{"id":20978,"date":"2025-02-13T10:28:21","date_gmt":"2025-02-13T02:28:21","guid":{"rendered":"https:\/\/www.aisharenet.com\/?p=20978"},"modified":"2025-02-13T10:28:21","modified_gmt":"2025-02-13T02:28:21","slug":"moxinglianghuashishencheng","status":"publish","type":"post","link":"https:\/\/www.kdjingpai.com\/en\/moxinglianghuashishencheng\/","title":{"rendered":"\u6a21\u578b\u91cf\u5316\u662f\u4ec0\u4e48\uff1aFP32, FP16, INT8, INT4 \u6570\u636e\u7c7b\u578b\u8be6\u89e3"},"content":{"rendered":"<p><strong>\u5bfc\u8bfb<\/strong><\/p>\n<p>\u5728\u4eba\u5de5\u667a\u80fd\u79d1\u6280\u7684\u6d69\u701a\u661f\u7a7a\u4e2d\uff0c\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u4ee5\u5176\u5353\u8d8a\u7684\u6027\u80fd\uff0c\u9a71\u52a8\u7740\u8bf8\u591a\u9886\u57df\u7684\u521b\u65b0\u53d1\u5c55\u3002\u7136\u800c\uff0c\u6a21\u578b\u89c4\u6a21\u7684\u6301\u7eed\u81a8\u80c0\uff0c\u5982\u540c\u53cc\u5203\u5251\uff0c\u5728\u63d0\u5347\u6027\u80fd\u7684\u540c\u65f6\uff0c\u4e5f\u5e26\u6765\u4e86\u7b97\u529b\u9700\u6c42\u4e0e\u5b58\u50a8\u538b\u529b\u7684\u5267\u589e\u3002\u7279\u522b\u662f\u5728\u8d44\u6e90\u53d7\u9650\u7684\u5e94\u7528\u573a\u666f\u4e2d\uff0c\u6a21\u578b\u7684\u90e8\u7f72\u548c\u8fd0\u884c\u9762\u4e34\u4e25\u5cfb\u6311\u6218\u3002<\/p>\n<p>\u9762\u5bf9\u8fd9\u4e00\u56f0\u5883\uff0c\u4e00\u79cd\u540d\u4e3a\u201c\u91cf\u5316\u201d\u7684\u6280\u672f\u5e94\u8fd0\u800c\u751f\uff0c\u72b9\u5982\u4e00\u67c4\u7cbe\u5de7\u7684\u624b\u672f\u5200\uff0c\u5728\u6a21\u578b\u7cbe\u5ea6\u53ef\u63a5\u53d7\u7684\u8303\u56f4\u5185\uff0c\u5de7\u5999\u5730\u964d\u4f4e\u6a21\u578b\u5927\u5c0f\uff0c\u63d0\u5347\u8fd0\u7b97\u901f\u5ea6\uff0c\u5e76\u663e\u8457\u51cf\u5c11\u80fd\u8017\u3002\u91cf\u5316\u6280\u672f\uff0c\u80fd\u591f\u5c06\u6a21\u578b\u4e2d\u9ad8\u7cbe\u5ea6\u7684 FP32 \u6570\u636e\uff0c\u8f6c\u6362\u4e3a\u4f4e\u7cbe\u5ea6\u7684 INT4 \u6570\u636e\uff0c\u5b9e\u73b0\u6a21\u578b\u7684\u201c\u7626\u8eab\u201d\u4e0e\u201c\u52a0\u901f\u201d\u3002\u672c\u6587\u5c06\u6df1\u5165\u6d45\u51fa\u5730\u5256\u6790\u91cf\u5316\u6280\u672f\u7684\u539f\u7406\u3001\u65b9\u6cd5\u53ca\u5176\u5728\u6df1\u5ea6\u5b66\u4e60\u9886\u57df\u7684\u5e94\u7528\uff0c\u5373\u4f7f\u662f\u6280\u672f\u521d\u5b66\u8005\uff0c\u4ea6\u80fd\u8f7b\u677e\u7406\u89e3\u5176\u7cbe\u9ad3\u3002<\/p>\n<p>&nbsp;<\/p>\n<h2>1. \u6570\u5b57\u8868\u793a\u57fa\u7840<\/h2>\n<h3>1.1 \u4e8c\u8fdb\u5236\u4e0e\u5341\u8fdb\u5236\u8f6c\u6362<\/h3>\n<p>\u5728\u6570\u5b57\u4e16\u754c\u7684\u57fa\u77f3\u2014\u2014\u8ba1\u7b97\u673a\u79d1\u5b66\u4e2d\uff0c\u6240\u6709\u6570\u636e\u7686\u4ee5\u4e8c\u8fdb\u5236\u5f62\u5f0f\u5b58\u50a8\u3002\u4e8c\u8fdb\u5236\u7cfb\u7edf\u4ec5\u5305\u542b 0 \u548c 1 \u4e24\u4e2a\u6570\u5b57\uff0c\u800c\u6211\u4eec\u65e5\u5e38\u751f\u6d3b\u4e2d\u6240\u4f7f\u7528\u7684\u5341\u8fdb\u5236\u7cfb\u7edf\uff0c\u5219\u62e5\u6709 0 \u5230 9 \u5171\u5341\u4e2a\u6570\u5b57\u3002\u8fd9\u4e24\u79cd\u8fdb\u5236\u4e4b\u95f4\u7684\u8f6c\u6362\uff0c\u72b9\u5982\u4e0d\u540c\u8bed\u8a00\u95f4\u7684\u7ffb\u8bd1\uff0c\u662f\u7406\u89e3\u8ba1\u7b97\u673a\u6570\u636e\u8868\u793a\u7684\u5173\u952e\u3002<\/p>\n<p>\u4ee5\u5341\u8fdb\u5236\u6570\u5b57 13 \u4e3a\u4f8b\uff0c\u8f6c\u6362\u4e3a\u4e8c\u8fdb\u5236\u5f62\u5f0f\u4e3a 1101\u3002\u8f6c\u6362\u8fc7\u7a0b\u7c7b\u4f3c\u4e8e\u5c06\u5341\u8fdb\u5236\u7684\u201c\u6574\u4f53\u201d\u5206\u89e3\u4e3a\u4e8c\u8fdb\u5236\u7684\u201c\u7ec4\u6210\u90e8\u5206\u201d\u3002\u5177\u4f53\u6b65\u9aa4\u5982\u4e0b\uff1a<\/p>\n<p>13 \u9664\u4ee5 2\uff0c\u5546\u4e3a 6\uff0c\u4f59\u6570\u4e3a 1 \uff08\u4e8c\u8fdb\u5236\u6700\u4f4e\u4f4d\uff09<br \/>\n6 \u9664\u4ee5 2\uff0c\u5546\u4e3a 3\uff0c\u4f59\u6570\u4e3a 0<br \/>\n3 \u9664\u4ee5 2\uff0c\u5546\u4e3a 1\uff0c\u4f59\u6570\u4e3a 1<br \/>\n1 \u9664\u4ee5 2\uff0c\u5546\u4e3a 0\uff0c\u4f59\u6570\u4e3a 1 \uff08\u4e8c\u8fdb\u5236\u6700\u9ad8\u4f4d\uff09<\/p>\n<p>\u4f59\u6570\u4ece\u4e0b\u5f80\u4e0a\u6392\u5217\uff1a<br \/>\n\u21911<br \/>\n\u21910<br \/>\n\u21911<br \/>\n\u21911<br \/>\n\u5f97\u5230\u4e8c\u8fdb\u5236\u7ed3\u679c\uff1a1101<\/p>\n<p>\u53cd\u4e4b\uff0c\u5c06\u4e8c\u8fdb\u5236 1101 \u8f6c\u6362\u56de\u5341\u8fdb\u5236\uff0c\u5219\u5982\u540c\u5c06\u201c\u7ec4\u6210\u90e8\u5206\u201d\u91cd\u65b0\u7ec4\u5408\u8fd8\u539f\u4e3a\u201c\u6574\u4f53\u201d\u3002\u4ece\u53f3\u5411\u5de6\uff0c\u6bcf\u4e00\u4f4d\u7684\u6743\u91cd\u4ee5 2 \u7684\u5e42\u6b21\u65b9\u9012\u589e\uff0c\u6700\u53f3\u4f4d\u6743\u91cd\u4e3a \uff0c\u5411\u5de6\u4f9d\u6b21\u4e3a \u3001 \u3001 \u7b49\u3002\u56e0\u6b64\uff0c\u4e8c\u8fdb\u5236 1101 \u8f6c\u6362\u4e3a\u5341\u8fdb\u5236\u4e3a\uff1a 1\u00d7 + 1\u00d7 + 0\u00d7 + 1\u00d7 = 8 + 4 + 0 + 1 = 13\u3002<\/p>\n<h3>1.2 \u6d6e\u70b9\u6570\u4e0e\u6574\u6570\u533a\u522b<\/h3>\n<h4>\uff08\u4e00\uff09\u6574\u6570\u7c7b\u578b\uff08INT\uff09<\/h4>\n<p>INT \u662f Integer \u7684\u7f29\u5199\uff0c\u4ee3\u8868\u6574\u6570\u7c7b\u578b\u3002\u6574\u6570\uff0c\u987e\u540d\u601d\u4e49\uff0c\u662f\u4e0d\u5305\u542b\u5c0f\u6570\u90e8\u5206\u7684\u6570\u5b57\uff0c\u4f8b\u5982 1\u30012\u30013 \u7b49\u3002<\/p>\n<p>INT4 \u8868\u793a\u4f7f\u7528 4 \u4f4d\u4e8c\u8fdb\u5236\u6570\u6765\u8868\u793a\u4e00\u4e2a\u6574\u6570\uff0cINT8 \u5219\u4f7f\u7528 8 \u4f4d\u4e8c\u8fdb\u5236\u6570\u8868\u793a\u6574\u6570\u3002\u4f4d\u6570\u51b3\u5b9a\u4e86\u6574\u6570\u7684\u8868\u793a\u8303\u56f4\u3002<\/p>\n<p>INT4 \u80fd\u591f\u8868\u793a\u7684\u6574\u6570\u8303\u56f4\u662f\u6709\u9650\u7684\uff0c\u56e0\u4e3a 4 \u4f4d\u4e8c\u8fdb\u5236\u6570\u6700\u591a\u80fd\u8868\u793a \u4e2a\u4e0d\u540c\u7684\u6570\u3002\u5bf9\u4e8e\u6709\u7b26\u53f7 INT4\uff0c\u5176\u8868\u793a\u8303\u56f4\u901a\u5e38\u4e3a -8 \u5230 7\uff1b\u5bf9\u4e8e\u65e0\u7b26\u53f7 INT4\uff0c\u8303\u56f4\u5219\u4e3a 0 \u5230 15\u3002INT8 \u4ea6\u7136\uff0c\u6709\u7b26\u53f7 INT8 \u7684\u8868\u793a\u8303\u56f4\u4e3a -128 \u5230 127\uff0c\u65e0\u7b26\u53f7 INT8 \u7684\u8303\u56f4\u4e3a 0 \u5230 255\uff0c\u56e0\u4e3a 8 \u4f4d\u4e8c\u8fdb\u5236\u6570\u80fd\u8868\u793a = 256 \u4e2a\u4e0d\u540c\u7684\u6570\u3002<\/p>\n<h4>\uff08\u4e8c\uff09\u6d6e\u70b9\u6570\u7c7b\u578b\uff08FP\uff09<\/h4>\n<p>FP \u662f Floating Point \u7684\u7f29\u5199\uff0c\u5373\u6d6e\u70b9\u6570\u7c7b\u578b\u3002\u6d6e\u70b9\u6570\uff0c\u4e0e\u6574\u6570\u76f8\u5bf9\uff0c\u7528\u4e8e\u8868\u793a\u5e26\u6709\u5c0f\u6570\u90e8\u5206\u7684\u6570\u5b57\u3002<\/p>\n<p>\u6d6e\u70b9\u6570\u7531\u7b26\u53f7\u4f4d\u3001\u6307\u6570\u4f4d\u548c\u5c3e\u6570\u4f4d\u4e09\u90e8\u5206\u6784\u6210\u3002\u4ee5 32 \u4f4d\u6d6e\u70b9\u6570\uff08FP32\uff09\u4e3a\u4f8b\uff0c\u5b83\u5305\u542b 1 \u4f4d\u7b26\u53f7\u4f4d\u30018 \u4f4d\u6307\u6570\u4f4d\u548c 23 \u4f4d\u5c3e\u6570\u4f4d\u3002\u8fd9\u79cd\u7cbe\u5de7\u7684\u8bbe\u8ba1\u4f7f\u5f97\u6d6e\u70b9\u6570\u80fd\u591f\u8868\u793a\u6781\u5176\u5e7f\u6cdb\u7684\u6570\u503c\u8303\u56f4\uff0c\u4ece\u6781\u5c0f\u7684\u6570\u5230\u6781\u5927\u7684\u6570\uff0c\u72b9\u5982\u4e00\u628a\u4f38\u7f29\u81ea\u5982\u7684\u5377\u5c3a\u3002<\/p>\n<p>\u4f8b\u5982\uff0cFP32 \u53ef\u4ee5\u8868\u793a\u6781\u5c0f\u7684\u6570\uff08\u7ea6 \uff09\uff0c\u4e5f\u53ef\u4ee5\u8868\u793a\u6781\u5927\u7684\u6570\uff08\u7ea6 \uff09\u3002\u800c INT8 (8 \u4f4d\u6574\u6570) \u4ec5\u80fd\u8868\u793a -128 \u5230 127 \u4e4b\u95f4\u7684\u6574\u6570\u3002\u8fd9\u79cd\u5dee\u5f02\uff0c\u5982\u540c\u4f7f\u7528\u56fa\u5b9a\u957f\u5ea6\u7684\u76f4\u5c3a\uff08\u6574\u6570\uff09\u4e0e\u53ef\u4f38\u7f29\u7684\u5377\u5c3a\uff08\u6d6e\u70b9\u6570\uff09\u6d4b\u91cf\u957f\u5ea6\uff0c\u6d6e\u70b9\u6570\u5728\u6570\u503c\u8868\u793a\u7684\u7075\u6d3b\u6027\u548c\u8303\u56f4\u4e0a\u8fdc\u80dc\u4e8e\u6574\u6570\u3002<\/p>\n<h4>\uff08\u4e09\uff09\u5e38\u7528\u7684\u6570\u636e\u7c7b\u578b<\/h4>\n<p><img decoding=\"async\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/02\/40f81dad82708dc.png\" alt=\"\" \/><\/p>\n<p>\u5728\u6df1\u5ea6\u5b66\u4e60\u548c\u901a\u7528\u8ba1\u7b97\u9886\u57df\uff0c\u5e38\u89c1\u7684\u6570\u636e\u7c7b\u578b\u5305\u62ec\uff1a<\/p>\n<ul>\n<li><strong>Float32 (FP32)<\/strong>: \u8fd9\u662f\u6807\u51c6\u7684 32 \u4f4d\u6d6e\u70b9\u6570\u683c\u5f0f\uff0c\u4ee5\u9ad8\u7cbe\u5ea6\u548c\u5e7f\u9614\u7684\u6570\u503c\u8303\u56f4\u8457\u79f0\u3002FP32 \u8fd0\u7b97\u56e0\u5176\u901a\u7528\u6027\uff0c\u5f97\u5230\u5e7f\u6cdb\u786c\u4ef6\u7684\u5e7f\u6cdb\u652f\u6301\uff0c\u6545\u800c\u5728\u6a21\u578b\u8bad\u7ec3\u548c\u63a8\u7406\u4e2d\u5360\u636e\u4e3b\u5bfc\u5730\u4f4d\u3002<\/li>\n<li><strong>Float16 (FP16)<\/strong>: FP16 \u662f\u4e00\u79cd 16 \u4f4d\u534a\u7cbe\u5ea6\u6d6e\u70b9\u6570\uff0c\u76f8\u8f83\u4e8e FP32\uff0c\u5176\u7cbe\u5ea6\u6709\u6240\u964d\u4f4e\uff0c\u4f46\u5185\u5b58\u5360\u7528\u663e\u8457\u51cf\u5c11\uff0c\u8ba1\u7b97\u901f\u5ea6\u5927\u5e45\u63d0\u5347\u3002FP16 \u7684\u6570\u503c\u8868\u793a\u8303\u56f4\u76f8\u5bf9\u8f83\u7a84\uff0c\u53ef\u80fd\u9762\u4e34\u4e0a\u6ea2\u548c\u4e0b\u6ea2\u7684\u98ce\u9669\u3002\u7136\u800c\uff0c\u5728\u6df1\u5ea6\u5b66\u4e60\u5b9e\u8df5\u4e2d\uff0c\u8bf8\u5982<strong>\u635f\u5931\u7f29\u653e (Loss Scaling)<\/strong>\u00a0\u7b49\u6280\u5de7\u7684\u5e94\u7528\uff0c\u80fd\u591f\u6709\u6548\u7f13\u89e3\u8fd9\u4e9b\u95ee\u9898\u3002<\/li>\n<li><strong>BFloat16 (BF16)<\/strong>: BF16 \u662f\u53e6\u4e00\u79cd 16 \u4f4d\u6d6e\u70b9\u6570\u683c\u5f0f\u3002\u5176\u72ec\u7279\u4e4b\u5904\u5728\u4e8e\uff0cBF16 \u4fdd\u7559\u4e86\u4e0e FP32 \u76f8\u540c\u7684 8 \u4f4d\u6307\u6570\u4f4d\uff0c\u4ece\u800c\u62e5\u6709\u4e0e FP32 \u76f8\u5f53\u7684\u52a8\u6001\u8303\u56f4\u3002\u7136\u800c\uff0cBF16 \u7684\u5c3e\u6570\u4f4d\u4ec5\u4e3a 7 \u4f4d\uff0c\u7cbe\u5ea6\u4f4e\u4e8e FP16\u3002\u5728\u5904\u7406\u6570\u503c\u8303\u56f4\u8f83\u5927\u7684\u573a\u666f\u65f6\uff0cBF16 \u8868\u73b0\u51fa\u8272\uff0c\u4f46\u5728\u7cbe\u5ea6\u654f\u611f\u578b\u4efb\u52a1\u4e2d\uff0c\u53ef\u80fd\u9700\u8981\u6743\u8861\u8003\u8651\u3002<\/li>\n<li><strong>Int8<\/strong>: INT8 \u662f\u4e00\u79cd 8 \u4f4d\u6574\u6570\u7c7b\u578b\uff0c\u5176\u6570\u503c\u8868\u793a\u8303\u56f4\u6709\u9650\uff0c\u4f46\u5185\u5b58\u5360\u7528\u6781\u4f4e\u3002INT8 \u4e3b\u8981\u5e94\u7528\u4e8e\u6a21\u578b\u91cf\u5316\u6280\u672f\uff0c\u901a\u8fc7\u5c06\u6a21\u578b\u53c2\u6570\u7531\u9ad8\u7cbe\u5ea6\u7684 FP32 \u6216 FP16 \u8f6c\u6362\u4e3a INT8\uff0c\u53ef\u4ee5\u5927\u5e45\u964d\u4f4e\u6a21\u578b\u5b58\u50a8\u7a7a\u95f4\u9700\u6c42\u548c\u8ba1\u7b97\u590d\u6742\u5ea6\uff0c\u4e3a\u6a21\u578b\u5728\u8d44\u6e90\u53d7\u9650\u8bbe\u5907\u4e0a\u7684\u9ad8\u6548\u90e8\u7f72\u94fa\u5e73\u9053\u8def\u3002<\/li>\n<\/ul>\n<p>&nbsp;<\/p>\n<h2>2. \u91cf\u5316\u6982\u5ff5<\/h2>\n<h3>2.1 \u91cf\u5316\u5b9a\u4e49<\/h3>\n<p>\u91cf\u5316\u6307\u7684\u662f\u5c06\u6a21\u578b\u4e2d\u7684\u6570\u636e\u4ece\u9ad8\u7cbe\u5ea6\u8868\u793a\u8f6c\u6362\u4e3a\u4f4e\u7cbe\u5ea6\u8868\u793a\uff0c\u5c31\u50cf<strong>\u5728\u753b\u8d28\u548c\u6587\u4ef6\u5927\u5c0f\u4e4b\u95f4\u505a\u6743\u8861<\/strong>\uff0c\u5c06\u9ad8\u7cbe\u5ea6\u7684\u753b\u4f5c\u538b\u7f29\u6210\u4f4e\u7cbe\u5ea6\u7684 JPEG \u56fe\u50cf\uff0c\u5728\u6587\u4ef6\u5927\u5c0f\u5927\u5e45\u964d\u4f4e\u7684\u540c\u65f6\uff0c\u753b\u9762\u4e3b\u4f53\u4fe1\u606f\u4ecd\u7136\u4fdd\u7559\u3002\u5728\u6df1\u5ea6\u5b66\u4e60\u9886\u57df\uff0c\u91cf\u5316\u901a\u5e38\u6307\u5c06\u6a21\u578b\u6743\u91cd\u548c\u6fc0\u6d3b\u503c\u4ece FP32 (32 \u4f4d\u6d6e\u70b9\u6570) \u964d\u81f3 FP16 (16 \u4f4d\u6d6e\u70b9\u6570) \u4e43\u81f3 INT8 (8 \u4f4d\u6574\u6570) \u6216\u66f4\u4f4e\u7cbe\u5ea6\u3002<\/p>\n<p>FP32 \u4f5c\u4e3a\u4e00\u79cd\u9ad8\u7cbe\u5ea6\u6d6e\u70b9\u6570\u683c\u5f0f\uff0c\u4ee5 32 \u4f4d\u4e8c\u8fdb\u5236\u4f4d (1 \u4f4d\u7b26\u53f7\u4f4d\u30018 \u4f4d\u6307\u6570\u4f4d\u300123 \u4f4d\u5c3e\u6570\u4f4d) \u7cbe\u786e\u8868\u793a\u6570\u503c\u3002FP32 \u5b9b\u5982\u4e00\u628a\u7cbe\u5bc6\u7684\u6807\u5c3a\uff0c\u80fd\u591f\u7cbe\u786e\u8861\u91cf\u5fae\u5c0f\u81f3\u5b8f\u5927\u7684\u6570\u503c\uff0c\u4f46\u5176\u9ad8\u7cbe\u5ea6\u7279\u6027\u4ea6\u5e26\u6765\u5b58\u50a8\u5f00\u9500\u5927\u3001\u8ba1\u7b97\u901f\u5ea6\u76f8\u5bf9\u8f83\u6162\u7684\u7f3a\u70b9\u3002<\/p>\n<p>FP16 \u4f5c\u4e3a\u4e00\u79cd\u534a\u7cbe\u5ea6\u6d6e\u70b9\u6570\u683c\u5f0f\uff0c\u4ec5\u9700 16 \u4f4d\u4e8c\u8fdb\u5236\u4f4d (1 \u4f4d\u7b26\u53f7\u4f4d\u30015 \u4f4d\u6307\u6570\u4f4d\u300110 \u4f4d\u5c3e\u6570\u4f4d) \u5373\u53ef\u8868\u793a\u6570\u503c\u3002\u76f8\u8f83\u4e8e FP32\uff0cFP16 \u7684\u7cbe\u5ea6\u6709\u6240\u727a\u7272\uff0c\u4f46\u6362\u6765\u4e86\u5b58\u50a8\u7a7a\u95f4\u51cf\u534a\u548c\u8ba1\u7b97\u6548\u7387\u63d0\u5347\u3002FP16 \u72b9\u5982\u4e00\u628a\u523b\u5ea6\u7a0d\u7c97\u7684\u6807\u5c3a\uff0c\u5728\u7cbe\u5ea6\u4e0e\u6548\u7387\u4e4b\u95f4\u5b9e\u73b0\u4e86\u826f\u597d\u7684\u5e73\u8861\u3002<\/p>\n<p>INT8 \u5219\u662f\u4e00\u79cd 8 \u4f4d\u6574\u6570\u7c7b\u578b\uff0c\u4ec5\u80fd\u8868\u793a -128 \u81f3 127 \u8303\u56f4\u5185\u7684\u6574\u6570\u3002INT8 \u4ee5\u6781\u4f4e\u7684\u5b58\u50a8\u7a7a\u95f4\u5360\u7528\u548c\u6781\u9ad8\u7684\u8ba1\u7b97\u901f\u5ea6\u4e3a\u4f18\u52bf\uff0c\u4f46\u5176\u6570\u503c\u8868\u793a\u7cbe\u5ea6\u4ea6\u662f\u6700\u4f4e\u7684\u3002INT8 \u7c7b\u4f3c\u4e8e\u4e00\u4e2a\u7b80\u5355\u7684\u8ba1\u6570\u5668\uff0c\u4ec5\u80fd\u8fdb\u884c\u6574\u6570\u8ba1\u6570\uff0c\u4f46\u80dc\u5728\u5feb\u901f\u4fbf\u6377\u3002<\/p>\n<h3>2.2 \u91cf\u5316\u76ee\u7684<\/h3>\n<p>\u91cf\u5316\u7684\u6838\u5fc3\u76ee\u7684\u5728\u4e8e<strong>\u964d\u4f4e\u6a21\u578b\u7684\u5b58\u50a8\u9700\u6c42\u548c\u8ba1\u7b97\u590d\u6742\u5ea6<\/strong>\uff0c\u540c\u65f6\u529b\u6c42\u5c06\u7cbe\u5ea6\u635f\u5931\u63a7\u5236\u5728\u53ef\u63a5\u53d7\u8303\u56f4\u5185\u3002\u5177\u4f53\u800c\u8a00\uff0c\u91cf\u5316\u65e8\u5728\u5b9e\u73b0\u4ee5\u4e0b\u76ee\u6807\uff1a<\/p>\n<p><strong>\u964d\u4f4e\u5b58\u50a8\u9700\u6c42<\/strong>: \u73b0\u4ee3\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\uff0c\u7279\u522b\u662f\u53c2\u6570\u89c4\u6a21\u5e9e\u5927\u7684\u5927\u578b\u6a21\u578b\uff0c\u52a8\u8f84\u62e5\u6709\u6570\u4ebf\u4e43\u81f3\u6570\u5343\u4ebf\u53c2\u6570\uff0c\u5bf9\u5b58\u50a8\u7a7a\u95f4\u6784\u6210\u5de8\u5927\u538b\u529b\u3002\u4ee5 FP32 \u6a21\u578b\u4e3a\u4f8b\uff0c\u6bcf\u4e2a\u53c2\u6570\u5360\u7528 4 \u5b57\u8282\u5b58\u50a8\u7a7a\u95f4\u3002\u82e5\u5c06\u6a21\u578b\u91cf\u5316\u4e3a FP16\uff0c\u5219\u6bcf\u4e2a\u53c2\u6570\u4ec5\u9700 2 \u5b57\u8282\uff0c\u5b58\u50a8\u9700\u6c42\u7acb\u51cf\u4e00\u534a\u3002\u8fdb\u4e00\u6b65\u91cf\u5316\u81f3 INT8\uff0c\u6bcf\u4e2a\u53c2\u6570\u4ec5\u9700 1 \u5b57\u8282\uff0c\u5b58\u50a8\u7a7a\u95f4\u53ef\u7f29\u51cf 75%\u3002<\/p>\n<p><strong>\u63d0\u5347\u8ba1\u7b97\u6548\u7387<\/strong>: \u91cf\u5316\u540e\u7684\u6a21\u578b\u5728\u63a8\u7406\u9636\u6bb5\uff0c\u8ba1\u7b97\u91cf\u663e\u8457\u964d\u4f4e\uff0c\u4ece\u800c\u5b9e\u73b0\u63a8\u7406\u901f\u5ea6\u7684\u63d0\u5347\u3002\u4f8b\u5982\uff0cFP32 \u6a21\u578b\u5728 GPU \u4e0a\u8fd0\u884c\u65f6\uff0c\u8ba1\u7b97\u901f\u5ea6\u53ef\u80fd\u53d7\u9650\u4e8e\u5185\u5b58\u5e26\u5bbd\u3002\u800c FP16 \u6216 INT8 \u6a21\u578b\uff0c\u5f97\u76ca\u4e8e\u786c\u4ef6\u5bf9\u4f4e\u7cbe\u5ea6\u8ba1\u7b97\u7684\u4f18\u5316\u52a0\u901f\uff0c\u8ba1\u7b97\u901f\u5ea6\u53ef\u5927\u5e45\u63d0\u5347\u3002\u5c24\u5176\u5728\u8fb9\u7f18\u8bbe\u5907\u6216\u79fb\u52a8\u8bbe\u5907\u7b49\u8ba1\u7b97\u8d44\u6e90\u53d7\u9650\u7684\u573a\u666f\u4e2d\uff0c\u91cf\u5316\u6240\u5e26\u6765\u7684\u6027\u80fd\u63d0\u5347\u5c24\u4e3a\u663e\u8457\u3002<\/p>\n<p><strong>\u964d\u4f4e\u529f\u8017<\/strong>: \u6a21\u578b\u8ba1\u7b97\u8d44\u6e90\u9700\u6c42\u7684\u964d\u4f4e\uff0c\u76f4\u63a5\u8f6c\u5316\u4e3a\u80fd\u8017\u7684\u51cf\u5c11\u3002\u5bf9\u4e8e\u79fb\u52a8\u8bbe\u5907\u548c\u5d4c\u5165\u5f0f\u7cfb\u7edf\u800c\u8a00\uff0c\u529f\u8017\u662f\u81f3\u5173\u91cd\u8981\u7684\u8003\u91cf\u56e0\u7d20\u3002\u91cf\u5316\u6280\u672f\u80fd\u591f\u6709\u6548\u964d\u4f4e\u6a21\u578b\u529f\u8017\uff0c\u5ef6\u957f\u8bbe\u5907\u7eed\u822a\u65f6\u95f4\uff0c\u5e76\u964d\u4f4e\u6563\u70ed\u9700\u6c42\u3002<\/p>\n<p><strong>\u51cf\u5c11\u5e26\u5bbd\u9700\u6c42<\/strong>: \u5728\u5206\u5e03\u5f0f\u8ba1\u7b97\u7cfb\u7edf\u4e2d\uff0c\u6a21\u578b\u4f53\u79ef\u7684\u7f29\u51cf\u4ea6\u610f\u5473\u7740\u6570\u636e\u4f20\u8f93\u5e26\u5bbd\u9700\u6c42\u7684\u964d\u4f4e\u3002\u5728\u591a\u670d\u52a1\u5668\u534f\u540c\u5de5\u4f5c\u7684\u573a\u666f\u4e0b\uff0c\u91cf\u5316\u6a21\u578b\u80fd\u591f\u4ee5\u66f4\u5feb\u7684\u901f\u5ea6\u5b8c\u6210\u6a21\u578b\u5206\u53d1\u548c\u53c2\u6570\u540c\u6b65\uff0c\u4ece\u800c\u63d0\u5347\u6574\u4f53\u6570\u636e\u4f20\u8f93\u6548\u7387\u3002<\/p>\n<p>&nbsp;<\/p>\n<h2>3. INT4\u3001INT8 \u91cf\u5316<\/h2>\n<h3>3.1 INT4 \u4e0e INT8 \u8868\u793a\u8303\u56f4<\/h3>\n<p>INT4 \u548c INT8 \u5747\u4e3a\u6574\u6570\u7c7b\u578b\u7684\u91cf\u5316\u65b9\u6cd5\uff0c\u5b83\u4eec\u5728\u8ba1\u7b97\u673a\u7cfb\u7edf\u4e2d\u4ee5\u4e8c\u8fdb\u5236\u5f62\u5f0f\u5b58\u50a8\u6570\u636e\uff0c\u4f46\u5728\u6570\u503c\u8868\u793a\u8303\u56f4\u548c\u7cbe\u5ea6\u4e0a\u5b58\u5728\u5dee\u5f02\u3002<\/p>\n<ul>\n<li><strong>INT8<\/strong>: INT8 \u662f\u4e00\u79cd 8 \u4f4d\u6574\u6570\u7c7b\u578b\uff0c\u5176\u8868\u793a\u8303\u56f4\u4e3a -128 \u81f3 127\u3002\u53ef\u4ee5\u5c06\u5176\u7c7b\u6bd4\u4e3a\u4e00\u4e2a 8 \u4f4d\u8ba1\u6570\u5668\uff0c\u6bcf\u4e00\u4f4d\u53ef\u4ee5\u662f 0 \u6216 1\uff0c\u901a\u8fc7\u4e0d\u540c\u7684 0\/1 \u7ec4\u5408\u6765\u8868\u793a\u4e0d\u540c\u7684\u6574\u6570\u503c\u3002\u4f8b\u5982\uff0c\u5728<strong>\u6709\u7b26\u53f7<\/strong>\u00a0INT8 \u8868\u793a\u4e2d\uff0c\u8303\u56f4\u662f -128 \u5230 127\uff0c\u4e8c\u8fdb\u5236\u7684 11111111 \u8868\u793a\u5341\u8fdb\u5236\u7684 -1\u3002 \u5982\u679c\u662f<strong>\u65e0\u7b26\u53f7<\/strong>\u00a0INT8\uff0c\u8303\u56f4\u662f 0 \u5230 255\uff0c\u6b64\u65f6 11111111 \u5219\u8868\u793a\u5341\u8fdb\u5236\u7684 255\u3002INT8 \u7684\u8868\u793a\u8303\u56f4\u8db3\u4ee5\u6ee1\u8db3\u8bf8\u591a\u5e94\u7528\u573a\u666f\u7684\u9700\u6c42\uff0c\u4f8b\u5982\u56fe\u50cf\u5904\u7406\u4e2d\u7684\u50cf\u7d20\u503c\uff0c\u901a\u5e38\u5728 0 \u81f3 255 \u8303\u56f4\u5185\uff0cINT8 \u5373\u53ef\u6709\u6548\u8868\u793a\u3002<\/li>\n<li><strong>INT4<\/strong>: INT4 \u662f\u4e00\u79cd 4 \u4f4d\u6574\u6570\u7c7b\u578b\uff0c\u5176\u8868\u793a\u8303\u56f4\u8f83 INT8 \u66f4\u5c0f\uff0c\u4e3a -8 \u81f3 7\u3002INT4 \u4ee5\u727a\u7272\u6570\u503c\u8868\u793a\u8303\u56f4\u4e3a\u4ee3\u4ef7\uff0c\u6362\u53d6\u66f4\u5c0f\u7684\u5b58\u50a8\u7a7a\u95f4\u5360\u7528\u548c\u66f4\u5feb\u7684\u8ba1\u7b97\u901f\u5ea6\u3002INT4 \u72b9\u5982\u4e00\u4e2a\u66f4\u5c0f\u5de7\u7684\u8ba1\u6570\u5668\uff0c\u867d\u7136\u6570\u503c\u8868\u793a\u8303\u56f4\u53d7\u9650\uff0c\u4f46\u201c\u4f53\u79ef\u201d\u66f4\u5c0f\uff0c\u66f4\u8282\u7701\u8d44\u6e90\u3002\u5728\u4e00\u4e9b\u5bf9\u7cbe\u5ea6\u8981\u6c42\u76f8\u5bf9\u5bbd\u677e\u7684\u5e94\u7528\u573a\u666f\u4e2d\uff0c\u4f8b\u5982\u67d0\u4e9b\u8f7b\u91cf\u7ea7\u795e\u7ecf\u7f51\u7edc\u5c42\uff0c\u91c7\u7528 INT4 \u91cf\u5316\u80fd\u591f\u663e\u8457\u964d\u4f4e\u5b58\u50a8\u548c\u8ba1\u7b97\u6210\u672c\u3002<strong>\u4f8b\u5982\uff0c\u5728\u79fb\u52a8\u7aef\u90e8\u7f72\u8f7b\u91cf\u7ea7\u56fe\u50cf\u5206\u7c7b\u6a21\u578b\u65f6\uff0c\u6a21\u578b\u53c2\u6570\u548c\u4e2d\u95f4\u8ba1\u7b97\u7ed3\u679c\u53ef\u4ee5\u4f7f\u7528 INT4 \u683c\u5f0f\u5b58\u50a8\u548c\u8ba1\u7b97\uff0c\u4ee5\u51cf\u5c11\u5185\u5b58\u5360\u7528\u548c\u52a0\u901f\u63a8\u7406\u3002<\/strong><\/li>\n<\/ul>\n<h3>3.2 \u91cf\u5316\u516c\u5f0f\u4e0e\u793a\u4f8b<\/h3>\n<p>\u91cf\u5316\u7684\u8fc7\u7a0b\uff0c\u672c\u8d28\u4e0a\u662f\u5c06\u9ad8\u7cbe\u5ea6\u7684\u6d6e\u70b9\u6570\u6620\u5c04\u8f6c\u6362\u4e3a\u4f4e\u7cbe\u5ea6\u7684\u6574\u6570\u3002\u4ee5 FP32 \u91cf\u5316\u81f3 INT8 \u4e3a\u4f8b\uff0c\u5176\u91cf\u5316\u516c\u5f0f\u5982\u4e0b\uff1a<\/p>\n<p>\u662f\u539f\u59cb\u7684\u6d6e\u70b9\u6570\u3002<\/p>\n<p>\u662f\u91cf\u5316\u540e\u7684\u6574\u6570\u3002<\/p>\n<p>\u662f\u7f29\u653e\u56e0\u5b50\uff0c\u7528\u4e8e\u5c06\u6d6e\u70b9\u6570\u6620\u5c04\u5230\u6574\u6570\u8303\u56f4\u3002<\/p>\n<p>\u8868\u793a\u56db\u820d\u4e94\u5165\u5230\u6700\u8fd1\u7684\u6574\u6570\u3002<\/p>\n<p>\u8868\u793a\u5c06\u7ed3\u679c\u9650\u5236\u5728 INT8 \u7684\u8303\u56f4\u5185\uff0c\u5373 \uff3b-128, 127\uff3d\u3002<\/p>\n<h4>\u8ba1\u7b97\u7f29\u653e\u56e0\u5b50<\/h4>\n<p>\u7f29\u653e\u56e0\u5b50 (scale) \u7684\u8ba1\u7b97\uff0c\u901a\u5e38\u4f9d\u636e\u6d6e\u70b9\u6570\u7edd\u5bf9\u503c\u7684\u6700\u5927\u503c\u786e\u5b9a\u3002\u5047\u8bbe\u5b58\u5728\u4e00\u7ec4\u6d6e\u70b9\u6570 \uff0c\u8ba1\u7b97\u6b65\u9aa4\u5982\u4e0b\uff1a<\/p>\n<ol>\n<li>\u627e\u51fa\u8be5\u7ec4\u6d6e\u70b9\u6570\u7684\u6700\u5927\u7edd\u5bf9\u503c\uff1a<\/li>\n<li>\u8ba1\u7b97\u7f29\u653e\u56e0\u5b50\uff1a \u3002<\/li>\n<\/ol>\n<p><strong>\u5728\u5b9e\u9645\u5e94\u7528\u4e2d\uff0c\u7f29\u653e\u56e0\u5b50\u7684\u8ba1\u7b97\u65b9\u6cd5\u6709\u591a\u79cd\uff0c\u4f8b\u5982\u6700\u5927\u503c\u91cf\u5316 (Max Quantization)\u3001\u5747\u503c-\u6807\u51c6\u5dee\u91cf\u5316 (Mean-Std Quantization) \u7b49\u3002 \u6700\u5927\u503c\u91cf\u5316\u76f4\u63a5\u4f7f\u7528\u5f20\u91cf\u4e2d\u7684\u6700\u5927\u7edd\u5bf9\u503c\u6765\u8ba1\u7b97\u7f29\u653e\u56e0\u5b50\uff0c\u5b9e\u73b0\u7b80\u5355\uff0c\u4f46\u53ef\u80fd\u5bf9\u5f02\u5e38\u503c\u654f\u611f\u3002 \u5747\u503c-\u6807\u51c6\u5dee\u91cf\u5316\u5219\u5229\u7528\u6570\u636e\u7684\u5747\u503c\u548c\u6807\u51c6\u5dee\u4fe1\u606f\uff0c\u66f4\u7a33\u5065\u5730\u786e\u5b9a\u7f29\u653e\u56e0\u5b50\uff0c\u4f46\u8ba1\u7b97\u590d\u6742\u5ea6\u7a0d\u9ad8\u3002\u9009\u62e9\u5408\u9002\u7684\u7f29\u653e\u56e0\u5b50\u8ba1\u7b97\u65b9\u6cd5\uff0c\u9700\u8981\u5728\u7cbe\u5ea6\u548c\u8ba1\u7b97\u5f00\u9500\u4e4b\u95f4\u8fdb\u884c\u6743\u8861\u3002<\/strong><\/p>\n<h4>\u793a\u4f8b<\/h4>\n<p>\u5047\u8bbe\u6211\u4eec\u6709\u4e00\u7ec4\u6d6e\u70b9\u6570 [-0.5, 0.3, 1.2, -2.1]\uff0c\u4ee5\u4e0b\u5c06\u9010\u6b65\u6f14\u793a\u91cf\u5316\u8fc7\u7a0b\uff1a<\/p>\n<p><strong>1. \u8ba1\u7b97\u6700\u5927\u7edd\u5bf9\u503c\uff1a<\/strong><\/p>\n<p><strong>2. \u8ba1\u7b97\u7f29\u653e\u56e0\u5b50\uff1a<\/strong><\/p>\n<p><strong>3. \u91cf\u5316\u6bcf\u4e2a\u6d6e\u70b9\u6570\uff1a<\/strong><\/p>\n<ul>\n<li>\u5bf9\u4e8e -0.5:<\/li>\n<li>\u5bf9\u4e8e 0.3:<\/li>\n<li>\u5bf9\u4e8e 1.2:<\/li>\n<li>\u5bf9\u4e8e -2.1:<\/li>\n<\/ul>\n<p>\u6700\u7ec8\uff0c\u91cf\u5316\u540e\u7684 INT8 \u8868\u793a\u4e3a [-1, 1, 4, -7]\u3002<\/p>\n<p>\u901a\u8fc7\u4e0a\u8ff0\u6b65\u9aa4\uff0c\u6211\u4eec\u53ef\u4ee5\u6e05\u6670\u5730\u770b\u5230\u91cf\u5316\u6280\u672f\u5982\u4f55\u5c06\u6d6e\u70b9\u6570\u8f6c\u6362\u4e3a\u6574\u6570\u3002\u5c3d\u7ba1\u91cf\u5316\u8fc7\u7a0b\u4e0d\u53ef\u907f\u514d\u5730\u4f1a\u5f15\u5165\u4e00\u5b9a\u7684\u7cbe\u5ea6\u635f\u5931\uff0c\u4f46\u901a\u8fc7\u5408\u7406\u5730\u9009\u62e9\u7f29\u653e\u56e0\u5b50\uff0c\u53ef\u4ee5\u5728\u663e\u8457\u964d\u4f4e\u5b58\u50a8\u548c\u8ba1\u7b97\u6210\u672c\u7684\u540c\u65f6\uff0c\u6700\u5927\u9650\u5ea6\u5730\u4fdd\u6301\u6a21\u578b\u7684\u6027\u80fd\u6c34\u51c6\u3002\u00a0<strong>\u4e3a\u4e86\u8fdb\u4e00\u6b65\u51cf\u5c0f\u91cf\u5316\u8bef\u5dee\uff0c\u8fd8\u53ef\u4ee5\u5f15\u5165\u96f6\u70b9 (zero-point) \u91cf\u5316\u3002 \u96f6\u70b9\u91cf\u5316\u5728\u91cf\u5316\u516c\u5f0f\u4e2d\u52a0\u5165\u96f6\u70b9\u504f\u79fb\uff0c\u4f7f\u5f97\u6d6e\u70b9\u6570 0 \u7cbe\u786e\u6620\u5c04\u5230\u6574\u6570 0\uff0c \u4ece\u800c\u63d0\u9ad8\u91cf\u5316\u7cbe\u5ea6\uff0c\u5c24\u5176\u662f\u5728\u6fc0\u6d3b\u503c\u4e2d\u5b58\u5728\u5927\u91cf 0 \u503c\u65f6\u3002<\/strong><\/p>\n<p>&nbsp;<\/p>\n<h2>4. FP8\u3001FP16\u3001FP32 \u91cf\u5316<\/h2>\n<h3>4.1 FP8\u3001FP16\u3001FP32 \u8868\u793a\u65b9\u6cd5<\/h3>\n<p>FP8\u3001FP16 \u548c FP32 \u5747\u4e3a\u6d6e\u70b9\u6570\u8868\u793a\u683c\u5f0f\uff0c\u5b83\u4eec\u5728\u8ba1\u7b97\u673a\u4e2d\u4ee5\u4e8c\u8fdb\u5236\u5f62\u5f0f\u5b58\u50a8\uff0c\u4f46\u56e0\u4f4d\u5bbd\u4e0d\u540c\uff0c\u6545\u800c\u6570\u503c\u8868\u793a\u8303\u56f4\u548c\u7cbe\u5ea6\u5404\u5f02\u3002<\/p>\n<h4>FP32<\/h4>\n<p>FP32 \u4f5c\u4e3a\u4e00\u79cd\u6807\u51c6 32 \u4f4d\u6d6e\u70b9\u6570\u683c\u5f0f\uff0c\u7531\u4ee5\u4e0b\u4e09\u90e8\u5206\u6784\u6210\uff1a<\/p>\n<ul>\n<li><strong>\u7b26\u53f7\u4f4d (1 \u4f4d)<\/strong>: \u7528\u4e8e\u6807\u8bc6\u6570\u503c\u7684\u6b63\u8d1f\uff0c0 \u4ee3\u8868\u6b63\u6570\uff0c1 \u4ee3\u8868\u8d1f\u6570\u3002<\/li>\n<li><strong>\u6307\u6570\u4f4d (8 \u4f4d)<\/strong>: \u7528\u4e8e\u5b9a\u4e49\u6570\u503c\u7684\u5927\u5c0f\u8303\u56f4\uff0c\u4f7f\u5f97 FP32 \u80fd\u591f\u8868\u793a\u4ece\u6781\u5c0f\u5230\u6781\u5927\u7684\u6570\u503c\u3002<\/li>\n<li><strong>\u5c3e\u6570\u4f4d (23 \u4f4d)<\/strong>: \u7528\u4e8e\u786e\u5b9a\u6570\u503c\u7684\u7cbe\u5ea6\uff0c\u5c3e\u6570\u4f4d\u8d8a\u591a\uff0c\u7cbe\u5ea6\u8d8a\u9ad8\u3002<\/li>\n<\/ul>\n<p>FP32 \u62e5\u6709\u6781\u5bbd\u5e7f\u7684\u6570\u503c\u8868\u793a\u8303\u56f4\uff0c\u7ea6\u4e3a \u81f3 \uff0c\u5176\u7cbe\u5ea6\u53ef\u8fbe\u5c0f\u6570\u70b9\u540e 6 \u4f4d\u5de6\u53f3\u3002FP32 \u5b9b\u5982\u4e00\u628a\u9ad8\u7cbe\u5ea6\u7684\u5377\u5c3a\uff0c\u65e2\u80fd\u91cf\u5ea6\u7ec6\u5fae\u81f3\u5c18\u57c3\u7684\u5c3a\u5ea6\uff0c\u4ea6\u53ef\u4e08\u91cf\u5e7f\u9614\u5982\u5c71\u5ddd\u7684\u8ddd\u79bb\u3002\u00a0<strong>\u5728\u79d1\u5b66\u8ba1\u7b97\u3001\u91d1\u878d\u5efa\u6a21\u7b49\u5bf9\u7cbe\u5ea6\u8981\u6c42\u6781\u9ad8\u7684\u9886\u57df\uff0cFP32 \u662f\u4e0d\u53ef\u6216\u7f3a\u7684\u6570\u636e\u7c7b\u578b\u3002<\/strong><\/p>\n<h4>FP16<\/h4>\n<p><img decoding=\"async\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/02\/efbb5b72e7f7d59.png\" alt=\"\" \/><\/p>\n<p>FP16 \u662f\u4e00\u79cd\u534a\u7cbe\u5ea6\u6d6e\u70b9\u6570\u683c\u5f0f\uff0c\u4ec5\u5360\u7528 16 \u4f4d\u5b58\u50a8\u7a7a\u95f4\uff0c\u4e3a FP32 \u7684\u4e00\u534a\u3002FP16 \u7684\u7ed3\u6784\u5982\u4e0b\uff1a<\/p>\n<ul>\n<li><strong>\u7b26\u53f7\u4f4d (1 \u4f4d)<\/strong>: \u6807\u8bc6\u6570\u503c\u6b63\u8d1f\u3002<\/li>\n<li><strong>\u6307\u6570\u4f4d (5 \u4f4d)<\/strong>: \u5b9a\u4e49\u6570\u503c\u5927\u5c0f\u8303\u56f4\u3002<\/li>\n<li><strong>\u5c3e\u6570\u4f4d (10 \u4f4d)<\/strong>: \u51b3\u5b9a\u6570\u503c\u7cbe\u5ea6\u3002<\/li>\n<\/ul>\n<p>FP16 \u7684\u6570\u503c\u8868\u793a\u8303\u56f4\u7ea6\u4e3a \u5230 \uff0c\u7cbe\u5ea6\u76f8\u8f83 FP32 \u964d\u4f4e\uff0c\u7ea6\u4e3a\u5c0f\u6570\u70b9\u540e 3 \u4f4d\u3002FP16 \u72b9\u5982\u4e00\u628a\u523b\u5ea6\u7565\u7c97\u7684\u76f4\u5c3a\uff0c\u7cbe\u5ea6\u6709\u6240\u964d\u4f4e\uff0c\u4f46\u5728\u5b58\u50a8\u7a7a\u95f4\u548c\u8ba1\u7b97\u6548\u7387\u4e0a\u66f4\u5177\u4f18\u52bf\u3002\u00a0<strong>FP16 \u5e38\u7528\u4e8e\u5bf9\u8ba1\u7b97\u901f\u5ea6\u548c\u5185\u5b58\u5e26\u5bbd\u6709\u8f83\u9ad8\u8981\u6c42\u7684\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u8bad\u7ec3\u548c\u63a8\u7406\u4e2d\u3002 \u5c24\u5176\u662f\u5728 GPU \u52a0\u901f\u7684\u573a\u666f\u4e0b\uff0c\u4f7f\u7528 FP16 \u53ef\u4ee5\u5145\u5206\u5229\u7528 Tensor Core \u7b49\u786c\u4ef6\u52a0\u901f\u5355\u5143\uff0c \u4ece\u800c\u5b9e\u73b0\u663e\u8457\u7684\u6027\u80fd\u63d0\u5347\u3002<\/strong><\/p>\n<h4>FP8<\/h4>\n<p>FP8 \u662f\u4e00\u79cd\u65b0\u5174\u7684\u4f4e\u7cbe\u5ea6\u6d6e\u70b9\u6570\u683c\u5f0f\uff0c\u4e3b\u8981\u5e94\u7528\u4e8e\u6df1\u5ea6\u5b66\u4e60\u9886\u57df\uff0c\u65e8\u5728\u5b9e\u73b0\u9ad8\u6548\u8ba1\u7b97\u3002FP8 \u7684\u5178\u578b\u7ed3\u6784\u5982\u4e0b\uff1a<\/p>\n<ul>\n<li><strong>\u7b26\u53f7\u4f4d (1 \u4f4d)<\/strong>: \u6807\u8bc6\u6570\u503c\u6b63\u8d1f\u3002<\/li>\n<li><strong>\u6307\u6570\u4f4d (3 \u4f4d\u6216 4 \u4f4d)<\/strong>: \u5b9a\u4e49\u6570\u503c\u5927\u5c0f\u8303\u56f4 (\u5b58\u5728\u4e24\u79cd FP8 \u53d8\u4f53)\u3002<\/li>\n<li><strong>\u5c3e\u6570\u4f4d (4 \u4f4d\u6216 3 \u4f4d)<\/strong>: \u51b3\u5b9a\u6570\u503c\u7cbe\u5ea6 (\u4e0e\u6307\u6570\u4f4d\u4f4d\u6570\u5bf9\u5e94)\u3002<\/li>\n<\/ul>\n<p>FP8 \u7684\u6570\u503c\u8868\u793a\u8303\u56f4\u548c\u7cbe\u5ea6\u5747\u8fdb\u4e00\u6b65\u964d\u4f4e\uff0c\u4f46\u5176\u4f18\u52bf\u5728\u4e8e\u66f4\u5c0f\u7684\u5b58\u50a8\u7a7a\u95f4\u5360\u7528\u548c\u66f4\u5feb\u7684\u8ba1\u7b97\u901f\u5ea6\u3002\u5982\u679c\u8bf4 FP32 \u662f\u4e00\u628a\u7cbe\u5bc6\u7684\u5377\u5c3a\uff0cFP16 \u662f\u523b\u5ea6\u7a0d\u7c97\u7684\u76f4\u5c3a\uff0c\u90a3\u4e48 FP8\u00a0<strong>\u5c31\u50cf\u662f\u4e00\u4e2a\u53ea\u6709\u5398\u7c73\u523b\u5ea6\u7684\u7b80\u6613\u76f4\u5c3a<\/strong>\uff0c\u91cf\u7a0b\u548c\u7cbe\u5ea6\u90fd\u8fdb\u4e00\u6b65\u964d\u4f4e\uff0c\u4f46\u5728\u7279\u5b9a\u573a\u5408\u4e0b\uff0c\u4ecd\u7136\u53ef\u4ee5\u5feb\u901f\u5b8c\u6210\u6d4b\u91cf\u4efb\u52a1\u3002\u00a0<strong>FP8 \u4f5c\u4e3a\u4e00\u79cd\u6781\u4f4e\u7cbe\u5ea6\u7684\u6570\u636e\u7c7b\u578b\uff0c \u5728\u5bf9\u5ef6\u8fdf\u548c\u541e\u5410\u91cf\u6709\u6781\u81f4\u8981\u6c42\u7684\u573a\u666f\u4e2d\u5c55\u73b0\u51fa\u5de8\u5927\u6f5c\u529b\uff0c\u4f8b\u5982\u5b9e\u65f6\u63a8\u7406\u3001\u8fb9\u7f18\u8ba1\u7b97\u7b49\u3002 \u4f46 FP8 \u7684\u5e94\u7528\u4e5f\u5bf9\u786c\u4ef6\u548c\u7b97\u6cd5\u63d0\u51fa\u4e86\u66f4\u9ad8\u7684\u8981\u6c42\uff0c \u9700\u8981\u4e13\u95e8\u7684\u786c\u4ef6\u652f\u6301\u548c\u91cf\u5316\u7b56\u7565\u6765\u4fdd\u8bc1\u7cbe\u5ea6\u3002<\/strong><\/p>\n<h3>4.2 \u91cf\u5316\u8fc7\u7a0b\u4e0e\u516c\u5f0f<\/h3>\n<p>\u91cf\u5316\u8fc7\u7a0b\uff0c\u662f\u5c06\u9ad8\u7cbe\u5ea6\u6d6e\u70b9\u6570\u8f6c\u6362\u4e3a\u4f4e\u7cbe\u5ea6\u6d6e\u70b9\u6570\u6216\u6574\u6570\u7684\u8fc7\u7a0b\u3002\u4ee5 FP32 \u91cf\u5316\u81f3 FP16 \u4e3a\u4f8b\uff0c\u5176\u91cf\u5316\u516c\u5f0f\u5982\u4e0b\uff1a<\/p>\n<p>\u662f\u539f\u59cb\u7684\u6d6e\u70b9\u6570\u3002<\/p>\n<p>\u662f\u91cf\u5316\u540e\u7684\u4f4e\u7cbe\u5ea6\u6d6e\u70b9\u6570\u3002<\/p>\n<p>\u662f\u7f29\u653e\u56e0\u5b50\uff0c\u7528\u4e8e\u5c06\u6d6e\u70b9\u6570\u6620\u5c04\u5230\u4f4e\u7cbe\u5ea6\u8303\u56f4\u3002<\/p>\n<p>\u8868\u793a\u56db\u820d\u4e94\u5165\u5230\u6700\u8fd1\u7684\u503c\u3002<\/p>\n<p><strong>\u8ba1\u7b97\u7f29\u653e\u56e0\u5b50<\/strong><\/p>\n<p>\u7f29\u653e\u56e0\u5b50 (scale) \u7684\u8ba1\u7b97\uff0c\u901a\u5e38\u4f9d\u636e\u6d6e\u70b9\u6570\u7edd\u5bf9\u503c\u7684\u6700\u5927\u503c\u786e\u5b9a\u3002\u5047\u8bbe\u5b58\u5728\u4e00\u7ec4\u6d6e\u70b9\u6570 \uff0c\u8ba1\u7b97\u6b65\u9aa4\u5982\u4e0b\uff1a<\/p>\n<ol>\n<li>\u627e\u51fa\u8be5\u7ec4\u6d6e\u70b9\u6570\u7684\u6700\u5927\u7edd\u5bf9\u503c\uff1a<\/li>\n<li>\u8ba1\u7b97\u7f29\u653e\u56e0\u5b50\uff1a \uff0c\u5176\u4e2d \u662f\u76ee\u6807\u4f4e\u7cbe\u5ea6\u683c\u5f0f\u7684\u6700\u5927<strong>\u53ef\u8868\u793a\u7edd\u5bf9\u503c<\/strong>\u3002\u5bf9\u4e8e FP16\uff0c\u8fd9\u4e2a\u503c\u7ea6\u4e3a\u00a0<strong>65504<\/strong>\u3002<\/li>\n<\/ol>\n<p><strong>\u7c7b\u4f3c\u4e8e INT8 \u91cf\u5316\uff0c FP16 \u91cf\u5316\u4e2d\u7f29\u653e\u56e0\u5b50\u7684\u8ba1\u7b97\u65b9\u6cd5\u4e5f\u5f88\u591a\u6837\uff0c \u53ef\u4ee5\u6839\u636e\u4e0d\u540c\u7684\u7cbe\u5ea6\u548c\u6027\u80fd\u9700\u6c42\u8fdb\u884c\u9009\u62e9\u3002 \u6b64\u5916\uff0c FP16 \u91cf\u5316\u4e5f\u5e38\u5e38\u7ed3\u5408\u6df7\u5408\u7cbe\u5ea6\u8bad\u7ec3 (Mixed Precision Training) \u6280\u672f\u4f7f\u7528\u3002 \u6df7\u5408\u7cbe\u5ea6\u8bad\u7ec3\u5728\u6a21\u578b\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0c \u67d0\u4e9b\u8ba1\u7b97\u5bc6\u96c6\u578b\u64cd\u4f5c (\u5982\u77e9\u9635\u4e58\u6cd5\u3001\u5377\u79ef) \u4f7f\u7528 FP16\uff0c \u800c\u5bf9\u7cbe\u5ea6\u8981\u6c42\u8f83\u9ad8\u7684\u64cd\u4f5c (\u5982\u635f\u5931\u8ba1\u7b97\u3001\u68af\u5ea6\u66f4\u65b0) \u4ecd\u4f7f\u7528 FP32\uff0c \u4ece\u800c\u5728\u4fdd\u8bc1\u6a21\u578b\u7cbe\u5ea6\u7684\u524d\u63d0\u4e0b\uff0c \u52a0\u901f\u8bad\u7ec3\u8fc7\u7a0b\u5e76\u51cf\u5c11\u5185\u5b58\u5360\u7528\u3002<\/strong><\/p>\n<p><strong>\u793a\u4f8b<\/strong><\/p>\n<p>\u5047\u8bbe\u6211\u4eec\u6709\u4e00\u7ec4 FP32 \u6d6e\u70b9\u6570 [-0.5, 0.3, 1.2, -2.1]\uff0c\u4ee5\u4e0b\u5c06\u9010\u6b65\u6f14\u793a FP32 \u91cf\u5316\u81f3 FP16 \u7684\u8fc7\u7a0b\uff1a<\/p>\n<p><strong>1. \u8ba1\u7b97\u6700\u5927\u7edd\u5bf9\u503c\uff1a<\/strong><\/p>\n<p><strong>2. \u8ba1\u7b97\u7f29\u653e\u56e0\u5b50\uff1a<\/strong><\/p>\n<p><strong>3. \u91cf\u5316\u6bcf\u4e2a\u6d6e\u70b9\u6570\uff1a<\/strong><\/p>\n<ul>\n<li>\u5bf9\u4e8e -0.5:<\/li>\n<li>\u5bf9\u4e8e 0.3:<\/li>\n<li>\u5bf9\u4e8e 1.2:<\/li>\n<li>\u5bf9\u4e8e -2.1:<\/li>\n<\/ul>\n<p>\u6700\u7ec8\uff0c\u91cf\u5316\u540e\u7684 FP16 \u8868\u793a\u4e3a [-0.5, 0.3, 1.2, -2.1]\u3002<\/p>\n<p>\u901a\u8fc7\u4ee5\u4e0a\u6b65\u9aa4\uff0c\u53ef\u4ee5\u89c2\u5bdf\u5230\u91cf\u5316\u6280\u672f\u5982\u4f55\u5c06\u9ad8\u7cbe\u5ea6\u6d6e\u70b9\u6570\u8f6c\u6362\u4e3a\u4f4e\u7cbe\u5ea6\u6d6e\u70b9\u6570\u3002\u91cf\u5316\u56fa\u7136\u4f1a\u5e26\u6765\u4e00\u5b9a\u7684\u7cbe\u5ea6\u635f\u5931\uff0c\u4f46\u901a\u8fc7\u5408\u7406\u9009\u62e9\u7f29\u653e\u56e0\u5b50\uff0c\u53ef\u4ee5\u5728\u964d\u4f4e\u5b58\u50a8\u548c\u8ba1\u7b97\u6210\u672c\u7684\u540c\u65f6\uff0c\u5c3d\u53ef\u80fd\u5730\u7ef4\u6301\u6a21\u578b\u7684\u6027\u80fd\u8868\u73b0\u3002\u00a0<strong>\u4e3a\u4e86\u8fdb\u4e00\u6b65\u63d0\u5347 FP16 \u91cf\u5316\u7684\u7cbe\u5ea6\uff0c \u53ef\u4ee5\u91c7\u7528\u52a8\u6001\u91cf\u5316 (Dynamic Quantization) \u6280\u672f\u3002 \u52a8\u6001\u91cf\u5316\u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d\uff0c \u6839\u636e\u8f93\u5165\u6570\u636e\u7684\u5b9e\u9645\u8303\u56f4\uff0c \u52a8\u6001\u8c03\u6574\u7f29\u653e\u56e0\u5b50\uff0c \u4ece\u800c\u66f4\u597d\u5730\u9002\u5e94\u6570\u636e\u7684\u5206\u5e03\u53d8\u5316\uff0c \u51cf\u5c0f\u91cf\u5316\u8bef\u5dee\u3002<\/strong><\/p>\n<p>&nbsp;<\/p>\n<h2>5. \u91cf\u5316\u5e94\u7528\u4e0e\u4f18\u52bf<\/h2>\n<h3>5.1 \u5728\u6df1\u5ea6\u5b66\u4e60\u4e2d\u7684\u5e94\u7528<\/h3>\n<p>\u91cf\u5316\u6280\u672f\u5728\u6df1\u5ea6\u5b66\u4e60\u9886\u57df\u62e5\u6709\u5e7f\u6cdb\u7684\u5e94\u7528\u524d\u666f\uff0c\u7279\u522b\u662f\u5728\u6a21\u578b\u8bad\u7ec3\u548c\u63a8\u7406\u9636\u6bb5\uff0c\u5176\u4ef7\u503c\u65e5\u76ca\u51f8\u663e\u3002\u4ee5\u4e0b\u662f\u91cf\u5316\u6280\u672f\u5728\u6df1\u5ea6\u5b66\u4e60\u4e2d\u7684\u51e0\u9879\u4e3b\u8981\u5e94\u7528\uff1a<\/p>\n<h4>\u6a21\u578b\u8bad\u7ec3\u52a0\u901f<\/h4>\n<p>\u5728\u6a21\u578b\u8bad\u7ec3\u9636\u6bb5\uff0c\u91c7\u7528\u4f4e\u7cbe\u5ea6\u6570\u636e\u7c7b\u578b (\u5982 FP16 \u6216 FP8) \u8fdb\u884c\u8ba1\u7b97\uff0c\u80fd\u591f\u663e\u8457\u52a0\u901f\u8bad\u7ec3\u8fdb\u7a0b\u3002\u4f8b\u5982\uff0cNVIDIA Hopper \u67b6\u6784 GPU \u652f\u6301 FP8 \u7cbe\u5ea6\u7684 Tensor Core \u8fd0\u7b97\u3002\u4e0e\u4f20\u7edf\u7684 FP32 \u8bad\u7ec3\u76f8\u6bd4\uff0cFP8 \u8bad\u7ec3\u901f\u5ea6\u53ef\u63d0\u5347 2 \u81f3 3 \u500d\u3002\u8fd9\u79cd\u8bad\u7ec3\u52a0\u901f\u5bf9\u4e8e\u53c2\u6570\u89c4\u6a21\u5e9e\u5927\u7684\u6a21\u578b\u5c24\u4e3a\u5173\u952e\uff0c\u80fd\u591f\u5927\u5e45\u7f29\u51cf\u8bad\u7ec3\u65f6\u95f4\u548c\u8ba1\u7b97\u8d44\u6e90\u6d88\u8017\u3002\u00a0<strong>\u4f8b\u5982\uff0c \u8bad\u7ec3 GPT-3 \u7b49\u5927\u578b\u8bed\u8a00\u6a21\u578b\u65f6\uff0c \u91c7\u7528 FP16 \u6216 BF16 \u6df7\u5408\u7cbe\u5ea6\u8bad\u7ec3\uff0c \u53ef\u4ee5\u663e\u8457\u7f29\u77ed\u8bad\u7ec3\u65f6\u95f4\uff0c \u8282\u7701\u5927\u91cf\u7684\u8ba1\u7b97\u8d44\u6e90\u3002<\/strong><\/p>\n<p>\u4ee5 <a href=\"https:\/\/www.kdjingpai.com\/piaizhinengxuexi\/\">Inflection AI<\/a> \u7684 Inflection-2 \u6a21\u578b\u4e3a\u4f8b\uff0c\u8be5\u6a21\u578b\u91c7\u7528\u4e86 FP8 \u6df7\u5408\u7cbe\u5ea6\u8bad\u7ec3\u7b56\u7565\uff0c\u5728 5000 \u5757 NVIDIA Hopper \u67b6\u6784 GPU \u4e0a\u5b8c\u6210\u4e86\u8bad\u7ec3\uff0c\u603b\u8ba1\u6d6e\u70b9\u8fd0\u7b97\u6b21\u6570\u9ad8\u8fbe FLOPs\u3002\u5728\u591a\u9879\u6807\u51c6\u4eba\u5de5\u667a\u80fd\u6027\u80fd\u57fa\u51c6\u6d4b\u8bd5\u4e2d\uff0cInflection-2 \u76f8\u8f83\u4e8e\u540c\u5c5e\u8bad\u7ec3\u8ba1\u7b97\u7c7b\u522b\u7684 Google \u65d7\u8230\u6a21\u578b PaLM 2\uff0c\u5c55\u73b0\u51fa\u5353\u8d8a\u7684\u6027\u80fd\u4f18\u52bf\u3002<\/p>\n<h4>\u6a21\u578b\u63a8\u7406\u4f18\u5316<\/h4>\n<p>\u5728\u6a21\u578b\u63a8\u7406\u9636\u6bb5\uff0c\u91cf\u5316\u6280\u672f\u80fd\u591f\u663e\u8457\u964d\u4f4e\u6a21\u578b\u7684\u5b58\u50a8\u9700\u6c42\u548c\u8ba1\u7b97\u590d\u6742\u5ea6\uff0c\u8fdb\u800c\u63d0\u5347\u63a8\u7406\u6548\u7387\u3002\u4f8b\u5982\uff0c\u5c06 FP32 \u6a21\u578b\u91cf\u5316\u4e3a INT8\uff0c\u6a21\u578b\u5b58\u50a8\u7a7a\u95f4\u53ef\u51cf\u5c11 75%\uff0c\u63a8\u7406\u901f\u5ea6\u53ef\u63d0\u5347\u6570\u500d\u3002\u8fd9\u5bf9\u4e8e\u5728\u8fb9\u7f18\u8bbe\u5907\u6216\u79fb\u52a8\u8bbe\u5907\u4e0a\u90e8\u7f72\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u81f3\u5173\u91cd\u8981\uff0c\u56e0\u4e3a\u8fd9\u4e9b\u8bbe\u5907\u901a\u5e38\u9762\u4e34\u8ba1\u7b97\u8d44\u6e90\u548c\u5b58\u50a8\u7a7a\u95f4\u53d7\u9650\u7684\u6311\u6218\u3002\u00a0<strong>\u4f8b\u5982\uff0c \u5728\u624b\u673a\u7aef\u90e8\u7f72\u56fe\u50cf\u8bc6\u522b\u6a21\u578b\u65f6\uff0c \u5c06\u6a21\u578b\u91cf\u5316\u4e3a INT8 \u53ef\u4ee5\u6709\u6548\u964d\u4f4e\u6a21\u578b\u5927\u5c0f\uff0c \u51cf\u5c11\u5185\u5b58\u5360\u7528\uff0c \u5e76\u52a0\u901f\u63a8\u7406\u901f\u5ea6\uff0c \u63d0\u5347\u7528\u6237\u4f53\u9a8c\u3002<\/strong><\/p>\n<p>\u4f8b\u5982\uff0cGoogle \u4e0e NVIDIA \u56e2\u961f\u7d27\u5bc6\u5408\u4f5c\uff0c\u5c06 TensorRT-LLM \u4f18\u5316\u6280\u672f\u5e94\u7528\u4e8e Gemma \u6a21\u578b\uff0c\u5e76\u7ed3\u5408 FP8 \u6280\u672f\u5b9e\u73b0\u4e86\u63a8\u7406\u52a0\u901f\u3002\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0c\u5728\u4f7f\u7528 Hopper GPU \u8fdb\u884c\u63a8\u7406\u65f6\uff0cFP8 \u8f83\u4e4b FP16 \u5728\u541e\u5410\u91cf\u65b9\u9762\u5b9e\u73b0\u4e86 3 \u500d\u4ee5\u4e0a\u7684\u63d0\u5347\u3002<\/p>\n<h4>\u6a21\u578b\u538b\u7f29\u4e0e\u90e8\u7f72<\/h4>\n<p>\u91cf\u5316\u6280\u672f\u4ea6\u53ef\u7528\u4e8e\u6a21\u578b\u538b\u7f29\u548c\u90e8\u7f72\u3002\u901a\u8fc7\u5c06\u9ad8\u7cbe\u5ea6\u6a21\u578b\u91cf\u5316\u4e3a\u4f4e\u7cbe\u5ea6\u6a21\u578b\uff0c\u53ef\u4ee5\u6709\u6548\u51cf\u5c0f\u6a21\u578b\u4f53\u79ef\uff0c\u4ece\u800c \u043e\u0431\u043b\u0435\u0433\u0447\u0438\u0442\u044c \u6a21\u578b\u5728\u8d44\u6e90\u53d7\u9650\u73af\u5883\u4e2d\u7684\u90e8\u7f72\u3002\u4f8b\u5982\uff0c\u96f6\u4e00\u4e07\u7269<strong>\u5229\u7528<\/strong>\u00a0NVIDIA \u8f6f\u786c\u4ef6\u7ed3\u5408\u7684\u6280\u672f\u6808\uff0c\u5b8c\u6210\u4e86\u5927\u6a21\u578b FP8 \u8bad\u7ec3\u548c\u9a8c\u8bc1\uff0c\u5176\u5927\u6a21\u578b\u8bad\u7ec3\u541e\u5410\u91cf\u76f8\u8f83 BF16 \u63d0\u5347\u4e86 1.3 \u500d\u3002\u00a0<strong>\u9664\u4e86 INT8 \u548c FP16\/FP8 \u91cf\u5316\u5916\uff0c \u8fd8\u5b58\u5728 INT4 \u751a\u81f3\u66f4\u4f4e\u6bd4\u7279\u7684\u91cf\u5316\u6280\u672f\uff0c \u4f8b\u5982\u4e8c\u503c\u5316\u795e\u7ecf\u7f51\u7edc (Binary Neural Network, BNN) \u548c\u4e09\u503c\u5316\u795e\u7ecf\u7f51\u7edc (Ternary Neural Network, TNN)\u3002 \u8fd9\u4e9b\u6781\u4f4e\u6bd4\u7279\u91cf\u5316\u6280\u672f\u53ef\u4ee5\u5c06\u6a21\u578b\u538b\u7f29\u5230\u6781\u81f4\uff0c \u4f46\u901a\u5e38\u4f1a\u5e26\u6765\u8f83\u5927\u7684\u7cbe\u5ea6\u635f\u5931\uff0c \u9002\u7528\u4e8e\u5bf9\u6a21\u578b\u5927\u5c0f\u548c\u901f\u5ea6\u6709\u6781\u7aef\u8981\u6c42\u7684\u573a\u666f\u3002<\/strong><\/p>\n<p>\u6b64\u5916\uff0c\u91cf\u5316\u540e\u7684\u6a21\u578b\u53ef\u4ee5\u501f\u52a9\u7279\u5b9a\u7684\u786c\u4ef6\u52a0\u901f\u6280\u672f\uff0c\u8fdb\u4e00\u6b65\u63d0\u5347\u6027\u80fd\u3002\u4f8b\u5982\uff0cNVIDIA <a href=\"https:\/\/www.kdjingpai.com\/transformer\/\">Transformer<\/a> Engine \u5df2\u88ab\u96c6\u6210\u81f3 PyTorch\u3001JAX\u3001PaddlePaddle \u7b49\u4e3b\u6d41\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\uff0c\u4e3a\u91cf\u5316\u6a21\u578b\u7684\u63a8\u7406\u63d0\u4f9b\u4e86\u9ad8\u6548\u7684\u786c\u4ef6\u7ea7\u652f\u6301\u3002\u00a0<strong>\u9664\u4e86 NVIDIA GPU \u5916\uff0c \u5176\u4ed6\u786c\u4ef6\u5e73\u53f0\uff0c \u4f8b\u5982 ARM \u67b6\u6784 CPU\u3001 \u79fb\u52a8\u7aef NPU \u7b49\uff0c \u4e5f\u9488\u5bf9\u91cf\u5316\u8ba1\u7b97\u8fdb\u884c\u4e86\u4e13\u95e8\u7684\u4f18\u5316\u548c\u52a0\u901f\uff0c \u4e3a\u91cf\u5316\u6a21\u578b\u7684\u5e7f\u6cdb\u90e8\u7f72\u63d0\u4f9b\u4e86\u786c\u4ef6\u57fa\u7840\u3002<\/strong><\/p>\n<h3>5.2 \u4f18\u52bf\u4e0e\u5c40\u9650\u6027<\/h3>\n<h4>\u4f18\u52bf<\/h4>\n<p><strong>\u8ba1\u7b97\u6548\u7387\u63d0\u5347<\/strong>: \u4f4e\u7cbe\u5ea6\u91cf\u5316\u80fd\u591f\u663e\u8457\u63d0\u5347\u8ba1\u7b97\u901f\u5ea6\uff0c\u964d\u4f4e\u8ba1\u7b97\u8d44\u6e90\u6d88\u8017\u3002FP16 \u548c FP8 \u7684\u8ba1\u7b97\u541e\u5410\u91cf\u5747\u6570\u500d\u9ad8\u4e8e FP32\u3002\u8fd9\u79cd\u52a0\u901f\u6548\u5e94\u5728\u5927\u89c4\u6a21\u6a21\u578b\u8bad\u7ec3\u548c\u63a8\u7406\u4e2d\u5c24\u4e3a\u7a81\u51fa\u3002<\/p>\n<p><strong>\u5b58\u50a8\u9700\u6c42\u964d\u4f4e<\/strong>: \u91cf\u5316\u6280\u672f\u53ef\u5927\u5e45\u964d\u4f4e\u6a21\u578b\u5b58\u50a8\u9700\u6c42\u3002\u4f8b\u5982\uff0c\u5c06 FP32 \u6a21\u578b\u91cf\u5316\u4e3a INT8\uff0c\u5b58\u50a8\u7a7a\u95f4\u53ef\u7f29\u51cf 75%\u3002\u8fd9\u5bf9\u4e8e\u5728\u5b58\u50a8\u8d44\u6e90\u53d7\u9650\u7684\u73af\u5883\u4e2d\u90e8\u7f72\u6a21\u578b\u5177\u6709\u91cd\u8981\u610f\u4e49\u3002<\/p>\n<p><strong>\u529f\u8017\u964d\u4f4e<\/strong>: \u4f4e\u7cbe\u5ea6\u8ba1\u7b97\u6240\u9700\u8ba1\u7b97\u8d44\u6e90\u66f4\u5c11\uff0c\u4ece\u800c\u964d\u4f4e\u4e86\u8bbe\u5907\u80fd\u8017\u3002\u5728\u79fb\u52a8\u8bbe\u5907\u548c\u5d4c\u5165\u5f0f\u7cfb\u7edf\u4e2d\uff0c\u529f\u8017\u662f\u5173\u952e\u7684\u8bbe\u8ba1\u8003\u91cf\u56e0\u7d20\u3002\u91cf\u5316\u540e\u7684\u6a21\u578b\u6709\u52a9\u4e8e\u5ef6\u957f\u8bbe\u5907\u7535\u6c60\u7eed\u822a\u65f6\u95f4\uff0c\u5e76\u51cf\u5c11\u6563\u70ed\u9700\u6c42\u3002<\/p>\n<p><strong>\u6a21\u578b\u4f18\u5316<\/strong>: \u91cf\u5316\u6280\u672f\u63a8\u52a8\u6a21\u578b\u5728\u8bad\u7ec3\u548c\u63a8\u7406\u8fc7\u7a0b\u4e2d\u8fdb\u884c\u4f18\u5316\u4e0e\u538b\u7f29\uff0c\u4ece\u800c\u8fdb\u4e00\u6b65\u964d\u4f4e\u90e8\u7f72\u6210\u672c\u3002\u4f8b\u5982\uff0cFP8 \u7684\u5e94\u7528\u4fc3\u4f7f\u6a21\u578b\u5728\u8bad\u7ec3\u9636\u6bb5\u8fdb\u884c\u66f4\u4e3a\u7cbe\u7ec6\u7684\u91cf\u5316\u7b56\u7565\u63a2\u7d22\uff0c\u8fdb\u800c\u63d0\u5347\u6a21\u578b\u6574\u4f53\u6548\u7387\u3002<\/p>\n<h4>\u5c40\u9650\u6027<\/h4>\n<p><strong>\u7cbe\u5ea6\u635f\u5931<\/strong>: \u91cf\u5316\u8fc7\u7a0b\u4e0d\u53ef\u907f\u514d\u5730\u4f34\u968f\u7cbe\u5ea6\u635f\u5931\uff0c\u5c24\u5176\u662f\u5728\u91c7\u7528\u6781\u4f4e\u7cbe\u5ea6\u683c\u5f0f (\u5982 FP8) \u65f6\uff0c\u7cbe\u5ea6\u635f\u5931\u53ef\u80fd\u8f83\u4e3a\u663e\u8457\uff0c\u5e76\u53ef\u80fd\u5bfc\u81f4\u6a21\u578b\u5728\u7279\u5b9a\u4efb\u52a1\u4e0a\u7684\u6027\u80fd\u4e0b\u964d\u3002\u5c3d\u7ba1\u53ef\u4ee5\u901a\u8fc7\u7cbe\u5de7\u5730\u9009\u62e9\u7f29\u653e\u56e0\u5b50\u7b49\u624b\u6bb5\u5728\u4e00\u5b9a\u7a0b\u5ea6\u4e0a\u7f13\u89e3\u7cbe\u5ea6\u635f\u5931\uff0c\u4f46\u5b8c\u5168\u6d88\u9664\u7cbe\u5ea6\u635f\u5931\u901a\u5e38\u96be\u4ee5\u5b9e\u73b0\u3002\u00a0<strong>\u4e3a\u4e86\u51cf\u8f7b\u91cf\u5316\u5e26\u6765\u7684\u7cbe\u5ea6\u635f\u5931\uff0c \u53ef\u4ee5\u91c7\u7528\u91cf\u5316\u611f\u77e5\u8bad\u7ec3 (Quantization-Aware Training, QAT) \u6280\u672f\u3002 QAT \u5728\u6a21\u578b\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0c \u6a21\u62df\u91cf\u5316\u64cd\u4f5c\uff0c \u5c06\u91cf\u5316\u8bef\u5dee\u7eb3\u5165\u8bad\u7ec3\u7684\u8003\u91cf\u8303\u56f4\uff0c \u4ece\u800c\u8bad\u7ec3\u51fa\u5bf9\u91cf\u5316\u66f4\u9c81\u68d2\u7684\u6a21\u578b\u3002 QAT \u901a\u5e38\u80fd\u591f\u663e\u8457\u63d0\u5347\u91cf\u5316\u6a21\u578b\u7684\u7cbe\u5ea6\uff0c \u4f46\u8bad\u7ec3\u6210\u672c\u4e5f\u4f1a\u76f8\u5e94\u589e\u52a0\u3002<\/strong><\/p>\n<p><strong>\u786c\u4ef6\u652f\u6301<\/strong>: \u5e76\u975e\u6240\u6709\u786c\u4ef6\u5e73\u53f0\u5747\u80fd\u5168\u9762\u652f\u6301\u4f4e\u7cbe\u5ea6\u8ba1\u7b97\u3002\u4f8b\u5982\uff0cFP8 \u548c FP16 \u8ba1\u7b97\u901a\u5e38\u9700\u8981\u7279\u5b9a\u786c\u4ef6 (\u5982 NVIDIA Hopper \u67b6\u6784 GPU) \u7684\u652f\u6301\u3002\u82e5\u786c\u4ef6\u5e73\u53f0\u7f3a\u4e4f\u5bf9\u4f4e\u7cbe\u5ea6\u8ba1\u7b97\u7684\u4f18\u5316\uff0c\u91cf\u5316\u6280\u672f\u7684\u4f18\u52bf\u5c06\u96be\u4ee5\u5145\u5206\u53d1\u6325\u3002\u00a0<strong>\u968f\u7740\u4f4e\u7cbe\u5ea6\u8ba1\u7b97\u7684\u666e\u53ca\uff0c \u8d8a\u6765\u8d8a\u591a\u7684\u786c\u4ef6\u5e73\u53f0\u5f00\u59cb\u652f\u6301 FP16\u3001 BF16 \u751a\u81f3 FP8 \u7b49\u4f4e\u7cbe\u5ea6\u6570\u636e\u7c7b\u578b\uff0c \u4e3a\u91cf\u5316\u6280\u672f\u7684\u5e7f\u6cdb\u5e94\u7528\u63d0\u4f9b\u4e86\u66f4\u575a\u5b9e\u7684\u786c\u4ef6\u57fa\u7840\u3002<\/strong><\/p>\n<p><strong>\u590d\u6742\u6027\u63d0\u5347<\/strong>: \u91cf\u5316\u6d41\u7a0b\u672c\u8eab\u4f1a\u589e\u52a0\u6a21\u578b\u5f00\u53d1\u7684\u590d\u6742\u6027\u3002\u4f8b\u5982\uff0c\u91cf\u5316\u8fc7\u7a0b\u9700\u8981\u7cbe\u7ec6\u5730\u8ba1\u7b97\u7f29\u653e\u56e0\u5b50\uff0c\u5e76\u53ef\u80fd\u9700\u8981\u5bf9\u6a21\u578b\u8fdb\u884c\u989d\u5916\u7684\u6821\u51c6\u548c\u5fae\u8c03\u3002\u8fd9\u65e0\u7591\u589e\u52a0\u4e86\u6a21\u578b\u5f00\u53d1\u548c\u90e8\u7f72\u7684\u96be\u5ea6\u3002\u00a0<strong>\u4e3a\u4e86\u964d\u4f4e\u91cf\u5316\u90e8\u7f72\u7684\u590d\u6742\u6027\uff0c \u4e1a\u754c\u4e5f\u6d8c\u73b0\u51fa\u8bb8\u591a\u81ea\u52a8\u5316\u91cf\u5316\u5de5\u5177\u548c\u5e73\u53f0\uff0c \u4f8b\u5982 NVIDIA TensorRT\u3001 Qualcomm AI Engine \u7b49\uff0c \u8fd9\u4e9b\u5de5\u5177\u53ef\u4ee5\u5e2e\u52a9\u5f00\u53d1\u8005\u5feb\u901f\u3001\u4fbf\u6377\u5730\u5c06\u6a21\u578b\u91cf\u5316\u5e76\u90e8\u7f72\u5230\u76ee\u6807\u786c\u4ef6\u5e73\u53f0\u3002<\/strong><\/p>\n<p><strong>\u5e94\u7528\u573a\u666f<\/strong>: \u91cf\u5316\u6280\u672f\u5e76\u975e\u5728\u6240\u6709\u5e94\u7528\u573a\u666f\u4e0b\u7686\u9002\u7528\u3002\u5728\u5bf9\u7cbe\u5ea6\u8981\u6c42\u6781\u9ad8\u7684\u4efb\u52a1\u4e2d (\u5982\u79d1\u5b66\u8ba1\u7b97\u3001\u91d1\u878d\u5efa\u6a21\u7b49)\uff0c\u91cf\u5316\u53ef\u80fd\u5bfc\u81f4\u65e0\u6cd5\u63a5\u53d7\u7684\u7cbe\u5ea6\u635f\u5931\u3002\u00a0<strong>\u5bf9\u4e8e\u7cbe\u5ea6\u654f\u611f\u578b\u4efb\u52a1\uff0c \u53ef\u4ee5\u5c1d\u8bd5\u91c7\u7528\u6df7\u5408\u7cbe\u5ea6\u91cf\u5316\u7b56\u7565\uff0c \u5bf9\u6a21\u578b\u4e2d\u4e0d\u540c\u5c42\u6216\u4e0d\u540c\u53c2\u6570\u91c7\u7528\u4e0d\u540c\u7684\u91cf\u5316\u7cbe\u5ea6\uff0c \u4f8b\u5982\u5bf9\u5173\u952e\u5c42\u6216\u53c2\u6570\u4ecd\u4fdd\u6301 FP32 \u6216 FP16 \u7cbe\u5ea6\uff0c \u800c\u5bf9\u5176\u4ed6\u5c42\u6216\u53c2\u6570\u5219\u91c7\u7528 INT8 \u6216\u66f4\u4f4e\u7cbe\u5ea6\u91cf\u5316\uff0c \u4ee5\u5728\u7cbe\u5ea6\u548c\u6548\u7387\u4e4b\u95f4\u53d6\u5f97\u66f4\u597d\u7684\u5e73\u8861\u3002<\/strong><\/p>\n<p>\u7efc\u4e0a\u6240\u8ff0\uff0c\u91cf\u5316\u6280\u672f\u4f5c\u4e3a\u6df1\u5ea6\u5b66\u4e60\u9886\u57df\u7684\u4e00\u9879\u5173\u952e\u6280\u672f\uff0c\u5728\u63d0\u5347\u8ba1\u7b97\u6548\u7387\u3001\u964d\u4f4e\u5b58\u50a8\u9700\u6c42\u548c\u529f\u8017\u7b49\u65b9\u9762\u5c55\u73b0\u51fa\u5de8\u5927\u7684\u6f5c\u529b\u3002\u7136\u800c\uff0c\u91cf\u5316\u6280\u672f\u4ea6\u5b58\u5728\u7cbe\u5ea6\u635f\u5931\u548c\u786c\u4ef6\u4f9d\u8d56\u6027\u7b49\u5c40\u9650\u6027\u3002\u56e0\u6b64\uff0c\u5728\u5b9e\u9645\u5e94\u7528\u4e2d\uff0c\u9700\u6839\u636e\u5177\u4f53\u7684\u5e94\u7528\u9700\u6c42\u548c\u573a\u666f\u7279\u70b9\uff0c\u5ba1\u614e\u9009\u62e9\u91cf\u5316\u7b56\u7565\uff0c\u4ee5\u671f\u5728\u6a21\u578b\u6027\u80fd\u4e0e\u6548\u7387\u4e4b\u95f4\u8fbe\u5230\u6700\u4f18\u5e73\u8861\u3002\u00a0<strong>\u672a\u6765\uff0c \u968f\u7740\u786c\u4ef6\u548c\u7b97\u6cd5\u7684\u4e0d\u65ad\u53d1\u5c55\uff0c \u91cf\u5316\u6280\u672f\u5c06\u5728\u6df1\u5ea6\u5b66\u4e60\u9886\u57df\u626e\u6f14\u66f4\u52a0\u91cd\u8981\u7684\u89d2\u8272\uff0c \u63a8\u52a8\u4eba\u5de5\u667a\u80fd\u5e94\u7528\u5728\u66f4\u5e7f\u6cdb\u7684\u573a\u666f\u4e2d\u843d\u5730\u751f\u6839\u3002<\/strong><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u5bfc\u8bfb \u5728\u4eba\u5de5\u667a\u80fd\u79d1\u6280\u7684\u6d69\u701a\u661f\u7a7a\u4e2d\uff0c\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u4ee5\u5176\u5353\u8d8a\u7684\u6027\u80fd\uff0c\u9a71\u52a8\u7740\u8bf8\u591a\u9886\u57df\u7684\u521b\u65b0\u53d1\u5c55\u3002\u7136\u800c\uff0c\u6a21\u578b\u89c4\u6a21\u7684\u6301\u7eed\u81a8\u80c0\uff0c\u5982\u540c\u53cc\u5203\u5251\uff0c\u5728\u63d0\u5347\u6027\u80fd\u7684\u540c\u65f6\uff0c\u4e5f\u5e26\u6765\u4e86\u7b97\u529b\u9700\u6c42\u4e0e\u5b58\u50a8\u538b\u529b\u7684\u5267\u589e\u3002\u7279\u522b\u662f\u5728\u8d44\u6e90\u53d7\u9650\u7684\u5e94\u7528\u573a\u666f\u4e2d\uff0c\u6a21\u578b\u7684\u90e8\u7f72\u548c\u8fd0\u884c\u9762\u4e34\u4e25\u5cfb\u6311\u6218\u3002 &#8230;<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[34],"tags":[],"class_list":["post-20978","post","type-post","status-publish","format-standard","hentry","category-knowledge"],"_links":{"self":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/posts\/20978","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=20978"}],"version-history":[{"count":0,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/posts\/20978\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/media?parent=20978"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/categories?post=20978"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/tags?post=20978"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}