{"id":1073484,"date":"2024-07-19T00:40:00","date_gmt":"2024-07-19T07:40:00","guid":{"rendered":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/?post_type=msr-blog-post&#038;p=1073484"},"modified":"2024-09-25T04:23:01","modified_gmt":"2024-09-25T11:23:01","slug":"q-sparse","status":"publish","type":"msr-blog-post","link":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/articles\/q-sparse\/","title":{"rendered":"\u5b8c\u5168\u6fc0\u6d3b\u7a00\u758f\u5927\u6a21\u578b\uff0cQ-Sparse\u7a81\u7834LLM\u63a8\u7406\u6548\u80fd"},"content":{"rendered":"\n<p><em>\u4f5c\u8005\uff1a\u91cf\u5b50\u4f4d<\/em><\/p>\n\n\n\n<p>\u7f16\u8005\u6309\uff1a\u6fc0\u6d3b\u7a00\u758f\u6027\u662f\u89e3\u51b3\u5927\u8bed\u8a00\u6a21\u578b\uff08LLMs\uff09\u5728\u63a8\u7406\u9636\u6bb5\u51fa\u73b0\u7684\u8ba1\u7b97\u6210\u672c\u9ad8\u3001\u5185\u5b58\u5360\u7528\u5927\u7b49\u95ee\u9898\u7684\u6709\u6548\u65b9\u6cd5\uff0c\u53ef\u4ee5\u6709\u6548\u51cf\u5c11\u6fc0\u6d3b\u5f20\u91cf\u4e2d\u6fc0\u6d3b\u5143\u7d20\u7684\u6570\u91cf\u3002\u7136\u800c\u8be5\u65b9\u6cd5\u65e0\u6cd5\u5b9e\u73b0 LLMs \u6fc0\u6d3b\u7684\u5b8c\u5168\u7a00\u758f\u6027\uff0c\u4ece\u800c\u9650\u5236\u4e86\u63a8\u7406\u9636\u6bb5\u7684\u6548\u7387\u63d0\u5347\u3002<\/p>\n\n\n\n<p>\u5bf9\u6b64\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u63d0\u51fa Q-Sparse \u5b9e\u73b0\u4e86 LLMs \u6fc0\u6d3b\u7684\u5b8c\u5168\u7a00\u758f\u6027\u3002\u8be5\u7814\u7a76\u4e0d\u4ec5\u63ed\u793a\u4e86\u5305\u62ec\u63a8\u7406\u4f18\u5316\u89c4\u6a21\u6cd5\u5219\uff08inference-optimal scaling law\uff09\u5728\u5185\u7684\u591a\u9879\u65b0\u53d1\u73b0\uff0c\u8fd8\u5c55\u73b0\u51fa\u5e7f\u6cdb\u7684\u517c\u5bb9\u6027\uff0c\u5176\u4e0e BitNet \u6b63\u4ea4\u4e14\u4e92\u8865\uff0c\u4e3a LLMs \u63a8\u7406\u4e2d\u7684\u6570\u636e\u7c7b\u578b\u63d0\u4f9b\u4e86\u5168\u9762\u4f18\u5316\u3002\u6b64\u5916\uff0cQ-Sparse \u5f3a\u5927\u7684\u9002\u7528\u6027\uff0c\u5bf9\u4ece\u5934\u5f00\u59cb\u8bad\u7ec3\u3001\u6301\u7eed\u8bad\u7ec3\u4ee5\u53ca\u5fae\u8c03\u90fd\u53ef\u4ee5\u63d0\u4f9b\u6709\u6548\u7684\u652f\u6301\u3002<\/p>\n\n\n\n<p>\u672c\u6587\u8f6c\u8f7d\u81ea\u516c\u4f17\u53f7\u201c\u91cf\u5b50\u4f4d\u201d\uff0c\u539f\u6587\u6807\u9898\u300a\u53ea\u6fc0\u6d3b3.8B\u53c2\u6570\uff0c\u6027\u80fd\u6bd4\u80a9\u540c\u6b3e7B\u6a21\u578b\uff01\u8bad\u7ec3\u5fae\u8c03\u90fd\u80fd\u7528\uff0c\u6765\u81ea\u5fae\u8f6f\u300b\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<p>\u53ea\u9700\u6fc0\u6d3b60%\u7684\u53c2\u6570\uff0c\u5c31\u80fd\u5b9e\u73b0\u4e0e\u5168\u6fc0\u6d3b\u7a20\u5bc6\u6a21\u578b\u76f8\u5f53\u7684\u6027\u80fd\u3002<\/p>\n\n\n\n<p>\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u4e00\u9879\u65b0\u7814\u7a76\uff0c\u5b9e\u73b0\u4e86\u6a21\u578b\u7684\u5b8c\u5168\u7a00\u758f\u6fc0\u6d3b\uff0c\u8ba9\u63a8\u7406\u6210\u672c\u5927\u5e45\u4e0b\u964d\u3002<\/p>\n\n\n\n<p>\u800c\u4e14\u9002\u7528\u8303\u56f4\u5e7f\u6cdb\uff0c\u65e0\u8bba\u662f\u4ece\u5934\u8bad\u7ec3\u3001\u7ee7\u7eed\u8bad\u7ec3\u8fd8\u662f\u5fae\u8c03\uff0c\u90fd\u80fd\u63d0\u4f9b\u6709\u6548\u652f\u6301\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"836\" height=\"271\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-1.png\" alt=\"Q-Sparse: All Large Language Models can be Fully Sparsely-Activated\" class=\"wp-image-1073493\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-1.png 836w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-1-300x97.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-1-768x249.png 768w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-1-240x78.png 240w\" sizes=\"auto, (max-width: 836px) 100vw, 836px\" \/><\/figure>\n\n\n\n<p>\u8be5\u65b9\u6cd5\u540d\u4e3a Q-Sparse\uff0c\u5728\u795e\u7ecf\u5143\u7ea7\u522b\u4e0a\u5b9e\u73b0\u4e86\u6a21\u578b\u7a00\u758f\u5316\uff0c\u76f8\u6bd4\u4e8e\u5176\u4ed6\u65b9\u5f0f\u7c92\u5ea6\u66f4\u7ec6\uff0c\u5728\u76f8\u540c\u63a8\u7406\u5f00\u9500\u4e0b\uff0c\u65e0\u8bba\u6027\u80fd\u8fd8\u662f\u7a00\u758f\u7387\u90fd\u66f4\u597d\u3002<\/p>\n\n\n\n<p>\u540d\u79f0\u4e4b\u4e2d\uff0cQ \u6307\u7684\u662f\u91cf\u5316\uff08Quantization\uff09\uff0c\u610f\u5473\u7740\u5b83\u9664\u4e86\u666e\u901a\u6a21\u578b\u4e4b\u5916\uff0c\u4e5f\u517c\u5bb9\u91cf\u5316\u6280\u672f\uff0c\u9002\u7528\u4e8e\u5404\u79cd\u91cf\u5316\u65b9\u5f0f\u7684\u6a21\u578b\u3002\u5982\u679c\u628a Q-Sparse \u4e0e\u6a21\u578b\u91cf\u5316\u6280\u672f\u7ed3\u5408\uff0c\u8fd8\u53ef\u4ee5\u5b9e\u73b0\u66f4\u5927\u7a0b\u5ea6\u7684\u964d\u672c\u589e\u6548\u3002<\/p>\n\n\n\n<p>\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u5168\u7403\u7814\u7a76\u5408\u4f19\u4eba\u97e6\u798f\u5982\u8868\u793a\uff0cQ-Sparse \u5c55\u73b0\u51fa\u4e86\u5e7f\u6cdb\u7684\u517c\u5bb9\u6027\uff0c\u65e2\u9002\u7528\u4e8eFP16\/BF16\u7c7b\u578b\u7684LLMs\uff0c\u4e5f\u652f\u63011-bit LLMs\uff0c\u4f8b\u5982 BitNet\u3002\u503c\u5f97\u4e00\u63d0\u7684\u662f\uff0cQ-Sparse \u4e0e\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u6b64\u524d\u53d1\u5e03\u7684 BitNet \u6280\u672f\u6b63\u4ea4\u4e14\u4e92\u8865\uff0c\u4e3a LLMs \u63a8\u7406\u4e2d\u7684\u6570\u636e\u7c7b\u578b\u63d0\u4f9b\u4e86\u5168\u9762\u4f18\u5316\u2014\u2014BitNet \u4e13\u6ce8\u4e8e\u6a21\u578b\u6743\u91cd\u7684\u4f4e\u6bd4\u7279\uff08\u53731(.58)bit\uff09\uff0c\u800c Q-Sparse \u4e13\u6ce8\u4e8e\u7f51\u7edc\u6fc0\u6d3b\u7684\u7a00\u758f\u5316\u3002BitNet b1.58 \u548c Q-Sparse\uff08\u53ef\u914d\u5907 MoE\uff09\u7684\u7ed3\u5408\uff0c\u4e3a\u672a\u6765 LLMs \u5728\u63d0\u5347\u6548\u7387\u3001\u964d\u4f4e\u6210\u672c\u548c\u80fd\u8017\u65b9\u9762\u63d0\u4f9b\u4e86\u6e05\u6670\u7684\u8def\u5f84\u3002<\/p>\n\n\n\n<p>\u53e6\u5916\u5728\u7814\u7a76 Q-Sparse \u7684\u540c\u65f6\uff0c\u56e2\u961f\u4e5f\u5bf9\u53c2\u6570\u89c4\u6a21\u3001\u7a00\u758f\u7387\u548c\u6a21\u578b\u6027\u80fd\u4e09\u8005\u4e4b\u95f4\u7684\u5173\u7cfb\u8fdb\u884c\u4e86\u6df1\u5165\u63a2\u5bfb\uff0c\u5e76\u53d1\u73b0\u4e86\u9002\u7528\u4e8e\u6a21\u578b\u63a8\u7406\u4f18\u5316\u7684\u201cScaling Law\u201d\u3002<\/p>\n\n\n\n<p>\u6709\u7f51\u53cb\u8ba4\u4e3a\uff0c\u8fd9\u9879\u6280\u672f\u786e\u5b9e\u4e0d\u9519\uff0c\u800c\u4e14\u6bd4 ReLU \u8981\u66f4\u597d\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"582\" height=\"128\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-2.png\" alt=\"Q-Sparse - feedback post \"Achieving better results on finetuned pre-trained models [three fire icons]. And better than ReLU! Great work.\"\" class=\"wp-image-1073496\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-2.png 582w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-2-300x66.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-2-240x53.png 240w\" sizes=\"auto, (max-width: 582px) 100vw, 582px\" \/><\/figure>\n\n\n\n<p>\u8fd8\u6709\u4eba\u5f00\u542f\u4e86\u8bb8\u613f\u6a21\u5f0f\uff0c\u8868\u793a\u5982\u679c\uff08AMD\u7684\uff09ROCm \u80fd\u6bd4\u82f1\u4f1f\u8fbe\u66f4\u5feb\u652f\u6301\u8fd9\u9879\u6280\u672f\u5c31\u597d\u4e86\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"587\" height=\"88\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-3.png\" alt=\"Q-Sparse - feedback post \"I hope rocm supports this sooner than nvidia\"\" class=\"wp-image-1073499\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-3.png 587w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-3-300x45.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-3-240x36.png 240w\" sizes=\"auto, (max-width: 587px) 100vw, 587px\" \/><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"\u7528top-k\u51fd\u6570\u5b9e\u73b0\u7a00\u758f\u5316\">\u7528Top-K\u51fd\u6570\u5b9e\u73b0\u7a00\u758f\u5316<\/h2>\n\n\n\n<p>Q-Sparse \u6240\u505a\u7684\u6700\u6838\u5fc3\u7684\u64cd\u4f5c\uff0c\u662f\u5bf9\u8f93\u5165\u7684\u5f20\u91cf\u5e94\u7528 Top-K \u7a00\u758f\u5316\u51fd\u6570\u3002<\/p>\n\n\n\n<p>\u5177\u4f53\u6765\u8bf4\uff0cTransformer \u67b6\u6784\u5728\u6ce8\u610f\u529b\u5c42\u548c\u524d\u9988\u5c42\u4e2d\u90fd\u4f7f\u7528 nn.Linear \u7ebf\u6027\u5c42\uff08\u77e9\u9635\u4e58\u6cd5\uff09\u8fdb\u884c\u6295\u5f71\uff0c\u53ef\u4ee5\u8868\u793a\u4e3a Y=X\u00b7W^T\u3002\uff08\u5176\u4e2d X \u5c31\u662f\u8f93\u5165\u5f20\u91cf\uff0cW \u4ee3\u8868\u5176\u6743\u91cd\uff0cY \u4e3a\u8f93\u51fa\u5f20\u91cf\uff09<\/p>\n\n\n\n<p>Q-Sparse \u4e2d\uff0c\u5bf9\u4e8e\u4e00\u4e2a\u8f93\u5165\u6fc0\u6d3b\u5f20\u91cf X\uff0c\u9996\u5148\u4f1a\u8ba1\u7b97\u5176\u7edd\u5bf9\u503c |X| \u5e76\u8fdb\u884c\u6392\u5e8f\uff0c\u627e\u51fa\u5176\u4e2d\u7edd\u5bf9\u503c\u6700\u5927\u7684 K \u4e2a\u5143\u7d20\u3002<\/p>\n\n\n\n<p>\u8fd9\u91cc\u7684 K \u662f\u9884\u5148\u8bbe\u5b9a\u7684\u8d85\u53c2\u6570\uff0c\u51b3\u5b9a\u4e86\u7a00\u758f\u5316\u7684\u7a0b\u5ea6\u3002<\/p>\n\n\n\n<p>\u4e4b\u540e Q-Sparse \u4f1a\u521b\u5efa\u4e00\u4e2a\u4e0e X \u5f62\u72b6\u76f8\u540c\u7684\u4e8c\u8fdb\u5236\u63a9\u7801\u5f20\u91cf M\uff0c\u5bf9\u4e8e\u4e00\u7cfb\u5217 |X| \u4e2d\u7edd\u5bf9\u503c\u6700\u5927\u7684 K \u4e2a\u5143\u7d20\u5bf9\u5e94\u7684\u4f4d\u7f6e\uff0c\u5c06 M \u4e2d\u7684\u76f8\u5e94\u4f4d\u7f6e\u8bbe\u7f6e\u4e3a1\uff0c\u5176\u4f59\u4f4d\u7f6e\u8bbe\u7f6e\u4e3a0\u3002<\/p>\n\n\n\n<p>\u63a5\u7740\uff0c\u5c06\u8f93\u5165\u5f20\u91cf X \u4e0e\u63a9\u7801\u5f20\u91cf M \u8fdb\u884c Hadamard \u79ef\uff08\u9010\u5143\u7d20\u76f8\u4e58\uff09\u8fd0\u7b97\uff0c\u5c31\u5f97\u5230\u4e86\u7a00\u758f\u5316\u7684\u5f20\u91cf X_sparse\u3002<\/p>\n\n\n\n<p>\u5728\u524d\u5411\u4f20\u64ad\u8fc7\u7a0b\u4e2d\uff0c\u7a00\u758f\u5316\u540e\u7684\u5f20\u91cf X_sparse \u5c06\u4ee3\u66ff\u539f\u59cb\u7684\u8f93\u5165\u5f20\u91cf X \u53c2\u4e0e\u540e\u7eed\u7684\u8ba1\u7b97\uff08\u5982\u77e9\u9635\u4e58\u6cd5\uff09\u3002<\/p>\n\n\n\n<p>\u7531\u4e8e X_sparse \u4e2d\u5927\u90e8\u5206\u5143\u7d20\u5df2\u7ecf\u88ab\u8bbe\u7f6e\u4e3a\u96f6\uff0c\u56e0\u6b64\u53ef\u4ee5\u663e\u8457\u51cf\u5c11\u8ba1\u7b97\u91cf\u548c\u5185\u5b58\u5e26\u5bbd\u9700\u6c42\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"639\" height=\"357\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-4.png\" alt=\"Q-Sparse - diagram\" class=\"wp-image-1073502\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-4.png 639w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-4-300x168.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-4-240x134.png 240w\" sizes=\"auto, (max-width: 639px) 100vw, 639px\" \/><\/figure>\n\n\n\n<p>\u5728\u53cd\u5411\u4f20\u64ad\u8fc7\u7a0b\u4e2d\uff0cQ-Sparse\u4f7f\u7528\u4e86\u76f4\u901a\u4f30\u8ba1\u5668\uff08Straight-Through Estimator\uff0cSTE\uff09\u6765\u8ba1\u7b97 Top-K \u51fd\u6570\u7684\u68af\u5ea6\u3002<\/p>\n\n\n\n<p>\u4f20\u7edf\u7684\u8bad\u7ec3\u65b9\u5f0f\u4e2d\uff0c\u901a\u5e38\u9700\u8981\u8ba1\u7b97\u635f\u5931\u51fd\u6570\u5bf9\u7f51\u7edc\u53c2\u6570\u7684\u68af\u5ea6\uff0c\u5e76\u4f7f\u7528\u68af\u5ea6\u4e0b\u964d\u6cd5\u66f4\u65b0\u53c2\u6570\u4ee5\u6700\u5c0f\u5316\u635f\u5931\u3002<\/p>\n\n\n\n<p>\u4f46\u5f53\u7f51\u7edc\u4e2d\u5b58\u5728 Top-K\u3001\u7f6e\u96f6\u7b49\u4e00\u4e9b\u7a00\u758f\u5316\u7684\u64cd\u4f5c\u65f6\uff0c\u68af\u5ea6\u7684\u8ba1\u7b97\u5c31\u4f1a\u9047\u5230\u95ee\u9898\uff0c\u56e0\u4e3a\u8fd9\u4e9b\u64cd\u4f5c\u7684\u8f93\u51fa\u5bf9\u8f93\u5165\u7684\u68af\u5ea6\u5728\u5927\u591a\u6570\u70b9\u4e0a\u90fd\u662f0\uff0c\u5bfc\u81f4\u68af\u5ea6\u65e0\u6cd5\u6709\u6548\u4f20\u64ad\u3002<\/p>\n\n\n\n<p>STE \u901a\u8fc7\u76f4\u63a5\u5c06\u68af\u5ea6\u4f20\u9012\u7ed9\u7a00\u758f\u5316\u4e4b\u524d\u7684\u5f20\u91cf\uff0c\u907f\u514d\u4e86\u68af\u5ea6\u6d88\u5931\u7684\u95ee\u9898\u3002<\/p>\n\n\n\n<p>\u4e00\u822c\u7684\u53cd\u5411\u4f20\u64ad\u4e2d\uff0c\u635f\u5931\u51fd\u6570 L \u5bf9 x \u7684\u68af\u5ea6 \u2202L\/\u2202x=\u2202L\/\u2202y\u22c5\u2202y\/\u2202x\uff0c\u4f46\u7531\u4e8e\u4e0d\u53ef\u5fae\u5206\u65e0\u6cd5\u76f4\u63a5\u8ba1\u7b97\u3002<\/p>\n\n\n\n<p>STE \u7684\u89e3\u51b3\u65b9\u6848\u662f\u53ea\u8ba1\u7b97\u635f\u5931\u51fd\u6570\u5bf9\u7a00\u758f\u5316\u5f20\u91cfy\u7684\u68af\u5ea6\uff0c\u7136\u540e\u5c06\u5176\u76f4\u63a5\u590d\u5236\u7ed9\u539f\u59cb\u5f20\u91cf x\uff0c\u4e5f\u5c31\u662f\u76f4\u63a5\u5c06 \u2202L\/\u2202y \u4f5c\u4e3a \u2202L\/\u2202x \u7684\u4f30\u8ba1\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"742\" height=\"421\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-5.png\" alt=\"Q-Sparse - bar chart with Dense baseline, Q-Sparse (2\/ STE) and Q-Sparse (w\/o STE)\" class=\"wp-image-1073505\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-5.png 742w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-5-300x170.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-5-240x136.png 240w\" sizes=\"auto, (max-width: 742px) 100vw, 742px\" \/><figcaption class=\"wp-element-caption\"><em>\u6709\/\u65e0 STE \u65f6\u7684\u68af\u5ea6\u6bd4\u8f83<\/em><\/figcaption><\/figure>\n\n\n\n<p>\u5bf9\u4e8e\u524d\u9988\u5c42\uff0cQ-Sparse \u4f7f\u7528\u5e73\u65b9 ReLU \u51fd\u6570\u4ee3\u66ff\u5e38\u89c4\u7684 ReLU \u6fc0\u6d3b\u51fd\u6570\uff0c\u5e73\u65b9\u8fd0\u7b97\u53ef\u4ee5\u8fdb\u4e00\u6b65\u63d0\u9ad8\u6fc0\u6d3b\u7684\u7a00\u758f\u6027\uff08\u2299 \u8868\u793a Hadamard \u79ef\uff09\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"520\" height=\"59\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-6.png\" alt=\"Q-Sparse - formula\" class=\"wp-image-1073508\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-6.png 520w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-6-300x34.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-6-240x27.png 240w\" sizes=\"auto, (max-width: 520px) 100vw, 520px\" \/><\/figure>\n\n\n\n<p>\u53e6\u5916\uff0c\u4e3a\u4e86\u9002\u914d\u91cf\u5316\u6a21\u578b\uff0cQ-Sparse \u5728\u5e94\u7528 Top-K \u7a00\u758f\u5316\u4e4b\u524d\uff0c\u4f1a\u5148\u5bf9\u8f93\u5165\u5f20\u91cf\u8fdb\u884c\u91cf\u5316\uff0c\u4ee5\u786e\u4fdd\u7a00\u758f\u5316\u64cd\u4f5c\u4e0e\u91cf\u5316\u8868\u793a\u517c\u5bb9\uff0c\u5176\u51fd\u6570\u8868\u793a\u5982\u4e0b\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"608\" height=\"246\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-7.png\" alt=\"Q-Sparse - formula\" class=\"wp-image-1073511\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-7.png 608w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-7-300x121.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-7-240x97.png 240w\" sizes=\"auto, (max-width: 608px) 100vw, 608px\" \/><\/figure>\n\n\n\n<p>\u5176\u4e2d\uff0c\u03b5 \u662f\u4e00\u4e2a\u5c0f\u5e38\u6570\uff0c\u7528\u4e8e\u907f\u514d\u51fa\u73b0\u5206\u6bcd\u4e3a\u96f6\u7684\u60c5\u51b5\u3002<\/p>\n\n\n\n<p>\u7279\u522b\u7684\uff0c\u5bf9\u4e8e 1-bit \u91cf\u5316\u7684\u6743\u91cd\uff0cQ-Sparse \u4f7f\u7528\u4ee5\u4e0b\u91cf\u5316\u51fd\u6570\uff0c\u5176\u4e2d \u03b1 \u662f\u6743\u91cd\u5f20\u91cf W \u7684\u5e73\u5747\u7edd\u5bf9\u503c\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"551\" height=\"87\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-8.png\" alt=\"Q-Sparse - formula\" class=\"wp-image-1073514\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-8.png 551w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-8-300x47.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-8-240x38.png 240w\" sizes=\"auto, (max-width: 551px) 100vw, 551px\" \/><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"60-\u6fc0\u6d3b\u53c2\u6570\u8fbe\u5230\u76f8\u540c\u6548\u679c\">60%\u6fc0\u6d3b\u53c2\u6570\u8fbe\u5230\u76f8\u540c\u6548\u679c<\/h2>\n\n\n\n<p>\u5bf9\u6bd4\u5b9e\u9a8c\u8868\u660e\uff0c\u65e0\u8bba\u662f\u7a00\u758f\u7387\u8fd8\u662f\u6a21\u578b\u8868\u73b0\uff0cQ-Sparse \u90fd\u663e\u8457\u4f18\u4e8e\u6b64\u524d\u7684 ReLU \u65b9\u6cd5\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"807\" height=\"313\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-9.png\" alt=\"Q-Sparse - Sparsity Ratio line charts\" class=\"wp-image-1073556\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-9.png 807w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-9-300x116.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-9-768x298.png 768w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-9-240x93.png 240w\" sizes=\"auto, (max-width: 807px) 100vw, 807px\" \/><\/figure>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"807\" height=\"312\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-10.png\" alt=\"Q-Sparse - PPL and Sparsity Ratio charts\" class=\"wp-image-1073517\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-10.png 807w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-10-300x116.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-10-768x297.png 768w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-10-240x93.png 240w\" sizes=\"auto, (max-width: 807px) 100vw, 807px\" \/><\/figure>\n\n\n\n<p>\u9488\u5bf9 Q-Sparse \u7684\u5177\u4f53\u6548\u679c\uff0c\u4f5c\u8005\u5bf9\u5176\u5728\u4ece\u5934\u8bad\u7ec3\u3001\u7ee7\u7eed\u8bad\u7ec3\u548c\u5fae\u8c03\u4e09\u9879\u4efb\u52a1\u4e0a\u7684\u6027\u80fd\u8fdb\u884c\u4e86\u8bc4\u4f30\u3002<\/p>\n\n\n\n<p>\u4ece\u5934\u8bad\u7ec3\u5b9e\u9a8c\u4f7f\u7528\u7684\u6a21\u578b\u4e3a Llama\uff0c\u7ed3\u679c\u5728 700M \u548c 7B \u6a21\u578b\u4e0a\uff0c\u4f7f\u752870% top-K(\u537340%\u7684\u6574\u4f53\u7a00\u758f\u7387)\u7684 Q-Sparse \u53ef\u4ee5\u8fbe\u5230\u4e0e\u5bc6\u96c6 baseline \u76f8\u5f53\u7684\u8bad\u7ec3\u635f\u5931\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"797\" height=\"351\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-14.png\" alt=\"Q-Sparse - Loss line charts by model size\" class=\"wp-image-1073529\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-14.png 797w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-14-300x132.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-14-768x338.png 768w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-14-240x106.png 240w\" sizes=\"auto, (max-width: 797px) 100vw, 797px\" \/><\/figure>\n\n\n\n<p>\u7ee7\u7eed\u8bad\u7ec3\u7684\u76ee\u7684\u662f\u5c06\u7a20\u5bc6\u6a21\u578b\u7a00\u758f\u5316\uff0c\u8fd9\u91cc\u7684\u5b9e\u9a8c\u5bf9\u8c61\u662f Mistral-7B\u3002<\/p>\n\n\n\n<p>\u7ed3\u679c\uff0c\u5728\u6fc0\u6d3b\u53c2\u6570\u4e3a 2.9B \u548c 3.8B \u7684\u60c5\u51b5\u4e0b\uff0c\u6a21\u578b\u5728 ARC\u3001MMLU \u7b49\u6570\u636e\u96c6\u4e2d\u7684\u5f97\u5206\u5747\u672a\u53d1\u751f\u660e\u663e\u4e0b\u964d\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"853\" height=\"201\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-12.png\" alt=\"Q-Sparse - Models table\" class=\"wp-image-1073523\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-12.png 853w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-12-300x71.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-12-768x181.png 768w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-12-240x57.png 240w\" sizes=\"auto, (max-width: 853px) 100vw, 853px\" \/><\/figure>\n\n\n\n<p>\u5728\u5fae\u8c03\u5b9e\u9a8c\u4e2d\uff0c\u5bf9\u4e8e Qwen-7B \u548c Mistral-7B \u4e24\u79cd\u6a21\u578b\uff0cQ-Sparse \u663e\u793a\u51fa\u4e86\u4e0e\u7ee7\u7eed\u8bad\u7ec3\u76f8\u4f3c\u7684\u7ed3\u679c\uff0c\u752860%\u5de6\u53f3\u7684\u6fc0\u6d3b\u53c2\u6570\u5b9e\u73b0\u4e86\u4e0e\u5bc6\u96c6\u6a21\u578b\u5341\u5206\u63a5\u8fd1\u7684\u8868\u73b0\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"680\" height=\"257\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-13.png\" alt=\"Q-Sparse - Models table\" class=\"wp-image-1073526\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-13.png 680w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-13-300x113.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-13-240x91.png 240w\" sizes=\"auto, (max-width: 680px) 100vw, 680px\" \/><\/figure>\n\n\n\n<p>\u8fd9\u4e9b\u7ed3\u679c\u610f\u5473\u7740\uff0c\u5728\u76f8\u540c\u7684\u6027\u80fd\u4e0b\uff0c\u4e0e\u5bc6\u96c6\u6a21\u578b\u76f8\u6bd4\uff0c\u7a00\u758f\u6fc0\u6d3b\u6a21\u578b\u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d\u53ef\u4ee5\u663e\u8457\u51cf\u5c11\u6fc0\u6d3b\u53c2\u6570\uff0c\u8fdb\u800c\u964d\u4f4e\u6d88\u8017 FLOPS \u7684\u6570\u91cf\u3002<\/p>\n\n\n\n<p>\u5bf9\u4e8e\u91cf\u5316\u6a21\u578b\uff0c\u56e2\u961f\u5728\u81ea\u7814\u7684BitNet b1.58\u6a21\u578b\u4e0a\u5e94\u7528\u4e86 Q-Sparse\uff0c\u5e76\u5728\u591a\u4e2a\u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u4e86\u8bad\u7ec3\u548c\u8bc4\u4f30\u3002<\/p>\n\n\n\n<p>\u53ef\u4ee5\u770b\u5230\uff0c\u5728 700M \u548c 7B \u4e24\u79cd\u89c4\u6a21\u4e0b\uff0c\u4f7f\u7528 Q-Sparse \u7684\u91cf\u5316\u6a21\u578b\u7684\u6536\u655b\u901f\u5ea6\u548c\u6700\u7ec8\u635f\u5931\u51fd\u6570\u503c\u4e0e\u672a\u4f7f\u7528 Q-Sparse \u7684\u91cf\u5316\u6a21\u578b\uff08BitNet b1.58\uff09\u76f8\u5f53\u3002<\/p>\n\n\n\n<p>\u8fd9\u8bf4\u660e Q-Sparse \u53ef\u4ee5\u65e0\u7f1d\u96c6\u6210\u5230\u91cf\u5316\u6a21\u578b\u4e2d\uff0c\u800c\u4e0d\u4f1a\u663e\u8457\u5f71\u54cd\u6a21\u578b\u7684\u8bad\u7ec3\u548c\u6536\u655b\u3002<\/p>\n\n\n\n<p>\u636e\u6b64\u4f5c\u8005\u8ba4\u4e3a\uff0c\u5c06 Q-Sparse \u4e0e\u91cf\u5316\u6280\u672f\u76f8\u7ed3\u5408\uff0c\u53ef\u4ee5\u8fdb\u4e00\u6b65\u63d0\u9ad8\u5927\u8bed\u8a00\u6a21\u578b\u5728\u63a8\u7406\u9636\u6bb5\u7684\u6548\u7387\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"797\" height=\"351\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-14.png\" alt=\"Q-Sparse - Loss line charts by model size\" class=\"wp-image-1073529\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-14.png 797w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-14-300x132.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-14-768x338.png 768w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-14-240x106.png 240w\" sizes=\"auto, (max-width: 797px) 100vw, 797px\" \/><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"\u53d1\u73b0\u63a8\u7406\u4f18\u5316\u65b0-scaling-law\">\u53d1\u73b0\u63a8\u7406\u4f18\u5316\u65b0\u201cScaling Law\u201d<\/h2>\n\n\n\n<p>\u9664\u4e86\u6d4b\u8bc4\u8fd9\u4e9b\u6a21\u578b\u91c7\u53d6\u7a00\u758f\u6fc0\u6d3b\u65f6\u7684\u8868\u73b0\uff0c\u4f5c\u8005\u4e5f\u5bf9\u6a21\u578b\u6027\u80fd\u3001\u89c4\u6a21\u548c\u7a00\u758f\u7387\u4e09\u8005\u4e4b\u95f4\u7684\u5173\u7cfb\u8fdb\u884c\u4e86\u63a2\u7a76\uff0c\u5e76\u6709\u4e86\u4e00\u4e9b\u65b0\u7684\u53d1\u73b0\u3002<\/p>\n\n\n\n<p>\u7a00\u758f\u6fc0\u6d3b\u6a21\u578b\u7684\u6027\u80fd\u7f29\u653e\u5b9a\u5f8b: \u4f5c\u8005\u53d1\u73b0\uff0c\u4e0e\u5bc6\u96c6\u6a21\u578b\u7c7b\u4f3c\uff0c\u7a00\u758f\u6fc0\u6d3b\u6a21\u578b\u7684\u6027\u80fd\u4e5f\u9075\u5faa\u4e00\u4e2a\u5e42\u5f8b\u7f29\u653e\u5173\u7cfb\u3002<\/p>\n\n\n\n<p>\u5177\u4f53\u6765\u8bf4\uff0c\u7ed9\u5b9a\u7a00\u758f\u7387 S\uff0c\u6a21\u578b\u5728\u6536\u655b\u65f6\u7684\u635f\u5931\u51fd\u6570\u503c L(N,S) \u53ef\u4ee5\u7528\u4ee5\u4e0b\u516c\u5f0f\u8fd1\u4f3c\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"442\" height=\"87\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-15.png\" alt=\"Q-Sparse - formula\" class=\"wp-image-1073532\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-15.png 442w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-15-300x59.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-15-240x47.png 240w\" sizes=\"auto, (max-width: 442px) 100vw, 442px\" \/><\/figure>\n\n\n\n<p>\u5176\u4e2d\uff0cN \u662f\u6a21\u578b\u53c2\u6570\u7684\u6570\u91cf\uff1bE \u662f\u4e00\u4e2a\u5e38\u6570\uff0c\u8868\u793a\u6a21\u578b\u5728\u65e0\u9650\u5927\u65f6\u7684\u635f\u5931\uff1bA(S) \u662f\u4e00\u4e2a\u4e0e\u7a00\u758f\u7387 S \u6709\u5173\u7684\u7f29\u653e\u56e0\u5b50\u3002<\/p>\n\n\n\n<p>\u8fd9\u4e2a\u7f29\u653e\u5b9a\u5f8b\u8868\u660e\uff0c\u7a00\u758f\u6fc0\u6d3b\u6a21\u578b\u7684\u6027\u80fd\u968f\u7740\u6a21\u578b\u89c4\u6a21\u7684\u589e\u5927\u800c\u63d0\u9ad8\uff0c\u4f46\u63d0\u9ad8\u7684\u901f\u5ea6\u4f1a\u9010\u6e10\u53d8\u6162\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"420\" height=\"298\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-16.png\" alt=\"Q-Sparse - Loss line chart for Model Size (N)\" class=\"wp-image-1073535\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-16.png 420w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-16-300x213.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-16-240x170.png 240w\" sizes=\"auto, (max-width: 420px) 100vw, 420px\" \/><\/figure>\n\n\n\n<p>\u540c\u65f6\u4f5c\u8005\u53d1\u73b0\uff0c\u6a21\u578b\u7684\u6027\u80fd\u4e5f\u4f1a\u53d7\u5230\u7a00\u758f\u7387\u7684\u5f71\u54cd\u3002<\/p>\n\n\n\n<p>\u5728\u53c2\u6570\u89c4\u6a21\u4e0e\u6027\u80fd\u4e4b\u95f4\u5173\u7cfb\u7684\u90e8\u5206\u63d0\u5230\uff0cA(S) \u662f\u4e00\u4e2a\u4e0e\u7a00\u758f\u7387 S \u6709\u5173\u7684\u7f29\u653e\u56e0\u5b50\uff0c\u53ef\u4ee5\u7528\u4ee5\u4e0b\u516c\u5f0f\u8fd1\u4f3c\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"428\" height=\"101\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-17.png\" alt=\"Q-Sparse - formula\" class=\"wp-image-1073538\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-17.png 428w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-17-300x71.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-17-240x57.png 240w\" sizes=\"auto, (max-width: 428px) 100vw, 428px\" \/><\/figure>\n\n\n\n<p>\u5176\u4e2d B \u548c C \u662f\u5e38\u6570\uff0c\u03b2 \u662f\u4e00\u4e2a\u63a7\u5236\u6307\u6570\u8870\u51cf\u901f\u5ea6\u7684\u53c2\u6570\u3002<\/p>\n\n\n\n<p>\u8fd9\u4e2a\u516c\u5f0f\u8868\u660e\uff0c\u5f53\u7a00\u758f\u7387 S \u589e\u5927\uff08\u6a21\u578b\u53d8\u5f97\u66f4\u7a00\u758f\uff09\u65f6\uff0c\u610f\u5473\u7740\u66f4\u9ad8\u7684\u7a00\u758f\u7387\u4f1a\u5bfc\u81f4\u6027\u80fd\u7684\u4e0b\u964d\uff0c\u4e0b\u964d\u7684\u901f\u5ea6\u662f\u6307\u6570\u7ea7\u7684\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"424\" height=\"303\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-18.png\" alt=\"Q-Sparse - Loss line chart for Sparsity Ratio (S)\" class=\"wp-image-1073541\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-18.png 424w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-18-300x214.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-18-240x172.png 240w\" sizes=\"auto, (max-width: 424px) 100vw, 424px\" \/><\/figure>\n\n\n\n<p>\u57fa\u4e8e\u4e0a\u8ff0\u53d1\u73b0\uff0c\u4f5c\u8005\u5f97\u51fa\u4e86\u4e00\u4e2a\u63a8\u7406\u6700\u4f18\u7684\u7a00\u758f\u7387 S*\uff0c\u80fd\u5728\u9884\u7b97\uff08\u63a8\u7406\u65f6\u7684\u6d6e\u70b9\u64cd\u4f5c\u6570\uff09\u4e00\u5b9a\u65f6\uff0c\u5b9e\u73b0\u6a21\u578b\u635f\u5931\u51fd\u6570\u503c\u7684\u6700\u5c0f\u5316\u3002<\/p>\n\n\n\n<p>\u5bf9\u4e8e\u5168\u7cbe\u5ea6\uff08FP32\uff09\u6a21\u578b\uff0c\u6700\u4f18\u7a00\u758f\u7387\u7ea6\u4e3a45.58%\uff1b\u800c\u4f4e\u7cbe\u5ea6\uff08\u59821.58-bit\uff09\u6a21\u578b\u7684\u6700\u4f18\u7a00\u758f\u7387\u5219\u66f4\u9ad8\uff0c\u7ea6\u4e3a61.25%\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"846\" height=\"465\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-19.png\" alt=\"Q-Sparse - diagram\" class=\"wp-image-1073544\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-19.png 846w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-19-300x165.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-19-768x422.png 768w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/q-sparse-19-240x132.png 240w\" sizes=\"auto, (max-width: 846px) 100vw, 846px\" \/><\/figure>\n\n\n\n<p>\u4f5c\u8005\u89c2\u5bdf\u5230\uff0c\u968f\u7740\u6a21\u578b\u89c4\u6a21\u7684\u589e\u5927\uff0c\u7a00\u758f\u6fc0\u6d3b\u6a21\u578b\u4e0e\u5bc6\u96c6\u6a21\u578b\u4e4b\u95f4\u7684\u6027\u80fd\u5dee\u8ddd\u9010\u6e10\u7f29\u5c0f\u3002<\/p>\n\n\n\n<p>\u8fd9\u53ef\u4ee5\u4ece\u7f29\u653e\u5b9a\u5f8b\u4e2d\u5f97\u5230\u89e3\u91ca\uff1a\u5f53\u6a21\u578b\u89c4\u6a21 N \u8d8b\u4e8e\u65e0\u7a77\u5927\u65f6\uff0c\u7a00\u758f\u6fc0\u6d3b\u6a21\u578b\u7684\u635f\u5931\u51fd\u6570\u503c\u8d8b\u4e8e L(\u221e,S)=E\uff0c\u800c\u5bc6\u96c6\u6a21\u578b\u7684\u635f\u5931\u51fd\u6570\u503c\u8d8b\u4e8e L(\u221e,0)=E\u3002<\/p>\n\n\n\n<p>\u8fd9\u610f\u5473\u7740\uff0c\u5728\u6781\u5927\u89c4\u6a21\u4e0b\uff0c\u7a00\u758f\u6fc0\u6d3b\u6a21\u578b\u6709\u53ef\u80fd\u8fbe\u5230\u4e0e\u5bc6\u96c6\u6a21\u578b\u76f8\u5f53\u7684\u6027\u80fd\uff0c\u4e3a\u8bbe\u8ba1\u548c\u8bad\u7ec3\u5927\u89c4\u6a21\u7a00\u758f\u6fc0\u6d3b\u6a21\u578b\u63d0\u4f9b\u4e86\u4e00\u4e2a\u6709\u7528\u7684\u53c2\u8003\u3002<\/p>\n\n\n\n<p>\u8be6\u7ec6\u7ec6\u8282\u53ef\u4ee5\u67e5\u770b\u539f\u8bba\u6587\u3002<\/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\/2407.10969\" target=\"_blank\" rel=\"noopener noreferrer\">Q-Sparse: All Large Language Models can be Fully Sparsely-Activated<span class=\"sr-only\"> (opens in new tab)<\/span><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u4f5c\u8005\uff1a\u91cf\u5b50\u4f4d \u7f16\u8005\u6309\uff1a\u6fc0\u6d3b\u7a00\u758f\u6027\u662f\u89e3\u51b3\u5927\u8bed\u8a00\u6a21\u578b\uff08LLMs\uff09\u5728\u63a8\u7406\u9636\u6bb5\u51fa\u73b0\u7684\u8ba1\u7b97\u6210\u672c\u9ad8\u3001\u5185\u5b58\u5360\u7528\u5927\u7b49\u95ee\u9898\u7684\u6709\u6548\u65b9\u6cd5\uff0c\u53ef\u4ee5\u6709\u6548\u51cf\u5c11\u6fc0\u6d3b\u5f20\u91cf\u4e2d\u6fc0\u6d3b\u5143\u7d20\u7684\u6570\u91cf\u3002\u7136\u800c\u8be5\u65b9\u6cd5\u65e0\u6cd5\u5b9e\u73b0 LLMs \u6fc0\u6d3b\u7684\u5b8c\u5168\u7a00\u758f\u6027\uff0c\u4ece\u800c\u9650\u5236\u4e86\u63a8\u7406\u9636\u6bb5\u7684\u6548\u7387\u63d0\u5347\u3002 \u5bf9\u6b64\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u63d0\u51fa Q-Sparse \u5b9e\u73b0\u4e86 LLMs \u6fc0\u6d3b\u7684\u5b8c\u5168\u7a00\u758f\u6027\u3002\u8be5\u7814\u7a76\u4e0d\u4ec5\u63ed\u793a\u4e86\u5305\u62ec\u63a8\u7406\u4f18\u5316\u89c4\u6a21\u6cd5\u5219\uff08inference-optimal scaling law\uff09\u5728\u5185\u7684\u591a\u9879\u65b0\u53d1\u73b0\uff0c\u8fd8\u5c55\u73b0\u51fa\u5e7f\u6cdb\u7684\u517c\u5bb9\u6027\uff0c\u5176\u4e0e BitNet \u6b63\u4ea4\u4e14\u4e92\u8865\uff0c\u4e3a LLMs \u63a8\u7406\u4e2d\u7684\u6570\u636e\u7c7b\u578b\u63d0\u4f9b\u4e86\u5168\u9762\u4f18\u5316\u3002\u6b64\u5916\uff0cQ-Sparse \u5f3a\u5927\u7684\u9002\u7528\u6027\uff0c\u5bf9\u4ece\u5934\u5f00\u59cb\u8bad\u7ec3\u3001\u6301\u7eed\u8bad\u7ec3\u4ee5\u53ca\u5fae\u8c03\u90fd\u53ef\u4ee5\u63d0\u4f9b\u6709\u6548\u7684\u652f\u6301\u3002 \u672c\u6587\u8f6c\u8f7d\u81ea\u516c\u4f17\u53f7\u201c\u91cf\u5b50\u4f4d\u201d\uff0c\u539f\u6587\u6807\u9898\u300a\u53ea\u6fc0\u6d3b3.8B\u53c2\u6570\uff0c\u6027\u80fd\u6bd4\u80a9\u540c\u6b3e7B\u6a21\u578b\uff01\u8bad\u7ec3\u5fae\u8c03\u90fd\u80fd\u7528\uff0c\u6765\u81ea\u5fae\u8f6f\u300b\u3002 \u53ea\u9700\u6fc0\u6d3b60%\u7684\u53c2\u6570\uff0c\u5c31\u80fd\u5b9e\u73b0\u4e0e\u5168\u6fc0\u6d3b\u7a20\u5bc6\u6a21\u578b\u76f8\u5f53\u7684\u6027\u80fd\u3002 \u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u4e00\u9879\u65b0\u7814\u7a76\uff0c\u5b9e\u73b0\u4e86\u6a21\u578b\u7684\u5b8c\u5168\u7a00\u758f\u6fc0\u6d3b\uff0c\u8ba9\u63a8\u7406\u6210\u672c\u5927\u5e45\u4e0b\u964d\u3002 \u800c\u4e14\u9002\u7528\u8303\u56f4\u5e7f\u6cdb\uff0c\u65e0\u8bba\u662f\u4ece\u5934\u8bad\u7ec3\u3001\u7ee7\u7eed\u8bad\u7ec3\u8fd8\u662f\u5fae\u8c03\uff0c\u90fd\u80fd\u63d0\u4f9b\u6709\u6548\u652f\u6301\u3002 \u8be5\u65b9\u6cd5\u540d\u4e3a Q-Sparse\uff0c\u5728\u795e\u7ecf\u5143\u7ea7\u522b\u4e0a\u5b9e\u73b0\u4e86\u6a21\u578b\u7a00\u758f\u5316\uff0c\u76f8\u6bd4\u4e8e\u5176\u4ed6\u65b9\u5f0f\u7c92\u5ea6\u66f4\u7ec6\uff0c\u5728\u76f8\u540c\u63a8\u7406\u5f00\u9500\u4e0b\uff0c\u65e0\u8bba\u6027\u80fd\u8fd8\u662f\u7a00\u758f\u7387\u90fd\u66f4\u597d\u3002 \u540d\u79f0\u4e4b\u4e2d\uff0cQ \u6307\u7684\u662f\u91cf\u5316\uff08Quantization\uff09\uff0c\u610f\u5473\u7740\u5b83\u9664\u4e86\u666e\u901a\u6a21\u578b\u4e4b\u5916\uff0c\u4e5f\u517c\u5bb9\u91cf\u5316\u6280\u672f\uff0c\u9002\u7528\u4e8e\u5404\u79cd\u91cf\u5316\u65b9\u5f0f\u7684\u6a21\u578b\u3002\u5982\u679c\u628a Q-Sparse \u4e0e\u6a21\u578b\u91cf\u5316\u6280\u672f\u7ed3\u5408\uff0c\u8fd8\u53ef\u4ee5\u5b9e\u73b0\u66f4\u5927\u7a0b\u5ea6\u7684\u964d\u672c\u589e\u6548\u3002 \u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u5168\u7403\u7814\u7a76\u5408\u4f19\u4eba\u97e6\u798f\u5982\u8868\u793a\uff0cQ-Sparse \u5c55\u73b0\u51fa\u4e86\u5e7f\u6cdb\u7684\u517c\u5bb9\u6027\uff0c\u65e2\u9002\u7528\u4e8eFP16\/BF16\u7c7b\u578b\u7684LLMs\uff0c\u4e5f\u652f\u63011-bit LLMs\uff0c\u4f8b\u5982 BitNet\u3002\u503c\u5f97\u4e00\u63d0\u7684\u662f\uff0cQ-Sparse \u4e0e\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u6b64\u524d\u53d1\u5e03\u7684 BitNet \u6280\u672f\u6b63\u4ea4\u4e14\u4e92\u8865\uff0c\u4e3a LLMs \u63a8\u7406\u4e2d\u7684\u6570\u636e\u7c7b\u578b\u63d0\u4f9b\u4e86\u5168\u9762\u4f18\u5316\u2014\u2014BitNet \u4e13\u6ce8\u4e8e\u6a21\u578b\u6743\u91cd\u7684\u4f4e\u6bd4\u7279\uff08\u53731(.58)bit\uff09\uff0c\u800c Q-Sparse \u4e13\u6ce8\u4e8e\u7f51\u7edc\u6fc0\u6d3b\u7684\u7a00\u758f\u5316\u3002BitNet b1.58 \u548c Q-Sparse\uff08\u53ef\u914d\u5907 MoE\uff09\u7684\u7ed3\u5408\uff0c\u4e3a\u672a\u6765 LLMs \u5728\u63d0\u5347\u6548\u7387\u3001\u964d\u4f4e\u6210\u672c\u548c\u80fd\u8017\u65b9\u9762\u63d0\u4f9b\u4e86\u6e05\u6670\u7684\u8def\u5f84\u3002 \u53e6\u5916\u5728\u7814\u7a76 Q-Sparse \u7684\u540c\u65f6\uff0c\u56e2\u961f\u4e5f\u5bf9\u53c2\u6570\u89c4\u6a21\u3001\u7a00\u758f\u7387\u548c\u6a21\u578b\u6027\u80fd\u4e09\u8005\u4e4b\u95f4\u7684\u5173\u7cfb\u8fdb\u884c\u4e86\u6df1\u5165\u63a2\u5bfb\uff0c\u5e76\u53d1\u73b0\u4e86\u9002\u7528\u4e8e\u6a21\u578b\u63a8\u7406\u4f18\u5316\u7684\u201cScaling Law\u201d\u3002 \u6709\u7f51\u53cb\u8ba4\u4e3a\uff0c\u8fd9\u9879\u6280\u672f\u786e\u5b9e\u4e0d\u9519\uff0c\u800c\u4e14\u6bd4 ReLU \u8981\u66f4\u597d\u3002 \u8fd8\u6709\u4eba\u5f00\u542f\u4e86\u8bb8\u613f\u6a21\u5f0f\uff0c\u8868\u793a\u5982\u679c\uff08AMD\u7684\uff09ROCm \u80fd\u6bd4\u82f1\u4f1f\u8fbe\u66f4\u5feb\u652f\u6301\u8fd9\u9879\u6280\u672f\u5c31\u597d\u4e86\u3002 Q-Sparse \u6240\u505a\u7684\u6700\u6838\u5fc3\u7684\u64cd\u4f5c\uff0c\u662f\u5bf9\u8f93\u5165\u7684\u5f20\u91cf\u5e94\u7528 [&hellip;]<\/p>\n","protected":false},"author":42735,"featured_media":1076634,"template":"","meta":{"msr-url-field":"","msr-podcast-episode":"","msrModifiedDate":"","msrModifiedDateEnabled":false,"ep_exclude_from_search":false,"_classifai_error":"","msr-content-parent":1012650,"msr_hide_image_in_river":0,"footnotes":""},"research-area":[],"msr-locale":[268881],"msr-post-option":[],"class_list":["post-1073484","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:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1073484","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post"}],"about":[{"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/types\/msr-blog-post"}],"author":[{"embeddable":true,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/users\/42735"}],"version-history":[{"count":6,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1073484\/revisions"}],"predecessor-version":[{"id":1076637,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1073484\/revisions\/1076637"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/media\/1076634"}],"wp:attachment":[{"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/media?parent=1073484"}],"wp:term":[{"taxonomy":"msr-research-area","embeddable":true,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/research-area?post=1073484"},{"taxonomy":"msr-locale","embeddable":true,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-locale?post=1073484"},{"taxonomy":"msr-post-option","embeddable":true,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-post-option?post=1073484"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}