{"id":1078164,"date":"2024-08-20T19:16:00","date_gmt":"2024-08-21T02:16:00","guid":{"rendered":"https:\/\/www.microsoft.com\/en-us\/research\/?post_type=msr-blog-post&#038;p=1078164"},"modified":"2024-09-25T04:25:22","modified_gmt":"2024-09-25T11:25:22","slug":"low-bit-quantization","status":"publish","type":"msr-blog-post","link":"https:\/\/www.microsoft.com\/en-us\/research\/articles\/low-bit-quantization\/","title":{"rendered":"\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u591a\u9879\u521b\u65b0\u6280\u672f\uff0c\u5f25\u5408\u5927\u6a21\u578b\u4f4e\u6bd4\u7279\u91cf\u5316\u4e0e\u7ec8\u7aef\u90e8\u7f72\u95f4\u9e3f\u6c9f"},"content":{"rendered":"\n<p>\u7f16\u8005\u6309\uff1a\u5728\u4eba\u5de5\u667a\u80fd\u9886\u57df\uff0c\u6a21\u578b\u53c2\u6570\u7684\u589e\u591a\u5f80\u5f80\u610f\u5473\u7740\u6027\u80fd\u7684\u63d0\u5347\u3002\u4f46\u968f\u7740\u6a21\u578b\u89c4\u6a21\u7684\u6269\u5927\uff0c\u5176\u5bf9\u7ec8\u7aef\u8bbe\u5907\u7684\u7b97\u529b\u4e0e\u5185\u5b58\u9700\u6c42\u4e5f\u65e5\u76ca\u589e\u52a0\u3002\u4f4e\u6bd4\u7279\u91cf\u5316\u6280\u672f\uff0c\u7531\u4e8e\u53ef\u4ee5\u5927\u5e45\u964d\u4f4e\u5b58\u50a8\u548c\u8ba1\u7b97\u6210\u672c\u5e76\u63d0\u5347\u63a8\u7406\u6548\u7387\uff0c\u5df2\u6210\u4e3a\u5b9e\u73b0\u5927\u6a21\u578b\u5728\u8d44\u6e90\u53d7\u9650\u8bbe\u5907\u4e0a\u9ad8\u6548\u8fd0\u884c\u7684\u5173\u952e\u6280\u672f\u4e4b\u4e00\u3002\u7136\u800c\uff0c\u5982\u679c\u786c\u4ef6\u8bbe\u5907\u4e0d\u652f\u6301\u4f4e\u6bd4\u7279\u91cf\u5316\u540e\u7684\u6570\u636e\u6a21\u5f0f\uff0c\u90a3\u4e48\u4f4e\u6bd4\u7279\u91cf\u5316\u7684\u4f18\u52bf\u5c06\u65e0\u6cd5\u53d1\u6325\u3002<\/p>\n\n\n\n<p>\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e00\u95ee\u9898\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u63a8\u51fa\u4e86\u5168\u65b0\u7684\u6570\u636e\u7f16\u8bd1\u5668 Ladder \u548c\u7b97\u6cd5 T-MAC\uff0c\u4f7f\u5f53\u524d\u53ea\u652f\u6301\u5bf9\u79f0\u7cbe\u5ea6\u8ba1\u7b97\u7684\u786c\u4ef6\u80fd\u591f\u76f4\u63a5\u8fd0\u884c\u6df7\u5408\u7cbe\u5ea6\u77e9\u9635\u4e58\u6cd5\u3002\u6d4b\u8bd5\u7ed3\u679c\u8868\u660e\uff0cLadder \u5728\u652f\u6301 GPU \u539f\u672c\u4e0d\u652f\u6301\u7684\u81ea\u5b9a\u4e49\u6570\u636e\u7c7b\u578b\u65b9\u9762\uff0c\u6700\u9ad8\u63d0\u901f\u53ef\u8fbe14.6\u500d\uff1bT-MAC \u5728\u642d\u8f7d\u4e86\u6700\u65b0\u9ad8\u901a Snapdragon X Elite \u82af\u7247\u7ec4\u7684 Surface AI PC \u4e0a\uff0c\u4f7f CPU \u4e0a\u8fd0\u884c\u7684\u5927\u6a21\u578b\u541e\u5410\u7387\u6bd4\u4e13\u7528\u52a0\u901f\u5668 NPU \u5feb\u4e24\u500d\u3002\u6b64\u5916\uff0c\u7814\u7a76\u5458\u4eec\u8fd8\u8bbe\u8ba1\u4e86 LUT Tensor Core \u786c\u4ef6\u67b6\u6784\uff0c\u8fd9\u79cd\u7cbe\u7b80\u8bbe\u8ba1\u4f7f\u786c\u4ef6\u80fd\u591f\u76f4\u63a5\u652f\u6301\u5404\u79cd\u4f4e\u6bd4\u7279\u6df7\u5408\u7cbe\u5ea6\u8ba1\u7b97\uff0c\u4e3a\u4eba\u5de5\u667a\u80fd\u786c\u4ef6\u8bbe\u8ba1\u63d0\u4f9b\u4e86\u65b0\u601d\u8def\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<p>\u5927\u6a21\u578b\u5df2\u7ecf\u8d8a\u6765\u8d8a\u591a\u5730\u88ab\u90e8\u7f72\u5728\u667a\u80fd\u624b\u673a\u3001\u7b14\u8bb0\u672c\u7535\u8111\u3001\u673a\u5668\u4eba\u7b49\u7aef\u4fa7\u8bbe\u5907\u4e0a\uff0c\u4ee5\u63d0\u4f9b\u5148\u8fdb\u7684\u667a\u80fd\u53ca\u5b9e\u65f6\u54cd\u5e94\u670d\u52a1\u3002\u4f46\u5305\u542b\u4e0a\u4ebf\u53c2\u6570\u7684\u5927\u6a21\u578b\u5bf9\u7ec8\u7aef\u8bbe\u5907\u7684\u5185\u5b58\u548c\u8ba1\u7b97\u80fd\u529b\u63d0\u51fa\u4e86\u6781\u9ad8\u7684\u8981\u6c42\uff0c\u4e5f\u56e0\u6b64\u9650\u5236\u4e86\u5b83\u4eec\u7684\u5e7f\u6cdb\u5e94\u7528\u3002\u4f4e\u6bd4\u7279\u91cf\u5316\u6280\u672f\u56e0\u5176\u80fd\u663e\u8457\u538b\u7f29\u6a21\u578b\u89c4\u6a21\uff0c\u964d\u4f4e\u5bf9\u8ba1\u7b97\u8d44\u6e90\u7684\u9700\u6c42\uff0c\u6210\u4e3a\u4e86\u5927\u6a21\u578b\u5728\u7aef\u4fa7\u90e8\u7f72\u548c\u5b9e\u73b0\u9ad8\u6548\u63a8\u7406\u7684\u6709\u6548\u624b\u6bb5\u3002<\/p>\n\n\n\n<p>\u968f\u7740\u4f4e\u6bd4\u7279\u91cf\u5316\u6280\u672f\u7684\u53d1\u5c55\uff0c\u6570\u636e\u7c7b\u578b\u65e5\u76ca\u591a\u6837\u5316\uff0c\u5982 int4\u3001int2\u3001int1 \u7b49\u4f4e\u6bd4\u7279\u6570\u636e\uff0c\u4f7f\u5f97\u5927\u6a21\u578b\u5728\u63a8\u7406\u4e2d\u8d8a\u6765\u8d8a\u591a\u5730\u91c7\u7528\u4f4e\u6bd4\u7279\u6743\u91cd\u548c\u9ad8\u6bd4\u7279\u6743\u91cd\u8ba1\u7b97\u7684\u6df7\u5408\u7cbe\u5ea6\u77e9\u9635\u4e58\u6cd5\uff08mixed-precision matrix multiplication\uff0cmpGEMM\uff09\u3002\u7136\u800c\uff0c\u73b0\u6709\u7684 CPU\u3001GPU \u7b49\u786c\u4ef6\u8ba1\u7b97\u5355\u5143\u901a\u5e38\u53ea\u652f\u6301\u5bf9\u79f0\u8ba1\u7b97\u6a21\u5f0f\uff0c\u5e76\u4e0d\u517c\u5bb9\u8fd9\u79cd\u6df7\u5408\u7cbe\u5ea6\u7684\u77e9\u9635\u4e58\u6cd5\u3002<\/p>\n\n\n\n<p>\u6df7\u5408\u7cbe\u5ea6\u77e9\u9635\u4e58\u6cd5\u4e0e\u4f20\u7edf\u7684\u77e9\u9635\u4e58\u6cd5\u6709\u4f55\u4e0d\u540c\uff1f<\/p>\n\n\n\n<p>\u5728\u4f20\u7edf\u7684\u77e9\u9635\u4e58\u6cd5\u4e2d\uff0c\u53c2\u4e0e\u8fd0\u7b97\u7684\u4e24\u7aef\u6570\u503c\u662f\u5bf9\u79f0\u7684\uff0c\u4f8b\u5982 FP16*FP16\u3001int8*int8\u3002\u4f46\u5927\u6a21\u578b\u7684\u4f4e\u6bd4\u7279\u91cf\u5316\u6253\u7834\u4e86\u8fd9\u79cd\u5bf9\u79f0\u6027\uff0c\u4f7f\u4e58\u6cd5\u7684\u4e00\u7aef\u662f\u9ad8\u6bd4\u7279\uff0c\u53e6\u4e00\u7aef\u662f\u4f4e\u6bd4\u7279\uff0c\u4f8b\u5982\u5728 1-bit \u7684&nbsp;<a class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" target=\"_blank\" href=\"http:\/\/mp.weixin.qq.com\/s?__biz=MzAwMTA3MzM4Nw==&mid=2649498640&idx=1&sn=a860101ceee6bc3a777f465bdd1586da&chksm=82c7cd94b5b0448231f0017d2694e59f6e41369ea14a38a3a19a32a9ba18c3fe0f934e214bee&scene=21#wechat_redirect\" rel=\"noreferrer noopener\">BitNet<span class=\"sr-only\"> (opens in new tab)<\/span><\/a>&nbsp;\u6a21\u578b\u4e2d\u5b9e\u73b0\u7684 int8*int1 \u6216 int8*int2\uff0c\u4ee5\u53ca\u6d6e\u70b9\u6570\u4e0e\u6574\u6570\u7684\u6df7\u5408\u4e58\u6cd5 FP16*int4\u3002<\/p>\n\n\n\n<p>\u4e3a\u4e86\u5145\u5206\u53d1\u6325\u4f4e\u6bd4\u7279\u91cf\u5316\u7684\u4f18\u52bf\uff0c\u8ba9\u786c\u4ef6\u8bbe\u5907\u80fd\u591f\u76f4\u63a5\u652f\u6301\u6df7\u5408\u7cbe\u5ea6\u77e9\u9635\u4e58\u6cd5\uff0c\u786e\u4fdd\u5927\u6a21\u578b\u5728\u7aef\u4fa7\u8bbe\u5907\u4e0a\u7684\u9ad8\u901f\u6709\u6548\u8fd0\u884c\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u9488\u5bf9\u73b0\u6709 CPU\u3001GPU \u8ba1\u7b97\u7b97\u5b50\u548c\u786c\u4ef6\u67b6\u6784\u8fdb\u884c\u521b\u65b0\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u63a8\u51fa\u4e86\u6570\u636e\u7c7b\u578b\u7f16\u8bd1\u5668 Ladder\uff0c\u652f\u6301\u5404\u79cd\u4f4e\u7cbe\u5ea6\u6570\u636e\u7c7b\u578b\u7684\u8868\u8fbe\u548c\u76f8\u4e92\u8f6c\u6362\uff0c\u5c06\u786c\u4ef6\u4e0d\u652f\u6301\u7684\u6570\u636e\u7c7b\u578b\u65e0\u635f\u8f6c\u6362\u4e3a\u786c\u4ef6\u652f\u6301\u7684\u6570\u636e\u7c7b\u578b\u6307\u4ee4\uff0c\u5728\u4f20\u7edf\u8ba1\u7b97\u6a21\u5f0f\u4e0b\uff0c\u4f7f\u5f97\u786c\u4ef6\u80fd\u591f\u652f\u6301\u6df7\u5408\u7cbe\u5ea6\u7684 DNN\uff08\u6df1\u5ea6\u795e\u7ecf\u7f51\u7edc\uff09\u8ba1\u7b97\uff1b<\/li>\n\n\n\n<li>\u7814\u53d1\u4e86\u5168\u65b0\u7b97\u6cd5 T-MAC\uff0c\u57fa\u4e8e\u67e5\u627e\u8868\uff08Lookup Table\uff0cLUT\uff09\u7684\u65b9\u6cd5\uff0c\u5b9e\u73b0\u4e86\u786c\u4ef6\u5bf9\u6df7\u5408\u7cbe\u5ea6\u77e9\u9635\u4e58\u6cd5\u7684\u76f4\u63a5\u652f\u6301\uff0c\u8f6f\u4ef6\u5c42\u9762\uff0c\u5728 CPU \u4e0a\u7684\u8ba1\u7b97\u76f8\u6bd4\u4f20\u7edf\u8ba1\u7b97\u6a21\u5f0f\u53d6\u5f97\u4e86\u66f4\u597d\u7684\u52a0\u901f\uff1b<\/li>\n\n\n\n<li>\u63d0\u51fa\u4e86\u65b0\u7684\u786c\u4ef6\u67b6\u6784 LUT Tensor Core\uff0c\u4e3a\u4e0b\u4e00\u4ee3\u4eba\u5de5\u667a\u80fd\u786c\u4ef6\u8bbe\u8ba1\u6253\u5f00\u4e86\u65b0\u601d\u8def\u3002<\/li>\n<\/ul>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"ladder-\u81ea\u5b9a\u4e49\u6570\u636e\u7c7b\u578b\u65e0\u635f\u8f6c\u6362\u6210\u786c\u4ef6\u652f\u6301\u7684\u6570\u636e\u7c7b\u578b\">Ladder\uff1a\u81ea\u5b9a\u4e49\u6570\u636e\u7c7b\u578b\u65e0\u635f\u8f6c\u6362\u6210\u786c\u4ef6\u652f\u6301\u7684\u6570\u636e\u7c7b\u578b<\/h2>\n\n\n\n<p>\u5f53\u524d\uff0c\u524d\u6cbf\u52a0\u901f\u5668\u6b63\u5728\u5c06\u66f4\u4f4e\u6bd4\u7279\u7684\u8ba1\u7b97\u5355\u5143\uff0c\u5982 FP32\u3001FP16\uff0c\u751a\u81f3 FP8 \u7684\u8fd0\u7b97\u96c6\u6210\u5230\u65b0\u4e00\u4ee3\u7684\u67b6\u6784\u4e2d\u3002\u7136\u800c\uff0c\u53d7\u9650\u4e8e\u82af\u7247\u9762\u79ef\u548c\u9ad8\u6602\u7684\u786c\u4ef6\u6210\u672c\uff0c\u6bcf\u4e2a\u52a0\u901f\u5668\u53ea\u80fd\u4e3a\u6807\u51c6\u7684\u6570\u636e\u7c7b\u578b\u63d0\u4f9b\u6709\u9650\u7c7b\u578b\u7684\u8ba1\u7b97\u5355\u5143\uff0c\u6bd4\u5982 NVIDIA V100 TENSOR CORE GPU \u4ec5\u652f\u6301 FP16\uff0c\u800c A100 \u867d\u7136\u52a0\u5165\u4e86\u5bf9 int2\u3001int4\u3001int8 \u7684\u652f\u6301\uff0c\u4f46\u5e76\u672a\u6db5\u76d6\u66f4\u65b0\u7684 FP8 \u6216 OCP-MXFP \u7b49\u6570\u636e\u683c\u5f0f\u3002\u6b64\u5916\uff0c\u5927\u6a21\u578b\u7684\u5feb\u901f\u8fed\u4ee3\u4e0e\u786c\u4ef6\u5347\u7ea7\u7684\u7f13\u6162\u6b65\u4f10\u4e4b\u95f4\u5b58\u5728\u5dee\u8ddd\uff0c\u5bfc\u81f4\u8bb8\u591a\u65b0\u6570\u636e\u7c7b\u578b\u65e0\u6cd5\u5f97\u5230\u786c\u4ef6\u652f\u6301\uff0c\u8fdb\u800c\u5f71\u54cd\u5927\u6a21\u578b\u7684\u52a0\u901f\u548c\u8fd0\u884c\u3002<\/p>\n\n\n\n<p>\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u53d1\u73b0\uff0c\u5c3d\u7ba1\u786c\u4ef6\u52a0\u901f\u5668\u7f3a\u4e4f\u9488\u5bf9\u81ea\u5b9a\u4e49\u6570\u636e\u7c7b\u578b\u7684\u8ba1\u7b97\u6307\u4ee4\uff0c\u4f46\u5176\u5185\u5b58\u7cfb\u7edf\u53ef\u4ee5\u5c06\u5b83\u4eec\u8f6c\u6362\u4e3a\u56fa\u5b9a\u4f4d\u5bbd\u7684\u4e0d\u900f\u660e\u6570\u636e\u5757\u6765\u5b58\u50a8\u4efb\u610f\u6570\u636e\u7c7b\u578b\u3002\u540c\u65f6\uff0c\u5927\u591a\u6570\u81ea\u5b9a\u4e49\u6570\u636e\u7c7b\u578b\u53ef\u4ee5\u65e0\u635f\u5730\u8f6c\u6362\u4e3a\u73b0\u6709\u786c\u4ef6\u8ba1\u7b97\u5355\u5143\u652f\u6301\u7684\u66f4\u591a\u4f4d\u7684\u6807\u51c6\u6570\u636e\u7c7b\u578b\u3002\u4f8b\u5982\uff0cNF4 \u5f20\u91cf\u53ef\u4ee5\u8f6c\u6362\u6210 FP16 \u6216 FP32 \u4ee5\u6267\u884c\u6d6e\u70b9\u8fd0\u7b97\u3002<\/p>\n\n\n\n<p>\u57fa\u4e8e\u8fd9\u4e9b\u53d1\u73b0\uff0c\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd<strong>\u901a\u8fc7\u5206\u79bb\u6570\u636e\u5b58\u50a8\u548c\u8ba1\u7b97\u6765\u652f\u6301\u6240\u6709\u81ea\u5b9a\u4e49\u6570\u636e\u7c7b\u578b\u7684\u65b9\u6cd5\uff0c\u5e76\u7814\u53d1\u4e86\u6570\u636e\u7f16\u8bd1\u5668&nbsp;<\/strong><strong>Ladder\uff0c\u4ee5\u5f25\u5408\u4e0d\u65ad\u51fa\u73b0\u7684\u81ea\u5b9a\u4e49\u6570\u636e\u7c7b\u578b\u4e0e\u5f53\u524d\u786c\u4ef6\u652f\u6301\u7684\u56fa\u6709\u7cbe\u5ea6\u683c\u5f0f\u4e4b\u95f4\u7684\u5dee\u8ddd\u3002<\/strong><\/p>\n\n\n\n<p>Ladder \u5b9a\u4e49\u4e86\u4e00\u5957\u6570\u636e\u7c7b\u578b\u7cfb\u7edf\uff0c\u5305\u62ec\u6570\u636e\u7c7b\u578b\u4e4b\u95f4\u65e0\u635f\u8f6c\u6362\u7684\u62bd\u8c61\uff0c\u5b83\u80fd\u591f\u8868\u793a\u7b97\u6cd5\u548c\u786c\u4ef6\u652f\u6301\u7684\u5404\u79cd\u6570\u636e\u7c7b\u578b\uff0c\u5e76\u5b9a\u4e49\u4e86\u6570\u636e\u7c7b\u578b\u4e4b\u95f4\u7684\u8f6c\u6362\u89c4\u5219\u3002\u5f53\u5904\u7406\u4f4e\u6bd4\u7279\u7b97\u6cd5\u5e94\u7528\u65f6\uff0cLadder \u901a\u8fc7\u4e00\u7cfb\u5217\u4f18\u5316\uff0c\u5c06\u4f4e\u6bd4\u7279\u6570\u636e\u8f6c\u8bd1\u6210\u5f53\u524d\u786c\u4ef6\u4e0a\u6700\u9ad8\u6548\u7684\u6267\u884c\u683c\u5f0f\uff0c\u5305\u62ec\u5bf9\u8ba1\u7b97\u548c\u5b58\u50a8\u7684\u4f18\u5316\u2014\u2014\u5c06\u7b97\u6cd5\u6620\u5c04\u5230\u5339\u914d\u7684\u8ba1\u7b97\u6307\u4ee4\uff0c\u5e76\u5c06\u4e0d\u540c\u683c\u5f0f\u7684\u6570\u636e\u5b58\u50a8\u5230\u4e0d\u540c\u7ea7\u522b\u7684\u5b58\u50a8\u5355\u5143\u4e2d\uff0c\u4ee5\u5b9e\u73b0\u6700\u9ad8\u6548\u7684\u8fd0\u7b97\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"465\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/uploads\/prod\/2024\/08\/low-bit-quantization-1-1024x465.png\" alt=\"Ladder \u7684\u7cfb\u7edf\u67b6\u6784\" class=\"wp-image-1078170\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/uploads\/prod\/2024\/08\/low-bit-quantization-1-1024x465.png 1024w, https:\/\/www.microsoft.com\/en-us\/research\/uploads\/prod\/2024\/08\/low-bit-quantization-1-300x136.png 300w, https:\/\/www.microsoft.com\/en-us\/research\/uploads\/prod\/2024\/08\/low-bit-quantization-1-768x349.png 768w, https:\/\/www.microsoft.com\/en-us\/research\/uploads\/prod\/2024\/08\/low-bit-quantization-1-1536x698.png 1536w, https:\/\/www.microsoft.com\/en-us\/research\/uploads\/prod\/2024\/08\/low-bit-quantization-1-240x109.png 240w, https:\/\/www.microsoft.com\/en-us\/research\/uploads\/prod\/2024\/08\/low-bit-quantization-1.png 1607w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><figcaption class=\"wp-element-caption\">\u56fe1\uff1aLadder \u7684\u7cfb\u7edf\u67b6\u6784<\/figcaption><\/figure>\n\n\n\n<p><\/p>\n\n\n\n<p>\u5728 NVIDIA A100\u3001NVIDIA V100\u3001NVIDIA RTX A6000\u3001NVIDIA RTX 4090 \u548c AMD Instinct MI250 GPU \u4e0a\u8fd0\u884c\u7684 DNN \u63a8\u7406\u6027\u80fd\u8bc4\u4f30\u663e\u793a\uff0cLadder \u5728\u539f\u751f\u652f\u6301\u6570\u636e\u7c7b\u578b\u4e0a\u8d85\u8d8a\u4e86\u73b0\u6709\u6700\u5148\u8fdb\u7684 DNN \u7f16\u8bd1\u5668\uff0c\u5e76\u4e14\u5728\u652f\u6301 GPU \u539f\u672c\u4e0d\u652f\u6301\u7684\u81ea\u5b9a\u4e49\u6570\u636e\u7c7b\u578b\u65b9\u9762\u8868\u73b0\u51fa\u8272\uff0c\u6700\u9ad8\u63d0\u901f\u53ef\u8fbe14.6\u500d\u3002<\/p>\n\n\n\n<p><strong>Ladder \u662f\u9996\u4e2a\u5728\u73b0\u4ee3\u786c\u4ef6\u52a0\u901f\u5668\u4e0a\u8fd0\u884c DNN \u65f6\uff0c\u53ef\u4ee5\u7cfb\u7edf\u6027\u5730\u652f\u6301\u4ee5\u81ea\u5b9a\u4e49\u6570\u636e\u7c7b\u578b\u8868\u793a\u4f4e\u6bd4\u7279\u7cbe\u5ea6\u6570\u636e\u7684\u7cfb\u7edf\u3002<\/strong>\u8fd9\u4e3a\u6a21\u578b\u7814\u7a76\u8005\u63d0\u4f9b\u4e86\u66f4\u7075\u6d3b\u7684\u6570\u636e\u7c7b\u578b\u4f18\u5316\u65b9\u6cd5\uff0c\u540c\u65f6\u4e5f\u8ba9\u786c\u4ef6\u67b6\u6784\u5f00\u53d1\u8005\u5728\u4e0d\u6539\u53d8\u786c\u4ef6\u7684\u60c5\u51b5\u4e0b\uff0c\u652f\u6301\u66f4\u5e7f\u6cdb\u7684\u6570\u636e\u7c7b\u578b\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"t-mac-\u65e0\u9700\u4e58\u6cd5\u7684\u901a\u7528\u4f4e\u6bd4\u7279\u6df7\u5408\u7cbe\u5ea6\u77e9\u9635\u4e58\u8ba1\u7b97\">T-MAC\uff1a\u65e0\u9700\u4e58\u6cd5\u7684\u901a\u7528\u4f4e\u6bd4\u7279\u6df7\u5408\u7cbe\u5ea6\u77e9\u9635\u4e58\u8ba1\u7b97<\/h2>\n\n\n\n<p>\u4e3a\u4e86\u8ba9\u73b0\u6709\u786c\u4ef6\u8bbe\u5907\u652f\u6301\u4e0d\u540c\u7684\u6570\u636e\u6a21\u5f0f\u548c\u6df7\u5408\u7cbe\u5ea6\u77e9\u9635\u4e58\u6cd5\uff0c\u5728\u7aef\u4fa7\u90e8\u7f72\u5927\u6a21\u578b\u65f6\uff0c\u5e38\u89c1\u7684\u505a\u6cd5\u662f\u5bf9\u4f4e\u6bd4\u7279\u6a21\u578b\u8fdb\u884c\u53cd\u91cf\u5316\u3002\u7136\u800c\uff0c\u8fd9\u79cd\u65b9\u6cd5\u5b58\u5728\u4e24\u5927\u95ee\u9898\uff1a\u9996\u5148\uff0c\u4ece\u6027\u80fd\u89d2\u5ea6\u6765\u770b\uff0c\u53cd\u91cf\u5316\u8fc7\u7a0b\u4e2d\u7684\u8f6c\u6362\u5f00\u9500\u53ef\u80fd\u4f1a\u62b5\u6d88\u4f4e\u6bd4\u7279\u91cf\u5316\u5e26\u6765\u7684\u6027\u80fd\u63d0\u5347\uff1b\u5176\u6b21\uff0c\u4ece\u5f00\u53d1\u89d2\u5ea6\u6765\u770b\uff0c\u5f00\u53d1\u8005\u9700\u8981\u9488\u5bf9\u4e0d\u540c\u7684\u6df7\u5408\u7cbe\u5ea6\u91cd\u65b0\u8bbe\u8ba1\u6570\u636e\u5e03\u5c40\u548c\u8ba1\u7b97\u5185\u6838\u3002\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u8ba4\u4e3a\uff0c\u5728\u8bbe\u5907\u4e0a\u90e8\u7f72\u4f4e\u6bd4\u7279\u91cf\u5316\u7684\u5927\u6a21\u578b\uff0c\u5173\u952e\u5728\u4e8e\u5982\u4f55\u57fa\u4e8e\u4f4e\u6bd4\u7279\u7684\u7279\u70b9\u6765\u7a81\u7834\u4f20\u7edf\u77e9\u9635\u4e58\u6cd5\u7684\u5b9e\u73b0\u3002<\/p>\n\n\n\n<p>\u4e3a\u6b64\uff0c\u7814\u7a76\u5458\u4eec\u4ece\u7cfb\u7edf\u548c\u7b97\u6cd5\u5c42\u9762\u63d0\u51fa\u4e86<strong>\u4e00\u79cd\u57fa\u4e8e\u67e5\u627e\u8868\uff08<\/strong><strong>LUT\uff0cLook-Up Table\uff09\u7684\u65b9\u6cd5 T-MAC\uff0c\u5e2e\u52a9\u4f4e\u6bd4\u7279\u91cf\u5316\u7684\u5927\u6a21\u578b\u5728 CPU \u4e0a\u5b9e\u73b0\u9ad8\u6548\u63a8\u7406\u3002<\/strong>T-MAC \u7684\u6838\u5fc3\u601d\u60f3\u5728\u4e8e\u5229\u7528\u6df7\u5408\u7cbe\u5ea6\u77e9\u9635\u4e58\u6cd5\u7684\u4e00\u7aef\u4e3a\u6781\u4f4e\u6bd4\u7279\uff08\u59821\u6bd4\u7279\u62162\u6bd4\u7279\uff09\u7684\u7279\u70b9\u3002\u5b83\u4eec\u7684\u8f93\u51fa\u7ed3\u679c\u53ea\u67092\u76841\u6b21\u65b9\u548c2\u76842\u6b21\u65b9\u79cd\u53ef\u80fd\uff0c\u8fd9\u4e9b\u8f83\u5c11\u7684\u8f93\u51fa\u7ed3\u679c\u5b8c\u5168\u53ef\u4ee5\u63d0\u524d\u8ba1\u7b97\u5e76\u5b58\u50a8\u5728\u8868\u4e2d\uff0c\u5728\u8fd0\u7b97\u65f6\uff0c\u53ea\u9700\u4ece\u8868\u4e2d\u8bfb\u53d6\u7ed3\u679c\uff0c\u907f\u514d\u4e86\u91cd\u590d\u8ba1\u7b97\uff0c\u5927\u5e45\u51cf\u5c11\u4e86\u4e58\u6cd5\u548c\u52a0\u6cd5\u7684\u8fd0\u7b97\u6b21\u6570\u3002<\/p>\n\n\n\n<p>\u5177\u4f53\u800c\u8a00\uff0c<strong>T-MAC \u5c06\u4f20\u7edf\u7684\u4ee5\u6570\u636e\u7c7b\u578b\u4e3a\u4e2d\u5fc3\u7684\u4e58\u6cd5\u8f6c\u53d8\u4e3a\u57fa\u4e8e\u4f4d\u7684\u67e5\u627e\u8868\u64cd\u4f5c\uff0c\u5b9e\u73b0\u4e86\u4e00\u79cd\u7edf\u4e00\u4e14\u53ef\u6269\u5c55\u7684\u6df7\u5408\u7cbe\u5ea6\u77e9\u9635\u4e58\u6cd5\u89e3\u51b3\u65b9\u6848\uff0c\u51cf\u5c0f\u4e86\u8868\u7684\u5927\u5c0f\u5e76\u4f7f\u5176\u505c\u7559\u5728\u6700\u5feb\u7684\u5185\u5b58\u5355\u5143\u4e2d\uff0c\u964d\u4f4e\u4e86\u968f\u673a\u8bbf\u95ee\u8868\u7684\u6210\u672c\u3002<\/strong>\u8fd9\u4e00\u521b\u65b0\u4e3a\u5728\u8d44\u6e90\u53d7\u9650\u7684\u8fb9\u7f18\u8bbe\u5907\u4e0a\u90e8\u7f72\u4f4e\u6bd4\u7279\u91cf\u5316\u5927\u6a21\u578b\u94fa\u5e73\u4e86\u9053\u8def\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"568\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/uploads\/prod\/2024\/08\/low-bit-quantization-2-1024x568.png\" alt=\"T-MAC \u793a\u610f\u56fe\" class=\"wp-image-1078173\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/uploads\/prod\/2024\/08\/low-bit-quantization-2-1024x568.png 1024w, https:\/\/www.microsoft.com\/en-us\/research\/uploads\/prod\/2024\/08\/low-bit-quantization-2-300x166.png 300w, https:\/\/www.microsoft.com\/en-us\/research\/uploads\/prod\/2024\/08\/low-bit-quantization-2-768x426.png 768w, https:\/\/www.microsoft.com\/en-us\/research\/uploads\/prod\/2024\/08\/low-bit-quantization-2-1536x852.png 1536w, https:\/\/www.microsoft.com\/en-us\/research\/uploads\/prod\/2024\/08\/low-bit-quantization-2-240x133.png 240w, https:\/\/www.microsoft.com\/en-us\/research\/uploads\/prod\/2024\/08\/low-bit-quantization-2.png 1650w\" sizes=\"(max-width: 1024px) 100vw, 1024px\" \/><figcaption class=\"wp-element-caption\">\u56fe2\uff1aT-MAC \u793a\u610f\u56fe<\/figcaption><\/figure>\n\n\n\n<p><\/p>\n\n\n\n<p>\u5728\u9488\u5bf9\u4f4e\u6bd4\u7279\u91cf\u5316\u7684 Llama \u548c1\u6bd4\u7279\u7684 BitNet \u5927\u8bed\u8a00\u6a21\u578b\u7684\u6d4b\u8bd5\u4e2d\uff0cT-MAC \u5c55\u73b0\u51fa\u4e86\u663e\u8457\u7684\u6027\u80fd\u4f18\u52bf\u3002\u5728\u642d\u8f7d\u4e86\u6700\u65b0\u9ad8\u901a Snapdragon X Elite \u82af\u7247\u7ec4\u7684 Surface Laptop 7 \u4e0a\uff0cT-MAC \u8ba9 3B BitNet-b1.58 \u6a21\u578b\u7684\u751f\u6210\u901f\u7387\u8fbe\u5230\u6bcf\u79d248\u4e2a token\uff0c2bit 7B Llama \u6a21\u578b\u7684\u751f\u6210\u901f\u7387\u8fbe\u5230\u6bcf\u79d230\u4e2a token\uff0c4bit 7B Llama \u6a21\u578b\u7684\u751f\u6210\u901f\u7387\u53ef\u8fbe\u6bcf\u79d220\u4e2a token\uff0c\u8fd9\u4e9b\u901f\u7387\u5747\u8fdc\u8d85\u4eba\u7c7b\u7684\u5e73\u5747\u9605\u8bfb\u901f\u5ea6\u3002\u4e0e\u539f\u59cb\u7684 Llama.cpp \u6846\u67b6\u76f8\u6bd4\uff0c\u5176\u63d0\u5347\u4e864\u81f35\u500d\uff0c\u751a\u81f3\u6bd4\u4e13\u7528\u7684 NPU \u52a0\u901f\u5668\u8fd8\u5feb\u4e24\u500d\u3002<\/p>\n\n\n\n<p>\u5373\u4f7f\u662f\u5728\u6027\u80fd\u8f83\u4f4e\u7684\u8bbe\u5907\u4e0a\uff0c\u5982 Raspberry Pi\uff08\u6811\u8393\u6d3e\uff095\uff0cT-MAC \u4e5f\u80fd\u4f7f 3B &nbsp; &nbsp;BitNet-b1.58 \u6a21\u578b\u8fbe\u5230\u6bcf\u79d211\u4e2a token \u7684\u751f\u6210\u901f\u7387\u3002T-MAC \u8fd8\u5177\u6709\u663e\u8457\u7684\u529f\u8017\u4f18\u52bf\uff0c\u5728\u8d44\u6e90\u53d7\u9650\u7684\u8bbe\u5907\u4e0a\u53ef\u4ee5\u8fbe\u5230\u76f8\u540c\u7684\u751f\u6210\u901f\u7387\uff0c\u800c\u5b83\u6240\u9700\u7684\u6838\u5fc3\u6570\u4ec5\u4e3a\u539f\u59cb Llama.cpp \u76841\/4\u81f31\/6\u3002<\/p>\n\n\n\n<p>\u8fd9\u4e9b\u7ed3\u679c\u8868\u660e\uff0cT-MAC \u63d0\u4f9b\u4e86\u4e00\u79cd\u5b9e\u7528\u7684\u89e3\u51b3\u65b9\u6848\uff0c\u4f7f\u5f97\u5728\u4f7f\u7528\u901a\u7528 CPU \u7684\u8fb9\u7f18\u8bbe\u5907\u4e0a\u90e8\u7f72\u5927\u8bed\u8a00\u6a21\u578b\u66f4\u4e3a\u9ad8\u6548\uff0c\u4e14\u65e0\u9700\u4f9d\u8d56 GPU\uff0c\u8ba9\u5927\u6a21\u578b\u5728\u8d44\u6e90\u53d7\u9650\u7684\u8bbe\u5907\u4e0a\u4e5f\u80fd\u9ad8\u6548\u8fd0\u884c\uff0c\u4ece\u800c\u63a8\u52a8\u5927\u6a21\u578b\u5728\u66f4\u5e7f\u6cdb\u7684\u573a\u666f\u4e2d\u7684\u5e94\u7528\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"lut-tensor-core-\u63a8\u52a8\u4e0b\u4e00\u4ee3\u786c\u4ef6\u52a0\u901f\u5668\u539f\u751f\u652f\u6301\u6df7\u5408\u7cbe\u5ea6\u77e9\u9635\u4e58\u6cd5\">LUT Tensor Core\uff1a\u63a8\u52a8\u4e0b\u4e00\u4ee3\u786c\u4ef6\u52a0\u901f\u5668\u539f\u751f\u652f\u6301\u6df7\u5408\u7cbe\u5ea6\u77e9\u9635\u4e58\u6cd5<\/h2>\n\n\n\n<p>T-MAC \u548c Ladder \u90fd\u662f\u5728\u73b0\u6709 CPU \u548c GPU \u67b6\u6784\u4e0a\uff0c\u5b9e\u73b0\u5bf9\u6df7\u5408\u7cbe\u5ea6\u77e9\u9635\u4e58\u6cd5\u7684\u4f18\u5316\u652f\u6301\u3002\u5c3d\u7ba1\u8fd9\u4e9b\u8f6f\u4ef6\u5c42\u9762\u7684\u521b\u65b0\u663e\u8457\u63d0\u5347\u4e86\u8ba1\u7b97\u6548\u7387\uff0c\u4f46\u5b83\u4eec\u5728\u6548\u7387\u4e0a\u4ecd\u65e0\u6cd5\u4e0e\u80fd\u591f\u76f4\u63a5\u5b9e\u73b0\u4e00\u4e2a\u4e13\u95e8\u67e5\u627e\u8868\u7684\u786c\u4ef6\u52a0\u901f\u5668\u76f8\u6bd4\u3002\u7814\u7a76\u5458\u4eec\u8ba4\u4e3a\uff0c\u6700\u7406\u60f3\u7684\u65b9\u6cd5\u662f\u91cd\u65b0\u8bbe\u8ba1\u786c\u4ef6\u52a0\u901f\u5668\uff0c\u8ba9 CPU\u3001GPU \u7b49\u80fd\u591f\u539f\u751f\u652f\u6301\u6df7\u5408\u7cbe\u5ea6\u77e9\u9635\u4e58\u6cd5\uff0c\u4f46\u8fd9\u4e00\u76ee\u6807\u9762\u4e34\u4e09\u5927\u6311\u6218\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u6548\u7387\uff1a<\/strong>\u8bbe\u8ba1\u548c\u5b9e\u73b0\u65b9\u5f0f\u5fc5\u987b\u5177\u6709\u6210\u672c\u6548\u76ca\uff0c\u901a\u8fc7\u4f18\u5316\u82af\u7247\u7684\u5229\u7528\u9762\u79ef\uff0c\u6700\u5927\u9650\u5ea6\u5730\u63d0\u9ad8\u4f4e\u6bd4\u7279\u6570\u636e\u7684\u8ba1\u7b97\u6548\u76ca\u3002<\/li>\n\n\n\n<li><strong>\u7075\u6d3b\u6027\uff1a<\/strong>\u7531\u4e8e\u4e0d\u540c\u7684\u6a21\u578b\u548c\u573a\u666f\u9700\u8981\u4e0d\u540c\u7684\u6743\u91cd\u548c\u6fc0\u6d3b\u7cbe\u5ea6\uff0c\u56e0\u6b64\u786c\u4ef6\u4e2d\u7684\u6df7\u5408\u7cbe\u5ea6\u77e9\u9635\u4e58\u6cd5\u8bbe\u8ba1\u5fc5\u987b\u80fd\u591f\u5904\u7406\u5404\u79cd\u6743\u91cd\u7cbe\u5ea6(\u5982 int4\/2\/1)\u548c\u6fc0\u6d3b\u7cbe\u5ea6(\u5982 FP16\/8\u3001int8)\u53ca\u5176\u7ec4\u5408\u3002<\/li>\n\n\n\n<li><strong>\u517c\u5bb9\u6027\uff1a<\/strong>\u65b0\u8bbe\u8ba1\u5fc5\u987b\u4e0e\u73b0\u6709\u7684 GPU \u67b6\u6784\u548c\u8f6f\u4ef6\u751f\u6001\u7cfb\u7edf\u65e0\u7f1d\u96c6\u6210\uff0c\u4ee5\u52a0\u901f\u65b0\u6280\u672f\u7684\u5e94\u7528\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u4e3a\u4e86\u5e94\u5bf9\u8fd9\u4e9b\u6311\u6218\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u8bbe\u8ba1\u4e86&nbsp;<strong>LUT Tensor Core\uff0c\u8fd9\u662f\u4e00\u79cd\u5229\u7528\u67e5\u627e\u8868\u76f4\u63a5\u6267\u884c\u6df7\u5408\u7cbe\u5ea6\u77e9\u9635\u4e58\u6cd5\u7684 GPU Tensor Core \u5fae\u67b6\u6784\u3002<\/strong>\u4e00\u65b9\u9762\uff0c\u57fa\u4e8e\u67e5\u627e\u8868\u7684\u8bbe\u8ba1\u5c06\u4e58\u6cd5\u8fd0\u7b97\u7b80\u5316\u4e3a\u8868\u9884\u8ba1\u7b97\u64cd\u4f5c\uff0c\u53ef\u76f4\u63a5\u5728\u8868\u4e2d\u67e5\u627e\u7ed3\u679c\uff0c\u63d0\u9ad8\u8ba1\u7b97\u6548\u7387\u3002\u53e6\u4e00\u65b9\u9762\uff0c\u8fd9\u79cd\u65b9\u6cd5\u4e5f\u7b80\u5316\u4e86\u5bf9\u786c\u4ef6\u7684\u9700\u6c42\uff0c\u5b83\u53ea\u9700\u7528\u4e8e\u8868\u5b58\u50a8\u7684\u5bc4\u5b58\u5668\u548c\u7528\u4e8e\u67e5\u627e\u7684\u591a\u8def\u9009\u62e9\u5668\uff0c\u65e0\u9700\u4e58\u6cd5\u5668\u548c\u52a0\u6cd5\u5668\u3002\u540c\u65f6\uff0cLUT Tensor Core \u901a\u8fc7\u6bd4\u7279\u4e32\u884c\u8bbe\u8ba1\u5b9e\u73b0\u4e86\u6743\u91cd\u7cbe\u5ea6\u7684\u7075\u6d3b\u6027\uff0c\u5e76\u5229\u7528\u8868\u91cf\u5316\u5b9e\u73b0\u4e86\u6fc0\u6d3b\u7cbe\u5ea6\u7684\u7075\u6d3b\u6027\u3002<\/p>\n\n\n\n<p>\u6b64\u5916\uff0c\u4e3a\u4e86\u4e0e\u73b0\u6709 GPU \u5fae\u67b6\u6784\u548c\u8f6f\u4ef6\u5806\u6808\u96c6\u6210\uff0c\u7814\u7a76\u5458\u4eec\u6269\u5c55\u4e86 GPU \u4e2d\u73b0\u6709\u7684 MMA \u6307\u4ee4\u96c6\uff0c\u52a0\u5165\u4e86\u4e00\u7ec4 LMMA \u6307\u4ee4\uff0c\u5e76\u8bbe\u8ba1\u4e86\u4e00\u4e2a\u7c7b\u4f3c\u4e8e cuBLAS \u7684\u8f6f\u4ef6\u5806\u6808\uff0c\u7528\u4e8e\u96c6\u6210\u5230\u73b0\u6709\u7684 DNN \u6846\u67b6\u4e2d\u3002\u7814\u7a76\u5458\u4eec\u8fd8\u8bbe\u8ba1\u4e86\u4e00\u4e2a\u7f16\u8bd1\u5668\uff0c\u7528\u4e8e\u5728\u5177\u6709 LUT Tensor Core \u7684 GPU \u4e0a\u8fdb\u884c\u7aef\u5230\u7aef\u7684\u6267\u884c\u8ba1\u5212\u3002\u8fd9\u4e9b\u521b\u65b0\u65b9\u6cd5\u53ef\u4ee5\u8ba9 LUT Tensor Core \u88ab\u65e0\u7f1d\u3001\u5feb\u901f\u5730\u91c7\u7528\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"996\" height=\"350\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/uploads\/prod\/2024\/08\/low-bit-quantization-3.png\" alt=\"LUT Tensor Core \u5fae\u67b6\u6784\u6982\u8ff0\" class=\"wp-image-1078176\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/uploads\/prod\/2024\/08\/low-bit-quantization-3.png 996w, https:\/\/www.microsoft.com\/en-us\/research\/uploads\/prod\/2024\/08\/low-bit-quantization-3-300x105.png 300w, https:\/\/www.microsoft.com\/en-us\/research\/uploads\/prod\/2024\/08\/low-bit-quantization-3-768x270.png 768w, https:\/\/www.microsoft.com\/en-us\/research\/uploads\/prod\/2024\/08\/low-bit-quantization-3-240x84.png 240w\" sizes=\"(max-width: 996px) 100vw, 996px\" \/><figcaption class=\"wp-element-caption\">\u56fe3\uff1aLUT Tensor Core \u5fae\u67b6\u6784\u6982\u8ff0<\/figcaption><\/figure>\n\n\n\n<p><\/p>\n\n\n\n<p>\u5728 Llama \u548c BitNet \u6a21\u578b\u4e0a\u7684\u6d4b\u8bd5\u663e\u793a\uff0cLUT Tensor Core \u53ef\u4ee5\u63d0\u4f9b\u9ad8\u8fbe6.93\u500d\u7684\u63a8\u7406\u901f\u5ea6\uff0c\u4e14\u53ea\u5360\u4f20\u7edf Tensor Core \u9762\u79ef\u768438.7%\u3002\u5728\u51e0\u4e4e\u76f8\u540c\u7684\u6a21\u578b\u7cbe\u5ea6\u4e0b\uff0c\u8fd9\u76f8\u5f53\u4e8e20.7\u500d\u7684\u8ba1\u7b97\u5bc6\u5ea6\u548c19.1\u500d\u7684\u80fd\u6548\u63d0\u5347\u3002\u968f\u7740\u4eba\u5de5\u667a\u80fd\u5927\u6a21\u578b\u89c4\u6a21\u548c\u590d\u6742\u6027\u7684\u4e0d\u65ad\u589e\u957f\uff0cLUT Tensor Core \u6709\u52a9\u4e8e\u8fdb\u4e00\u6b65\u91ca\u653e\u4f4e\u6bd4\u7279\u5927\u8bed\u8a00\u6a21\u578b\u7684\u6f5c\u529b\uff0c\u63a8\u52a8\u4eba\u5de5\u667a\u80fd\u5728\u65b0\u573a\u666f\u4e2d\u7684\u5e94\u7528\u3002<\/p>\n\n\n\n<p>\u201c\u67e5\u627e\u8868\u65b9\u6cd5\u5f15\u9886\u4e86\u8ba1\u7b97\u8303\u5f0f\u7684\u8f6c\u53d8\u3002\u5728\u8fc7\u53bb\uff0c\u6211\u4eec\u4f9d\u8d56\u4e8e\u77e9\u9635\u4e58\u6cd5\u548c\u7d2f\u52a0\u8fd0\u7b97\uff0c\u800c\u5728\u5927\u6a21\u578b\u65f6\u4ee3\uff0c\u5f97\u76ca\u4e8e\u4f4e\u6bd4\u7279\u91cf\u5316\u6280\u672f\uff0c\u67e5\u627e\u8868\u65b9\u6cd5\u5c06\u6210\u4e3a\u4e3b\u6d41\u3002\u76f8\u8f83\u4e8e\u4f20\u7edf\u7684\u6d6e\u70b9\u8fd0\u7b97\u6216\u77e9\u9635\u4e58\u6cd5\uff0c\u67e5\u627e\u8868\u65b9\u6cd5\u5728\u8ba1\u7b97\u4e0a\u66f4\u8f7b\u4fbf\u9ad8\u6548\uff0c\u800c\u4e14\u5728\u786c\u4ef6\u5c42\u9762\u4e0a\u66f4\u6613\u4e8e\u6269\u5c55\uff0c\u80fd\u591f\u5b9e\u73b0\u66f4\u9ad8\u7684\u6676\u4f53\u7ba1\u5bc6\u5ea6\uff0c\u5728\u5355\u4f4d\u82af\u7247\u9762\u79ef\u4e0a\u63d0\u4f9b\u66f4\u5927\u7684\u541e\u5410\u91cf\uff0c\u4ece\u800c\u63a8\u52a8\u786c\u4ef6\u67b6\u6784\u7684\u9769\u65b0\u3002\u201d\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u9996\u5e2d\u7814\u7a76\u5458\u66f9\u5a77\u8868\u793a\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"\u4f4e\u6bd4\u7279\u91cf\u5316\u7684\u957f\u5c3e\u6548\u5e94-\u4e3a\u5177\u8eab\u667a\u80fd\u5e26\u6765\u65b0\u53ef\u80fd\">\u4f4e\u6bd4\u7279\u91cf\u5316\u7684\u957f\u5c3e\u6548\u5e94\uff1a\u4e3a\u5177\u8eab\u667a\u80fd\u5e26\u6765\u65b0\u53ef\u80fd<\/h2>\n\n\n\n<p>\u4f4e\u6bd4\u7279\u91cf\u5316\u6280\u672f\u4e0d\u4ec5\u4f18\u5316\u4e86\u5927\u6a21\u578b\u5728\u7aef\u4fa7\u8bbe\u5907\u4e0a\u7684\u8fd0\u884c\u6548\u7387\uff0c\u8fd8\u901a\u8fc7\u51cf\u5c11\u5355\u4e2a\u53c2\u6570\u7684\u201c\u4f53\u79ef\u201d\uff0c\u4e3a\u6a21\u578b\u53c2\u6570\u7684\u6269\u5c55\uff08Scale up\uff09\u63d0\u4f9b\u4e86\u65b0\u7684\u7a7a\u95f4\u3002\u8fd9\u79cd\u53c2\u6570\u6269\u5c55\u80fd\u529b\uff0c\u4f7f\u6a21\u578b\u62e5\u6709\u4e86\u66f4\u5f3a\u7684\u7075\u6d3b\u6027\u548c\u8868\u8fbe\u80fd\u529b\uff0c\u6b63\u5982 BitNet \u6a21\u578b\u6240\u5c55\u793a\u7684\uff0c\u4ece\u4f4e\u6bd4\u7279\u6a21\u578b\u51fa\u53d1\uff0c\u9010\u6b65\u6269\u5c55\u81f3\u66f4\u5927\u89c4\u6a21\u7684\u8bad\u7ec3\u3002<\/p>\n\n\n\n<p>\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684 T-MAC\u3001Ladder \u548c LUT Tensor Core \u7b49\u521b\u65b0\u6280\u672f\uff0c\u4e3a\u5404\u79cd\u4f4e\u6bd4\u7279\u91cf\u5316\u5927\u6a21\u578b\u63d0\u4f9b\u4e86\u9ad8\u6548\u80fd\u7684\u8fd0\u884c\u65b9\u6848\uff0c\u4f7f\u5f97\u8fd9\u4e9b\u6a21\u578b\u80fd\u591f\u5728\u5404\u79cd\u8bbe\u5907\u4e0a\u9ad8\u6548\u8fd0\u884c\uff0c\u5e76\u63a8\u52a8\u79d1\u7814\u4eba\u5458\u4ece\u4f4e\u6bd4\u7279\u89d2\u5ea6\u8bbe\u8ba1\u548c\u4f18\u5316\u5927\u6a21\u578b\u3002\u5176\u4e2d\u90e8\u5206\u6280\u672f\u5df2\u7ecf\u5728\u5fae\u8f6f\u5fc5\u5e94\uff08Bing\uff09\u641c\u7d22\u53ca\u5176\u5e7f\u544a\u4e1a\u52a1\u7b49\u641c\u7d22\u5927\u6a21\u578b\u4e2d\u53d1\u6325\u4f5c\u7528<strong>\u3002\u968f\u7740\u5bf9\u5185\u5b58\u548c\u8ba1\u7b97\u8d44\u6e90\u7684\u964d\u4f4e\uff0c\u4f4e\u6bd4\u7279\u5927\u6a21\u578b\u5728\u673a\u5668\u4eba\u7b49\u5177\u8eab\u667a\u80fd\u7cfb\u7edf\u4e0a\u7684\u90e8\u7f72\u4e5f\u5c06\u6210\u4e3a\u53ef\u80fd\uff0c\u53ef\u4ee5\u4f7f\u8fd9\u4e9b\u8bbe\u5907\u66f4\u597d\u5730\u5b9e\u73b0\u4e0e\u73af\u5883\u7684\u52a8\u6001\u611f\u77e5\u548c\u5b9e\u65f6\u4ea4\u4e92\u3002<\/strong><\/p>\n\n\n\n<p>\u76ee\u524d\uff0cT-MAC \u548c Ladder \u5df2\u7ecf\u5728 GitHub \u4e0a\u5f00\u6e90\uff0c\u6b22\u8fce\u76f8\u5173\u7814\u53d1\u4eba\u5458\u6d4b\u8bd5\u5e94\u7528\uff0c\u4e0e\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u5171\u540c\u63a2\u7d22\u4eba\u5de5\u667a\u80fd\u6280\u672f\u7684\u66f4\u591a\u53ef\u80fd\u3002<\/p>\n\n\n\n<p>\u76f8\u5173\u94fe\u63a5\uff1a<\/p>\n\n\n\n<p>Ladder \u8bba\u6587\u94fe\u63a5\uff1a<\/p>\n\n\n\n<p><a class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" href=\"https:\/\/www.usenix.org\/conference\/osdi24\/presentation\/wang-lei\">https:\/\/www.usenix.org\/conference\/osdi24\/presentation\/wang-lei<span class=\"sr-only\"> (opens in new tab)<\/span><\/a><\/p>\n\n\n\n<p>BitBLAS\/Ladder GitHub \u94fe\u63a5\uff1a<\/p>\n\n\n\n<p><a class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" href=\"https:\/\/github.com\/microsoft\/BitBLAS\">https:\/\/github.com\/microsoft\/BitBLAS<span class=\"sr-only\"> (opens in new tab)<\/span><\/a><\/p>\n\n\n\n<p>T-MAC \u8bba\u6587\u94fe\u63a5\uff1a<\/p>\n\n\n\n<p><a class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" href=\"https:\/\/arxiv.org\/abs\/2407.00088\">https:\/\/arxiv.org\/abs\/2407.00088<span class=\"sr-only\"> (opens in new tab)<\/span><\/a><\/p>\n\n\n\n<p>T-MAC GitHub \u94fe\u63a5\uff1a<\/p>\n\n\n\n<p><a class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" href=\"https:\/\/github.com\/microsoft\/T-MAC\">https:\/\/github.com\/microsoft\/T-MAC<span class=\"sr-only\"> (opens in new tab)<\/span><\/a><\/p>\n\n\n\n<p>LUT Tensor Core \u8bba\u6587\u94fe\u63a5\uff1a<\/p>\n\n\n\n<p><a class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" href=\"https:\/\/arxiv.org\/abs\/2408.06003\">https:\/\/arxiv.org\/abs\/2408.06003<span class=\"sr-only\"> (opens in new tab)<\/span><\/a><\/p>\n\n\n\n<p>BitDistiller \u8bba\u6587\u94fe\u63a5\uff1a<\/p>\n\n\n\n<p><a class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" href=\"https:\/\/arxiv.org\/abs\/2402.10631\">https:\/\/arxiv.org\/abs\/2402.10631<span class=\"sr-only\"> (opens in new tab)<\/span><\/a><\/p>\n\n\n\n<p>BitDistiller GitHub \u94fe\u63a5\uff1a<\/p>\n\n\n\n<p><a class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" href=\"https:\/\/github.com\/DD-DuDa\/BitDistiller\">https:\/\/github.com\/DD-DuDa\/BitDistiller<span class=\"sr-only\"> (opens in new tab)<\/span><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u7f16\u8005\u6309\uff1a\u5728\u4eba\u5de5\u667a\u80fd\u9886\u57df\uff0c\u6a21\u578b\u53c2\u6570\u7684\u589e\u591a\u5f80\u5f80\u610f\u5473\u7740\u6027\u80fd\u7684\u63d0\u5347\u3002\u4f46\u968f\u7740\u6a21\u578b\u89c4\u6a21\u7684\u6269\u5927\uff0c\u5176\u5bf9\u7ec8\u7aef\u8bbe\u5907\u7684\u7b97\u529b\u4e0e\u5185\u5b58\u9700\u6c42\u4e5f\u65e5\u76ca\u589e\u52a0\u3002\u4f4e\u6bd4\u7279\u91cf\u5316\u6280\u672f\uff0c\u7531\u4e8e\u53ef\u4ee5\u5927\u5e45\u964d\u4f4e\u5b58\u50a8\u548c\u8ba1\u7b97\u6210\u672c\u5e76\u63d0\u5347\u63a8\u7406\u6548\u7387\uff0c\u5df2\u6210\u4e3a\u5b9e\u73b0\u5927\u6a21\u578b\u5728\u8d44\u6e90\u53d7\u9650\u8bbe\u5907\u4e0a\u9ad8\u6548\u8fd0\u884c\u7684\u5173\u952e\u6280\u672f\u4e4b\u4e00\u3002\u7136\u800c\uff0c\u5982\u679c\u786c\u4ef6\u8bbe\u5907\u4e0d\u652f\u6301\u4f4e\u6bd4\u7279\u91cf\u5316\u540e\u7684\u6570\u636e\u6a21\u5f0f\uff0c\u90a3\u4e48\u4f4e\u6bd4\u7279\u91cf\u5316\u7684\u4f18\u52bf\u5c06\u65e0\u6cd5\u53d1\u6325\u3002 \u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e00\u95ee\u9898\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u63a8\u51fa\u4e86\u5168\u65b0\u7684\u6570\u636e\u7f16\u8bd1\u5668 Ladder \u548c\u7b97\u6cd5 T-MAC\uff0c\u4f7f\u5f53\u524d\u53ea\u652f\u6301\u5bf9\u79f0\u7cbe\u5ea6\u8ba1\u7b97\u7684\u786c\u4ef6\u80fd\u591f\u76f4\u63a5\u8fd0\u884c\u6df7\u5408\u7cbe\u5ea6\u77e9\u9635\u4e58\u6cd5\u3002\u6d4b\u8bd5\u7ed3\u679c\u8868\u660e\uff0cLadder \u5728\u652f\u6301 GPU \u539f\u672c\u4e0d\u652f\u6301\u7684\u81ea\u5b9a\u4e49\u6570\u636e\u7c7b\u578b\u65b9\u9762\uff0c\u6700\u9ad8\u63d0\u901f\u53ef\u8fbe14.6\u500d\uff1bT-MAC \u5728\u642d\u8f7d\u4e86\u6700\u65b0\u9ad8\u901a Snapdragon X Elite \u82af\u7247\u7ec4\u7684 Surface AI PC \u4e0a\uff0c\u4f7f CPU \u4e0a\u8fd0\u884c\u7684\u5927\u6a21\u578b\u541e\u5410\u7387\u6bd4\u4e13\u7528\u52a0\u901f\u5668 NPU \u5feb\u4e24\u500d\u3002\u6b64\u5916\uff0c\u7814\u7a76\u5458\u4eec\u8fd8\u8bbe\u8ba1\u4e86 LUT Tensor Core \u786c\u4ef6\u67b6\u6784\uff0c\u8fd9\u79cd\u7cbe\u7b80\u8bbe\u8ba1\u4f7f\u786c\u4ef6\u80fd\u591f\u76f4\u63a5\u652f\u6301\u5404\u79cd\u4f4e\u6bd4\u7279\u6df7\u5408\u7cbe\u5ea6\u8ba1\u7b97\uff0c\u4e3a\u4eba\u5de5\u667a\u80fd\u786c\u4ef6\u8bbe\u8ba1\u63d0\u4f9b\u4e86\u65b0\u601d\u8def\u3002 \u5927\u6a21\u578b\u5df2\u7ecf\u8d8a\u6765\u8d8a\u591a\u5730\u88ab\u90e8\u7f72\u5728\u667a\u80fd\u624b\u673a\u3001\u7b14\u8bb0\u672c\u7535\u8111\u3001\u673a\u5668\u4eba\u7b49\u7aef\u4fa7\u8bbe\u5907\u4e0a\uff0c\u4ee5\u63d0\u4f9b\u5148\u8fdb\u7684\u667a\u80fd\u53ca\u5b9e\u65f6\u54cd\u5e94\u670d\u52a1\u3002\u4f46\u5305\u542b\u4e0a\u4ebf\u53c2\u6570\u7684\u5927\u6a21\u578b\u5bf9\u7ec8\u7aef\u8bbe\u5907\u7684\u5185\u5b58\u548c\u8ba1\u7b97\u80fd\u529b\u63d0\u51fa\u4e86\u6781\u9ad8\u7684\u8981\u6c42\uff0c\u4e5f\u56e0\u6b64\u9650\u5236\u4e86\u5b83\u4eec\u7684\u5e7f\u6cdb\u5e94\u7528\u3002\u4f4e\u6bd4\u7279\u91cf\u5316\u6280\u672f\u56e0\u5176\u80fd\u663e\u8457\u538b\u7f29\u6a21\u578b\u89c4\u6a21\uff0c\u964d\u4f4e\u5bf9\u8ba1\u7b97\u8d44\u6e90\u7684\u9700\u6c42\uff0c\u6210\u4e3a\u4e86\u5927\u6a21\u578b\u5728\u7aef\u4fa7\u90e8\u7f72\u548c\u5b9e\u73b0\u9ad8\u6548\u63a8\u7406\u7684\u6709\u6548\u624b\u6bb5\u3002 \u968f\u7740\u4f4e\u6bd4\u7279\u91cf\u5316\u6280\u672f\u7684\u53d1\u5c55\uff0c\u6570\u636e\u7c7b\u578b\u65e5\u76ca\u591a\u6837\u5316\uff0c\u5982 int4\u3001int2\u3001int1 \u7b49\u4f4e\u6bd4\u7279\u6570\u636e\uff0c\u4f7f\u5f97\u5927\u6a21\u578b\u5728\u63a8\u7406\u4e2d\u8d8a\u6765\u8d8a\u591a\u5730\u91c7\u7528\u4f4e\u6bd4\u7279\u6743\u91cd\u548c\u9ad8\u6bd4\u7279\u6743\u91cd\u8ba1\u7b97\u7684\u6df7\u5408\u7cbe\u5ea6\u77e9\u9635\u4e58\u6cd5\uff08mixed-precision matrix multiplication\uff0cmpGEMM\uff09\u3002\u7136\u800c\uff0c\u73b0\u6709\u7684 CPU\u3001GPU \u7b49\u786c\u4ef6\u8ba1\u7b97\u5355\u5143\u901a\u5e38\u53ea\u652f\u6301\u5bf9\u79f0\u8ba1\u7b97\u6a21\u5f0f\uff0c\u5e76\u4e0d\u517c\u5bb9\u8fd9\u79cd\u6df7\u5408\u7cbe\u5ea6\u7684\u77e9\u9635\u4e58\u6cd5\u3002 \u6df7\u5408\u7cbe\u5ea6\u77e9\u9635\u4e58\u6cd5\u4e0e\u4f20\u7edf\u7684\u77e9\u9635\u4e58\u6cd5\u6709\u4f55\u4e0d\u540c\uff1f \u5728\u4f20\u7edf\u7684\u77e9\u9635\u4e58\u6cd5\u4e2d\uff0c\u53c2\u4e0e\u8fd0\u7b97\u7684\u4e24\u7aef\u6570\u503c\u662f\u5bf9\u79f0\u7684\uff0c\u4f8b\u5982 FP16*FP16\u3001int8*int8\u3002\u4f46\u5927\u6a21\u578b\u7684\u4f4e\u6bd4\u7279\u91cf\u5316\u6253\u7834\u4e86\u8fd9\u79cd\u5bf9\u79f0\u6027\uff0c\u4f7f\u4e58\u6cd5\u7684\u4e00\u7aef\u662f\u9ad8\u6bd4\u7279\uff0c\u53e6\u4e00\u7aef\u662f\u4f4e\u6bd4\u7279\uff0c\u4f8b\u5982\u5728 1-bit \u7684&nbsp;BitNet (opens in new tab)&nbsp;\u6a21\u578b\u4e2d\u5b9e\u73b0\u7684 int8*int1 \u6216 int8*int2\uff0c\u4ee5\u53ca\u6d6e\u70b9\u6570\u4e0e\u6574\u6570\u7684\u6df7\u5408\u4e58\u6cd5 FP16*int4\u3002 \u4e3a\u4e86\u5145\u5206\u53d1\u6325\u4f4e\u6bd4\u7279\u91cf\u5316\u7684\u4f18\u52bf\uff0c\u8ba9\u786c\u4ef6\u8bbe\u5907\u80fd\u591f\u76f4\u63a5\u652f\u6301\u6df7\u5408\u7cbe\u5ea6\u77e9\u9635\u4e58\u6cd5\uff0c\u786e\u4fdd\u5927\u6a21\u578b\u5728\u7aef\u4fa7\u8bbe\u5907\u4e0a\u7684\u9ad8\u901f\u6709\u6548\u8fd0\u884c\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u9488\u5bf9\u73b0\u6709 CPU\u3001GPU \u8ba1\u7b97\u7b97\u5b50\u548c\u786c\u4ef6\u67b6\u6784\u8fdb\u884c\u521b\u65b0\uff1a \u5f53\u524d\uff0c\u524d\u6cbf\u52a0\u901f\u5668\u6b63\u5728\u5c06\u66f4\u4f4e\u6bd4\u7279\u7684\u8ba1\u7b97\u5355\u5143\uff0c\u5982 FP32\u3001FP16\uff0c\u751a\u81f3 FP8 \u7684\u8fd0\u7b97\u96c6\u6210\u5230\u65b0\u4e00\u4ee3\u7684\u67b6\u6784\u4e2d\u3002\u7136\u800c\uff0c\u53d7\u9650\u4e8e\u82af\u7247\u9762\u79ef\u548c\u9ad8\u6602\u7684\u786c\u4ef6\u6210\u672c\uff0c\u6bcf\u4e2a\u52a0\u901f\u5668\u53ea\u80fd\u4e3a\u6807\u51c6\u7684\u6570\u636e\u7c7b\u578b\u63d0\u4f9b\u6709\u9650\u7c7b\u578b\u7684\u8ba1\u7b97\u5355\u5143\uff0c\u6bd4\u5982 NVIDIA V100 [&hellip;]<\/p>\n","protected":false},"author":34512,"featured_media":1078179,"template":"","meta":{"msr-url-field":"","msr-podcast-episode":"","msrModifiedDate":"","msrModifiedDateEnabled":false,"ep_exclude_from_search":false,"_classifai_error":"","msr-content-parent":1012650,"footnotes":""},"research-area":[],"msr-locale":[268881],"msr-post-option":[],"class_list":["post-1078164","msr-blog-post","type-msr-blog-post","status-publish","has-post-thumbnail","hentry","msr-locale-zh_cn"],"msr_assoc_parent":{"id":1012650,"type":"lab"},"_links":{"self":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1078164"}],"collection":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post"}],"about":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/types\/msr-blog-post"}],"author":[{"embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/users\/34512"}],"version-history":[{"count":3,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1078164\/revisions"}],"predecessor-version":[{"id":1078185,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1078164\/revisions\/1078185"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/media\/1078179"}],"wp:attachment":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/media?parent=1078164"}],"wp:term":[{"taxonomy":"msr-research-area","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/research-area?post=1078164"},{"taxonomy":"msr-locale","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-locale?post=1078164"},{"taxonomy":"msr-post-option","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-post-option?post=1078164"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}