u540c\u9886\u57df\u7684\u6570\u636e\u3002<\/p>\n<p><strong>Fox-1 \u7684\u8bad\u7ec3\u53ef\u4ee5\u5206\u4e3a\u4e09\u4e2a\u9636\u6bb5<\/strong>\u3002<\/p>\n<ul>\n<li>\u7b2c\u4e00\u9636\u6bb5\u5305\u62ec\u6574\u4e2a\u9884\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u7ea6 39% \u7684\u603b\u6570\u636e\u6837\u672c\uff0c\u5176\u4e2d 1.05 \u4e07\u4ebf token \u7684\u6570\u636e\u96c6\u88ab\u5206\u5272\u6210\u957f\u5ea6\u4e3a 2,000 \u7684\u6837\u672c\uff0cbatch size \u5927\u5c0f\u4e3a 2M\u3002\u5728\u6b64\u9636\u6bb5\u4f7f\u7528\u4e86 2,000epoch \u7684\u7ebf\u6027\u9884\u70ed\u3002<\/li>\n<li>\u7b2c\u4e8c\u9636\u6bb5\u5305\u62ec\u7ea6 59% \u7684\u6837\u672c\uff0c\u6709 1.58 \u4e07\u4ebf token\uff0c\u5e76\u5c06 chunk \u957f\u5ea6\u4ece 2K \u589e\u52a0\u5230 4K \u548c 8K\u3002\u5b9e\u9645\u7684 chunk \u957f\u5ea6\u56e0\u4e0d\u540c\u6570\u636e\u6e90\u800c\u5f02\u3002\u8003\u8651\u5230\u7b2c\u4e8c\u9636\u6bb5\u82b1\u8d39\u7684\u65f6\u95f4\u6700\u957f\uff0c\u4e14\u6d89\u53ca\u4e0d\u540c\u6570\u636e\u96c6\u7684\u4e0d\u540c\u6765\u6e90\uff0cbatch size \u4e5f\u589e\u52a0\u5230 4M \u4ee5\u63d0\u9ad8\u8bad\u7ec3\u6548\u7387\u3002<\/li>\n<li>\u6700\u540e\u5728\u7b2c\u4e09\u9636\u6bb5\uff0cFox \u6a21\u578b\u4f7f\u7528 62 \u4ebf token\uff08\u7ea6\u603b\u91cf\u7684 0.02%\uff09\u7684\u9ad8\u8d28\u91cf\u6570\u636e\u8fdb\u884c\u8bad\u7ec3\uff0c\u4e3a\u4e0d\u540c\u7684\u4e0b\u6e38\u4efb\u52a1\u80fd\u529b\u6253\u4e0b\u57fa\u7840\uff0c\u5982\u6307\u4ee4\u9075\u5faa\u3001\u95f2\u804a\u3001\u7279\u5b9a\u9886\u57df\u7684\u95ee\u7b54\u7b49\u3002<\/li>\n<\/ul>\n<p>&nbsp;<\/p>\n<h2>Fox-1 \u8868\u73b0\u5982\u4f55<\/h2>\n<p>\u548c\u5176\u4ed6 SLM \u6a21\u578b(Gemma-2B, Qwen1.5-1.8B, StableLM-2-1.6B \u00a0\u548c OpenELM1.1B)\u76f8\u6bd4\uff0cFOX-1 \u5728 ARC Challenge (25-shot)\uff0c HellaSwag (10-shot)\uff0c TruthfulQA(0-shot)\uff0cMMLU (5-shot)\uff0cWinogrande (5-shot)\uff0cGSM8k (5-shot)<strong>\u516d\u9879\u4efb\u52a1\u7684 benchmark \u7684\u5e73\u5747\u5206\u6570\u6700\u9ad8\uff0c\u4e14\u5728 GSM8k \u4e0a\u4f18\u52bf\u660e\u663e\u3002<\/strong><\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/b80b55fd660f973.png\" \/><\/p>\n<p>\u9664\u6b64\u4e4b\u5916\uff0cTensorOpera \u8fd8\u8bc4\u4f30\u4e86 Fox-1\u3001Qwen1.5-1.8B \u548c Gemma-2B \u4f7f\u7528 <a href=\"https:\/\/www.kdjingpai.com\/vllm\/\">vLLM<\/a> \u4e0e TensorOpera \u670d\u52a1\u5e73\u53f0\u5728\u5355\u4e2a NVIDIA H100 \u4e0a\u7684\u7aef\u5230\u7aef\u7684\u63a8\u7406\u6548\u7387\u3002<\/p>\n<p>Fox-1 \u5b9e\u73b0\u4e86\u6bcf\u79d2\u8d85\u8fc7 200 \u4e2a token \u7684\u541e\u5410\u91cf\uff0c\u8d85\u8fc7\u4e86 Gemma-2B\uff0c\u5e76\u5728\u76f8\u540c\u7684\u90e8\u7f72\u73af\u5883\u4e2d\u4e0e Qwen1.5-1.8B \u76f8\u5339\u914d\u3002\u5728 BF16 \u7cbe\u5ea6\u4e0b\uff0cFox-1 \u4ec5\u9700\u8981 3703MiB \u7684 GPU \u5185\u5b58\uff0c\u800c Qwen1.5-1.8B\u3001StableLM-2-1.6B \u548c Gemma-2B \u5206\u522b\u9700\u8981 4739MiB\u30013852MiB \u548c 5379MiB\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/12\/c4364056d0dd8cc.png\" \/><\/p>\n<p>&nbsp;<\/p>\n<h2>\u53c2\u6570\u867d\u5c0f\uff0c\u4f9d\u65e7\u6709\u7ade\u4e89\u529b<\/h2>\n<p>\u76ee\u524d\u5404\u5bb6 AI \u516c\u53f8\u90fd\u5728\u5927\u8bed\u8a00\u6a21\u578b\u4e0a\u5f00\u5c55\u7ade\u4e89\uff0c\u800c TensorOpera \u53e6\u8f9f\u8e4a\u5f84\uff0c\u5728 SLM \u9886\u57df\u8fdb\u884c\u7a81\u7834\uff0c\u4ec5\u7528 1.6B \u5b9e\u73b0\u4e86\u4e0e LLM \u76f8\u4f3c\u7684\u6548\u679c\uff0c\u5728\u5404\u79cd\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u8868\u73b0\u51fa\u8272\u3002<\/p>\n<p>\u5373\u4f7f\u5728\u6709\u9650\u7684\u6570\u636e\u8d44\u6e90\u4e0b\uff0cTensorOpera \u4e5f\u53ef\u4ee5\u9884\u8bad\u7ec3\u51fa\u5177\u6709\u7ade\u4e89\u529b\u8868\u73b0\u7684\u8bed\u8a00\u6a21\u578b\uff0c\u4e3a\u5176\u4ed6 AI \u516c\u53f8\u7684\u7814\u53d1\u63d0\u4f9b\u4e86\u4e00\u79cd\u65b0\u601d\u8def\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>&nbsp; \u81ea\u4ece Chatgpt \u8bde\u751f\u4ee5\u6765\uff0cLLM\uff08\u5927\u8bed\u8a00\u6a21\u578b\uff09\u7684\u53c2\u6570\u91cf\u4f3c\u4e4e\u5c31\u6210\u4e3a\u4e86\u5404\u4e2a\u516c\u53f8\u7684\u7ade\u8d5b\u6307\u6807\u3002GPT-1 \u53c2\u6570\u91cf\u4e3a 1.17 \u4ebf\uff08117M\uff09\uff0c\u800c\u5b83\u7684\u7b2c\u56db\u4ee3 GPT-4 \u53c2\u6570\u91cf\u5df2\u7ecf\u5237\u65b0\u5230\u4e86 1.8 \u4e07\u4ebf\uff081800B\uff09\u3002 \u50cf\u5176\u4ed6 L&#8230;<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[46],"tags":[],"class_list":["post-15166","post","type-post","status-publish","format-standard","hentry","category-news"],"_links":{"self":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/posts\/15166","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/comments?post=15166"}],"version-history":[{"count":0,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/posts\/15166\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/media?parent=15166"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/categories?post=15166"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/tags?post=15166"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}