{"id":1072113,"date":"2023-05-17T20:46:00","date_gmt":"2023-05-18T03:46:00","guid":{"rendered":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/?post_type=msr-blog-post&#038;p=1072113"},"modified":"2024-09-25T03:37:36","modified_gmt":"2024-09-25T10:37:36","slug":"llm-accelerator","status":"publish","type":"msr-blog-post","link":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/articles\/llm-accelerator\/","title":{"rendered":"LLM Accelerator\uff1a\u4f7f\u7528\u53c2\u8003\u6587\u672c\u65e0\u635f\u52a0\u901f\u5927\u8bed\u8a00\u6a21\u578b\u63a8\u7406"},"content":{"rendered":"\n<p><em>\u4f5c\u8005\uff1a\u6768\u5357<\/em><\/p>\n\n\n\n<p>\u7f16\u8005\u6309\uff1a\u5982\u4eca\uff0c\u57fa\u7840\u5927\u6a21\u578b\u6b63\u5728\u8bf8\u591a\u5e94\u7528\u4e2d\u53d1\u6325\u7740\u65e5\u76ca\u91cd\u8981\u7684\u4f5c\u7528\u3002\u5927\u591a\u6570\u5927\u8bed\u8a00\u6a21\u578b\u7684\u8bad\u7ec3\u90fd\u662f\u91c7\u53d6\u81ea\u56de\u5f52\u7684\u65b9\u5f0f\u8fdb\u884c\u751f\u6210\uff0c\u867d\u7136\u81ea\u56de\u5f52\u6a21\u578b\u751f\u6210\u7684\u6587\u672c\u8d28\u91cf\u6709\u6240\u4fdd\u8bc1\uff0c\u4f46\u5374\u5bfc\u81f4\u4e86\u9ad8\u6602\u7684\u63a8\u7406\u6210\u672c\u548c\u957f\u65f6\u95f4\u7684\u5ef6\u8fdf\u3002\u7531\u4e8e\u5927\u6a21\u578b\u7684\u53c2\u6570\u91cf\u5de8\u5927\u3001\u63a8\u7406\u6210\u672c\u9ad8\uff0c\u56e0\u6b64\u5982\u4f55\u5728\u5927\u89c4\u6a21\u90e8\u7f72\u5927\u6a21\u578b\u7684\u8fc7\u7a0b\u4e2d\u964d\u4f4e\u6210\u672c\u3001\u51cf\u5c0f\u5ef6\u8fdf\u662f\u4e00\u4e2a\u5173\u952e\u8bfe\u9898\u3002\u9488\u5bf9\u6b64\u95ee\u9898\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u4f7f\u7528\u53c2\u8003\u6587\u672c\u65e0\u635f\u52a0\u901f\u5927\u8bed\u8a00\u6a21\u578b\u63a8\u7406\u7684\u65b9\u6cd5 LLM Accelerator\uff0c\u5728\u5927\u6a21\u578b\u5178\u578b\u7684\u5e94\u7528\u573a\u666f\u4e2d\u53ef\u4ee5\u53d6\u5f97\u4e24\u5230\u4e09\u500d\u7684\u52a0\u901f\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<p>\u968f\u7740\u4eba\u5de5\u667a\u80fd\u6280\u672f\u7684\u5feb\u901f\u53d1\u5c55\uff0cChatGPT\u3001New Bing\u3001GPT-4 \u7b49\u65b0\u4ea7\u54c1\u548c\u65b0\u6280\u672f\u9646\u7eed\u53d1\u5e03\uff0c\u57fa\u7840\u5927\u6a21\u578b\u5728\u8bf8\u591a\u5e94\u7528\u4e2d\u5c06\u53d1\u6325\u65e5\u76ca\u91cd\u8981\u7684\u4f5c\u7528\u3002\u76ee\u524d\u7684\u5927\u8bed\u8a00\u6a21\u578b\u5927\u591a\u662f\u81ea\u56de\u5f52\u6a21\u578b\u3002\u81ea\u56de\u5f52\u662f\u6307\u6a21\u578b\u5728\u8f93\u51fa\u65f6\u5f80\u5f80\u91c7\u7528\u9010\u8bcd\u8f93\u51fa\u7684\u65b9\u5f0f\uff0c\u5373\u5728\u8f93\u51fa\u6bcf\u4e2a\u8bcd\u65f6\uff0c\u6a21\u578b\u9700\u8981\u5c06\u4e4b\u524d\u8f93\u51fa\u7684\u8bcd\u4f5c\u4e3a\u8f93\u5165\u3002\u800c\u8fd9\u79cd\u81ea\u56de\u5f52\u6a21\u5f0f\u901a\u5e38\u5728\u8f93\u51fa\u65f6\u5236\u7ea6\u7740\u5e76\u884c\u52a0\u901f\u5668\u7684\u5145\u5206\u5229\u7528\u3002<\/p>\n\n\n\n<p>\u5728\u8bb8\u591a\u5e94\u7528\u573a\u666f\u4e2d\uff0c\u5927\u6a21\u578b\u7684\u8f93\u51fa\u5e38\u5e38\u4e0e\u4e00\u4e9b\u53c2\u8003\u6587\u672c\u6709\u5f88\u5927\u7684\u76f8\u4f3c\u6027\uff0c\u4f8b\u5982\u5728\u4ee5\u4e0b\u4e09\u4e2a\u5e38\u89c1\u7684\u573a\u666f\u4e2d\uff1a<\/p>\n\n\n\n<p>1. \u68c0\u7d22\u589e\u5f3a\u7684\u751f\u6210\u3002New Bing \u7b49\u68c0\u7d22\u5e94\u7528\u5728\u54cd\u5e94\u7528\u6237\u8f93\u5165\u7684\u5185\u5bb9\u65f6\uff0c\u4f1a\u5148\u8fd4\u56de\u4e00\u4e9b\u4e0e\u7528\u6237\u8f93\u5165\u76f8\u5173\u7684\u4fe1\u606f\uff0c\u7136\u540e\u7528\u8bed\u8a00\u6a21\u578b\u603b\u7ed3\u68c0\u7d22\u51fa\u7684\u4fe1\u606f\uff0c\u518d\u56de\u7b54\u7528\u6237\u8f93\u5165\u7684\u5185\u5bb9\u3002\u5728\u8fd9\u79cd\u573a\u666f\u4e2d\uff0c\u6a21\u578b\u7684\u8f93\u51fa\u5f80\u5f80\u5305\u542b\u5927\u91cf\u68c0\u7d22\u7ed3\u679c\u4e2d\u7684\u6587\u672c\u7247\u6bb5\u3002<\/p>\n\n\n\n<p>2. \u4f7f\u7528\u7f13\u5b58\u7684\u751f\u6210\u3002\u5927\u89c4\u6a21\u90e8\u7f72\u8bed\u8a00\u6a21\u578b\u7684\u8fc7\u7a0b\u4e2d\uff0c\u5386\u53f2\u7684\u8f93\u5165\u8f93\u51fa\u4f1a\u88ab\u7f13\u5b58\u3002\u5728\u5904\u7406\u65b0\u7684\u8f93\u5165\u65f6\uff0c\u68c0\u7d22\u5e94\u7528\u4f1a\u5728\u7f13\u5b58\u4e2d\u5bfb\u627e\u76f8\u4f3c\u7684\u8f93\u5165\u3002\u56e0\u6b64\uff0c\u6a21\u578b\u7684\u8f93\u51fa\u5f80\u5f80\u548c\u7f13\u5b58\u4e2d\u5bf9\u5e94\u7684\u8f93\u51fa\u6709\u5f88\u5927\u7684\u76f8\u4f3c\u6027\u3002<\/p>\n\n\n\n<p>3. \u591a\u8f6e\u5bf9\u8bdd\u4e2d\u7684\u751f\u6210\u3002\u5728\u4f7f\u7528 ChatGPT \u7b49\u5e94\u7528\u65f6\uff0c\u7528\u6237\u5f80\u5f80\u4f1a\u6839\u636e\u6a21\u578b\u7684\u8f93\u51fa\u53cd\u590d\u63d0\u51fa\u4fee\u6539\u8981\u6c42\u3002\u5728\u8fd9\u79cd\u591a\u8f6e\u5bf9\u8bdd\u7684\u573a\u666f\u4e0b\uff0c\u6a21\u578b\u7684\u591a\u6b21\u8f93\u51fa\u5f80\u5f80\u53ea\u6709\u5c11\u91cf\u7684\u53d8\u5316\uff0c\u91cd\u590d\u5ea6\u8f83\u9ad8\u3002<\/p>\n\n\n\n<p><\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"437\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/llm-accelerator-1-1024x437.png\" alt=\"\u56fe1\uff1a\u5927\u6a21\u578b\u7684\u8f93\u51fa\u4e0e\u53c2\u8003\u6587\u672c\u5b58\u5728\u76f8\u4f3c\u6027\u7684\u5e38\u89c1\u573a\u666f\" class=\"wp-image-1072260\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/llm-accelerator-1-1024x437.png 1024w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/llm-accelerator-1-300x128.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/llm-accelerator-1-768x328.png 768w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/llm-accelerator-1-240x102.png 240w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/llm-accelerator-1.png 1286w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><figcaption class=\"wp-element-caption\">\u56fe1\uff1a\u5927\u6a21\u578b\u7684\u8f93\u51fa\u4e0e\u53c2\u8003\u6587\u672c\u5b58\u5728\u76f8\u4f3c\u6027\u7684\u5e38\u89c1\u573a\u666f<\/figcaption><\/figure>\n\n\n\n<p id=\"caption-attachment-44990\"><\/p>\n\n\n\n<p>\u57fa\u4e8e\u4ee5\u4e0a\u89c2\u5bdf\uff0c\u7814\u7a76\u5458\u4eec\u4ee5\u53c2\u8003\u6587\u672c\u4e0e\u6a21\u578b\u8f93\u51fa\u7684\u91cd\u590d\u6027\u4f5c\u4e3a\u7a81\u7834\u81ea\u56de\u5f52\u74f6\u9888\u7684\u7740\u529b\u70b9\uff0c\u5e0c\u671b\u53ef\u4ee5\u63d0\u9ad8\u5e76\u884c\u52a0\u901f\u5668\u5229\u7528\u7387\uff0c\u52a0\u901f\u5927\u8bed\u8a00\u6a21\u578b\u63a8\u7406\uff0c\u8fdb\u800c\u63d0\u51fa\u4e86\u4e00\u79cd\u5229\u7528\u8f93\u51fa\u4e0e\u53c2\u8003\u6587\u672c\u7684\u91cd\u590d\u6027\u6765\u5b9e\u73b0\u4e00\u6b65\u8f93\u51fa\u591a\u4e2a\u8bcd\u7684\u65b9\u6cd5 LLM Accelerator\u3002<\/p>\n\n\n\n<p><\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"948\" height=\"436\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/llm-accelerator-2.png\" alt=\"\u56fe2\uff1aLLM Accelerator \u89e3\u7801\u7b97\u6cd5\" class=\"wp-image-1072263\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/llm-accelerator-2.png 948w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/llm-accelerator-2-300x138.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/llm-accelerator-2-768x353.png 768w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/llm-accelerator-2-240x110.png 240w\" sizes=\"auto, (max-width: 948px) 100vw, 948px\" \/><figcaption class=\"wp-element-caption\">\u56fe2\uff1aLLM Accelerator \u89e3\u7801\u7b97\u6cd5<\/figcaption><\/figure>\n\n\n\n<p id=\"caption-attachment-44991\"><\/p>\n\n\n\n<p>\u5177\u4f53\u6765\u8bf4\uff0c\u5728\u6bcf\u4e00\u6b65\u89e3\u7801\u65f6\uff0c\u8ba9\u6a21\u578b\u5148\u5339\u914d\u5df2\u6709\u7684\u8f93\u51fa\u7ed3\u679c\u4e0e\u53c2\u8003\u6587\u672c\uff0c\u5982\u679c\u53d1\u73b0\u67d0\u4e2a\u53c2\u8003\u6587\u672c\u4e0e\u5df2\u6709\u7684\u8f93\u51fa\u76f8\u7b26\uff0c\u90a3\u4e48\u6a21\u578b\u5f88\u53ef\u80fd\u987a\u5ef6\u5df2\u6709\u7684\u53c2\u8003\u6587\u672c\u7ee7\u7eed\u8f93\u51fa\u3002\u56e0\u6b64\uff0c\u7814\u7a76\u5458\u4eec\u5c06\u53c2\u8003\u6587\u672c\u7684\u540e\u7eed\u8bcd\u4e5f\u4f5c\u4e3a\u8f93\u5165\u52a0\u5165\u5230\u6a21\u578b\u4e2d\uff0c\u4ece\u800c\u4f7f\u5f97\u4e00\u4e2a\u89e3\u7801\u6b65\u9aa4\u53ef\u4ee5\u8f93\u51fa\u591a\u4e2a\u8bcd\u3002<\/p>\n\n\n\n<p>\u4e3a\u4e86\u4fdd\u8bc1\u8f93\u5165\u8f93\u51fa\u51c6\u786e\uff0c\u7814\u7a76\u5458\u4eec\u8fdb\u4e00\u6b65\u5bf9\u6bd4\u4e86\u6a21\u578b\u8f93\u51fa\u7684\u8bcd\u4e0e\u4ece\u53c2\u8003\u6587\u6863\u8f93\u5165\u7684\u8bcd\u3002\u5982\u679c\u4e24\u8005\u4e0d\u4e00\u81f4\uff0c\u90a3\u4e48\u4e0d\u6b63\u786e\u7684\u8f93\u5165\u8f93\u51fa\u7ed3\u679c\u5c06\u88ab\u820d\u5f03\u3002\u4ee5\u4e0a\u65b9\u6cd5\u80fd\u591f\u4fdd\u8bc1\u89e3\u7801\u7ed3\u679c\u4e0e\u57fa\u51c6\u65b9\u6cd5\u5b8c\u5168\u4e00\u81f4\uff0c\u5e76\u53ef\u4ee5\u63d0\u9ad8\u6bcf\u4e2a\u89e3\u7801\u6b65\u9aa4\u7684\u8f93\u51fa\u8bcd\u6570\uff0c\u4ece\u800c\u5b9e\u73b0\u5927\u6a21\u578b\u63a8\u7406\u7684\u65e0\u635f\u52a0\u901f\u3002LLM Accelerator \u65e0\u9700\u989d\u5916\u8f85\u52a9\u6a21\u578b\uff0c\u7b80\u5355\u6613\u7528\uff0c\u53ef\u4ee5\u65b9\u4fbf\u5730\u90e8\u7f72\u5230\u5404\u79cd\u5e94\u7528\u573a\u666f\u4e2d\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"702\" height=\"212\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/llm-accelerator-3.png\" alt=\"paper screenshot\" class=\"wp-image-1072266\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/llm-accelerator-3.png 702w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/llm-accelerator-3-300x91.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/llm-accelerator-3-240x72.png 240w\" sizes=\"auto, (max-width: 702px) 100vw, 702px\" \/><\/figure>\n\n\n\n<p>\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2304.04487.pdf<\/p>\n\n\n\n<p>\u9879\u76ee\u94fe\u63a5\uff1ahttps:\/\/github.com\/microsoft\/LMOps<\/p>\n\n\n\n<p>\u4f7f\u7528 LLM Accelerator\uff0c\u6709\u4e24\u4e2a\u8d85\u53c2\u6570\u9700\u8981\u8c03\u6574\u3002\u4e00\u662f\u89e6\u53d1\u5339\u914d\u673a\u5236\u6240\u9700\u7684\u8f93\u51fa\u4e0e\u53c2\u8003\u6587\u672c\u7684\u5339\u914d\u8bcd\u6570\uff1a\u5339\u914d\u8bcd\u6570\u8d8a\u957f\u5f80\u5f80\u8d8a\u51c6\u786e\uff0c\u53ef\u4ee5\u66f4\u597d\u5730\u4fdd\u8bc1\u4ece\u53c2\u8003\u6587\u672c\u62f7\u8d1d\u7684\u8bcd\u662f\u6b63\u786e\u7684\u8f93\u51fa\uff0c\u51cf\u5c11\u4e0d\u5fc5\u8981\u7684\u89e6\u53d1\u548c\u8ba1\u7b97\uff1b\u66f4\u77ed\u7684\u5339\u914d\uff0c\u89e3\u7801\u6b65\u9aa4\u66f4\u5c11\uff0c\u6f5c\u5728\u52a0\u901f\u66f4\u5feb\u3002\u4e8c\u662f\u6bcf\u6b21\u62f7\u8d1d\u8bcd\u7684\u6570\u91cf\uff1a\u62f7\u8d1d\u8bcd\u6570\u8d8a\u591a\uff0c\u52a0\u901f\u6f5c\u529b\u8d8a\u5927\uff0c\u4f46\u4e5f\u53ef\u80fd\u9020\u6210\u66f4\u591a\u4e0d\u6b63\u786e\u7684\u8f93\u51fa\u88ab\u820d\u5f03\uff0c\u6d6a\u8d39\u8ba1\u7b97\u8d44\u6e90\u3002\u7814\u7a76\u5458\u4eec\u901a\u8fc7\u5b9e\u9a8c\u53d1\u73b0\uff0c\u66f4\u52a0\u6fc0\u8fdb\u7684\u7b56\u7565\uff08\u5339\u914d\u5355\u4e2a\u8bcd\u89e6\u53d1\uff0c\u4e00\u6b21\u62f7\u8d1d15\u523020\u4e2a\u8bcd\uff09\u5f80\u5f80\u80fd\u591f\u53d6\u5f97\u66f4\u597d\u7684\u52a0\u901f\u6bd4\u3002<\/p>\n\n\n\n<p>\u4e3a\u4e86\u9a8c\u8bc1 LLM Accelerator \u7684\u6709\u6548\u6027\uff0c\u7814\u7a76\u5458\u4eec\u5728\u68c0\u7d22\u589e\u5f3a\u548c\u7f13\u5b58\u8f85\u52a9\u751f\u6210\u65b9\u9762\u8fdb\u884c\u4e86\u5b9e\u9a8c\uff0c\u5229\u7528 MS-MARCO \u6bb5\u843d\u68c0\u7d22\u6570\u636e\u96c6\u6784\u9020\u4e86\u5b9e\u9a8c\u6837\u672c\u3002\u5728\u68c0\u7d22\u589e\u5f3a\u5b9e\u9a8c\u4e2d\uff0c\u7814\u7a76\u5458\u4eec\u4f7f\u7528\u68c0\u7d22\u6a21\u578b\u5bf9\u6bcf\u4e2a\u67e5\u8be2\u8fd4\u56de10\u4e2a\u6700\u76f8\u5173\u7684\u6587\u6863\uff0c\u7136\u540e\u62fc\u63a5\u5230\u67e5\u8be2\u540e\u4f5c\u4e3a\u6a21\u578b\u8f93\u5165\uff0c\u5c06\u8fd910\u4e2a\u6587\u6863\u4f5c\u4e3a\u53c2\u8003\u6587\u672c\u3002\u5728\u7f13\u5b58\u8f85\u52a9\u751f\u6210\u5b9e\u9a8c\u4e2d\uff0c\u6bcf\u4e2a\u67e5\u8be2\u751f\u6210\u56db\u4e2a\u76f8\u4f3c\u7684\u67e5\u8be2\uff0c\u7136\u540e\u7528\u6a21\u578b\u8f93\u51fa\u5bf9\u5e94\u7684\u67e5\u8be2\u4f5c\u4e3a\u53c2\u8003\u6587\u672c\u3002<\/p>\n\n\n\n<p><\/p>\n\n\n\n<figure class=\"wp-block-image size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"965\" height=\"217\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/llm-accelerator-4.png\" alt=\"\u88681\uff1a\u68c0\u7d22\u589e\u5f3a\u7684\u751f\u6210\u573a\u666f\u4e0b\u7684\u65f6\u95f4\u5bf9\u6bd4\" class=\"wp-image-1072269\" style=\"width:658px;height:auto\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/llm-accelerator-4.png 965w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/llm-accelerator-4-300x67.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/llm-accelerator-4-768x173.png 768w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/llm-accelerator-4-240x54.png 240w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/llm-accelerator-4-960x217.png 960w\" sizes=\"auto, (max-width: 965px) 100vw, 965px\" \/><figcaption class=\"wp-element-caption\">\u88681\uff1a\u68c0\u7d22\u589e\u5f3a\u7684\u751f\u6210\u573a\u666f\u4e0b\u7684\u65f6\u95f4\u5bf9\u6bd4<\/figcaption><\/figure>\n\n\n\n<p id=\"caption-attachment-44993\"><\/p>\n\n\n\n<p><\/p>\n\n\n\n<figure class=\"wp-block-image size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"958\" height=\"217\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/llm-accelerator-5.png\" alt=\"\u88682\uff1a\u4f7f\u7528\u7f13\u5b58\u7684\u751f\u6210\u573a\u666f\u4e0b\u7684\u65f6\u95f4\u5bf9\u6bd4\" class=\"wp-image-1072272\" style=\"width:657px;height:auto\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/llm-accelerator-5.png 958w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/llm-accelerator-5-300x68.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/llm-accelerator-5-768x174.png 768w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/llm-accelerator-5-240x54.png 240w\" sizes=\"auto, (max-width: 958px) 100vw, 958px\" \/><figcaption class=\"wp-element-caption\">\u88682\uff1a\u4f7f\u7528\u7f13\u5b58\u7684\u751f\u6210\u573a\u666f\u4e0b\u7684\u65f6\u95f4\u5bf9\u6bd4<\/figcaption><\/figure>\n\n\n\n<p id=\"caption-attachment-44994\"><\/p>\n\n\n\n<p>\u7814\u7a76\u5458\u4eec\u4f7f\u7528\u901a\u8fc7 OpenAI \u63a5\u53e3\u5f97\u5230\u7684 Davinci-003 \u6a21\u578b\u7684\u8f93\u51fa\u7ed3\u679c\u4f5c\u4e3a\u76ee\u6807\u8f93\u51fa\uff0c\u4ee5\u83b7\u5f97\u9ad8\u8d28\u91cf\u7684\u8f93\u51fa\u3002\u5f97\u5230\u6240\u9700\u8f93\u5165\u3001\u8f93\u51fa\u548c\u53c2\u8003\u6587\u672c\u540e\uff0c\u7814\u7a76\u5458\u4eec\u5728\u5f00\u6e90\u7684 LLaMA \u8bed\u8a00\u6a21\u578b\u4e0a\u8fdb\u884c\u4e86\u5b9e\u9a8c\u3002\u7531\u4e8e LLaMA \u6a21\u578b\u7684\u8f93\u51fa\u4e0e Davinci-003 \u8f93\u51fa\u4e0d\u4e00\u81f4\uff0c\u6240\u4ee5\u7814\u7a76\u5458\u4eec\u91c7\u7528\u4e86\u76ee\u6807\u5bfc\u5411\u7684\u89e3\u7801\u65b9\u6cd5\u6765\u6d4b\u8bd5\u7406\u60f3\u8f93\u51fa\uff08Davinci-003 \u6a21\u578b\u7ed3\u679c\uff09\u7ed3\u679c\u4e0b\u7684\u52a0\u901f\u6bd4\u3002\u7814\u7a76\u5458\u4eec\u5229\u7528\u7b97\u6cd52\u5f97\u5230\u4e86\u8d2a\u5a6a\u89e3\u7801\u65f6\u751f\u6210\u76ee\u6807\u8f93\u51fa\u6240\u9700\u7684\u89e3\u7801\u6b65\u9aa4\uff0c\u5e76\u5f3a\u5236 LLaMA \u6a21\u578b\u6309\u7167\u5f97\u5230\u7684\u89e3\u7801\u6b65\u9aa4\u8fdb\u884c\u89e3\u7801\u3002<\/p>\n\n\n\n<p><\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"537\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/llm-accelerator-6-1024x537.png\" alt=\"\u56fe3\uff1a\u5229\u7528\u7b97\u6cd52\u5f97\u5230\u4e86\u8d2a\u5a6a\u89e3\u7801\u65f6\u751f\u6210\u76ee\u6807\u8f93\u51fa\u6240\u9700\u7684\u89e3\u7801\u6b65\u9aa4\" class=\"wp-image-1072275\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/llm-accelerator-6-1024x537.png 1024w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/llm-accelerator-6-300x157.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/llm-accelerator-6-768x403.png 768w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/llm-accelerator-6-240x126.png 240w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/llm-accelerator-6.png 1045w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><figcaption class=\"wp-element-caption\">\u56fe3\uff1a\u5229\u7528\u7b97\u6cd52\u5f97\u5230\u4e86\u8d2a\u5a6a\u89e3\u7801\u65f6\u751f\u6210\u76ee\u6807\u8f93\u51fa\u6240\u9700\u7684\u89e3\u7801\u6b65\u9aa4<\/figcaption><\/figure>\n\n\n\n<p id=\"caption-attachment-44995\"><\/p>\n\n\n\n<p>\u5bf9\u4e8e\u53c2\u6570\u91cf\u4e3a 7B \u548c 13B \u7684\u6a21\u578b\uff0c\u7814\u7a76\u5458\u4eec\u5728\u5355\u4e2a 32G NVIDIA V100 GPU \u4e0a\u8fdb\u884c\u5b9e\u9a8c\uff1b\u5bf9\u4e8e\u53c2\u6570\u91cf\u4e3a 30B \u7684\u6a21\u578b\uff0c\u5728\u56db\u5757\u540c\u6837\u7684 GPU \u4e0a\u8fdb\u884c\u5b9e\u9a8c\u3002\u6240\u6709\u7684\u5b9e\u9a8c\u5747\u91c7\u7528\u4e86\u534a\u7cbe\u5ea6\u6d6e\u70b9\u6570\uff0c\u89e3\u7801\u5747\u4e3a\u8d2a\u5a6a\u89e3\u7801\uff0c\u4e14\u6279\u91cf\u5927\u5c0f\u4e3a1\u3002\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0cLLM Accelerator \u5728\u4e0d\u540c\u6a21\u578b\u5927\u5c0f\uff087B\uff0c13B\uff0c30B\uff09\u4e0e\u4e0d\u540c\u7684\u5e94\u7528\u573a\u666f\u4e2d\uff08\u68c0\u7d22\u589e\u5f3a\u3001\u7f13\u5b58\u8f85\u52a9\uff09\u90fd\u53d6\u5f97\u4e86\u4e24\u5230\u4e09\u500d\u7684\u52a0\u901f\u6bd4\u3002<\/p>\n\n\n\n<p>\u8fdb\u4e00\u6b65\u5b9e\u9a8c\u5206\u6790\u53d1\u73b0\uff0cLLM Accelertator \u80fd\u663e\u8457\u51cf\u5c11\u6240\u9700\u7684\u89e3\u7801\u6b65\u9aa4\uff0c\u5e76\u4e14\u52a0\u901f\u6bd4\u4e0e\u89e3\u7801\u6b65\u9aa4\u7684\u51cf\u5c11\u6bd4\u4f8b\u5448\u6b63\u76f8\u5173\u3002\u66f4\u5c11\u7684\u89e3\u7801\u6b65\u9aa4\u4e00\u65b9\u9762\u610f\u5473\u7740\u6bcf\u4e2a\u89e3\u7801\u6b65\u9aa4\u751f\u6210\u7684\u8f93\u51fa\u8bcd\u6570\u66f4\u591a\uff0c\u53ef\u4ee5\u63d0\u9ad8 GPU \u8ba1\u7b97\u7684\u8ba1\u7b97\u6548\u7387\uff1b\u53e6\u4e00\u65b9\u9762\uff0c\u5bf9\u4e8e\u9700\u8981\u591a\u5361\u5e76\u884c\u768430B\u6a21\u578b\uff0c\u8fd9\u610f\u5473\u7740\u66f4\u5c11\u7684\u591a\u5361\u540c\u6b65\uff0c\u4ece\u800c\u8fbe\u5230\u66f4\u5feb\u7684\u901f\u5ea6\u63d0\u5347\u3002\u5728\u6d88\u878d\u5b9e\u9a8c\u4e2d\uff0c\u5728\u5f00\u53d1\u96c6\u4e0a\u5bf9 LLM Accelertator \u7684\u8d85\u53c2\u6570\u8fdb\u884c\u5206\u6790\u7684\u7ed3\u679c\u663e\u793a\uff0c\u5339\u914d\u5355\u4e2a\u5355\u8bcd\uff08\u5373\u89e6\u53d1\u62f7\u8d1d\u673a\u5236\uff09\u65f6\uff0c\u4e00\u6b21\u62f7\u8d1d15\u523020\u4e2a\u5355\u8bcd\u65f6\u7684\u52a0\u901f\u6bd4\u53ef\u8fbe\u5230\u6700\u5927 \uff08\u56fe4\u6240\u793a\uff09\u3002\u5728\u56fe5\u4e2d\u6211\u4eec\u53ef\u4ee5\u770b\u51fa\uff0c\u5339\u914d\u8bcd\u6570\u4e3a1\u80fd\u66f4\u591a\u5730\u89e6\u53d1\u62f7\u8d1d\u673a\u5236\uff0c\u5e76\u4e14\u968f\u7740\u62f7\u8d1d\u957f\u5ea6\u7684\u589e\u52a0\uff0c\u6bcf\u4e2a\u89e3\u7801\u6b65\u9aa4\u63a5\u53d7\u7684\u8f93\u51fa\u8bcd\u589e\u52a0\uff0c\u89e3\u7801\u6b65\u9aa4\u51cf\u5c11\uff0c\u4ece\u800c\u8fbe\u5230\u66f4\u9ad8\u7684\u52a0\u901f\u6bd4\u3002<\/p>\n\n\n\n<p><\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"472\" height=\"385\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/llm-accelerator-7.png\" alt=\"\u56fe4\uff1a\u6d88\u878d\u5b9e\u9a8c\u4e2d\uff0c\u5728\u5f00\u53d1\u96c6\u4e0a\u5bf9 LLM Accelertator \u7684\u8d85\u53c2\u6570\u7684\u5206\u6790\u7ed3\u679c\" class=\"wp-image-1072278\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/llm-accelerator-7.png 472w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/llm-accelerator-7-300x245.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/llm-accelerator-7-221x180.png 221w\" sizes=\"auto, (max-width: 472px) 100vw, 472px\" \/><figcaption class=\"wp-element-caption\">\u56fe4\uff1a\u6d88\u878d\u5b9e\u9a8c\u4e2d\uff0c\u5728\u5f00\u53d1\u96c6\u4e0a\u5bf9 LLM Accelertator \u7684\u8d85\u53c2\u6570\u7684\u5206\u6790\u7ed3\u679c<\/figcaption><\/figure>\n\n\n\n<p id=\"caption-attachment-44996\"><\/p>\n\n\n\n<p><\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"683\" height=\"360\" src=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/llm-accelerator-8.png\" alt=\"\u56fe5\uff1a\u5728\u5f00\u53d1\u96c6\u4e0a\uff0c\u5177\u6709\u4e0d\u540c\u5339\u914d\u8bcd\u6570 n \u548c\u62f7\u8d1d\u8bcd\u6570 k \u7684\u89e3\u7801\u6b65\u9aa4\u7edf\u8ba1\u6570\u636e\" class=\"wp-image-1072281\" srcset=\"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/llm-accelerator-8.png 683w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/llm-accelerator-8-300x158.png 300w, https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-content\/uploads\/2024\/08\/llm-accelerator-8-240x127.png 240w\" sizes=\"auto, (max-width: 683px) 100vw, 683px\" \/><figcaption class=\"wp-element-caption\">\u56fe5\uff1a\u5728\u5f00\u53d1\u96c6\u4e0a\uff0c\u5177\u6709\u4e0d\u540c\u5339\u914d\u8bcd\u6570 n \u548c\u62f7\u8d1d\u8bcd\u6570 k \u7684\u89e3\u7801\u6b65\u9aa4\u7edf\u8ba1\u6570\u636e<\/figcaption><\/figure>\n\n\n\n<p id=\"caption-attachment-44997\"><\/p>\n\n\n\n<p>LLM Accelertator \u662f\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u81ea\u7136\u8bed\u8a00\u8ba1\u7b97\u7ec4\u5728\u5927\u8bed\u8a00\u6a21\u578b\u52a0\u901f\u7cfb\u5217\u5de5\u4f5c\u7684\u4e00\u90e8\u5206\uff0c\u672a\u6765\uff0c\u7814\u7a76\u5458\u4eec\u5c06\u6301\u7eed\u5bf9\u76f8\u5173\u95ee\u9898\u8fdb\u884c\u66f4\u52a0\u6df1\u5165\u7684\u63a2\u7d22\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u4f5c\u8005\uff1a\u6768\u5357 \u7f16\u8005\u6309\uff1a\u5982\u4eca\uff0c\u57fa\u7840\u5927\u6a21\u578b\u6b63\u5728\u8bf8\u591a\u5e94\u7528\u4e2d\u53d1\u6325\u7740\u65e5\u76ca\u91cd\u8981\u7684\u4f5c\u7528\u3002\u5927\u591a\u6570\u5927\u8bed\u8a00\u6a21\u578b\u7684\u8bad\u7ec3\u90fd\u662f\u91c7\u53d6\u81ea\u56de\u5f52\u7684\u65b9\u5f0f\u8fdb\u884c\u751f\u6210\uff0c\u867d\u7136\u81ea\u56de\u5f52\u6a21\u578b\u751f\u6210\u7684\u6587\u672c\u8d28\u91cf\u6709\u6240\u4fdd\u8bc1\uff0c\u4f46\u5374\u5bfc\u81f4\u4e86\u9ad8\u6602\u7684\u63a8\u7406\u6210\u672c\u548c\u957f\u65f6\u95f4\u7684\u5ef6\u8fdf\u3002\u7531\u4e8e\u5927\u6a21\u578b\u7684\u53c2\u6570\u91cf\u5de8\u5927\u3001\u63a8\u7406\u6210\u672c\u9ad8\uff0c\u56e0\u6b64\u5982\u4f55\u5728\u5927\u89c4\u6a21\u90e8\u7f72\u5927\u6a21\u578b\u7684\u8fc7\u7a0b\u4e2d\u964d\u4f4e\u6210\u672c\u3001\u51cf\u5c0f\u5ef6\u8fdf\u662f\u4e00\u4e2a\u5173\u952e\u8bfe\u9898\u3002\u9488\u5bf9\u6b64\u95ee\u9898\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u4f7f\u7528\u53c2\u8003\u6587\u672c\u65e0\u635f\u52a0\u901f\u5927\u8bed\u8a00\u6a21\u578b\u63a8\u7406\u7684\u65b9\u6cd5 LLM Accelerator\uff0c\u5728\u5927\u6a21\u578b\u5178\u578b\u7684\u5e94\u7528\u573a\u666f\u4e2d\u53ef\u4ee5\u53d6\u5f97\u4e24\u5230\u4e09\u500d\u7684\u52a0\u901f\u3002 \u968f\u7740\u4eba\u5de5\u667a\u80fd\u6280\u672f\u7684\u5feb\u901f\u53d1\u5c55\uff0cChatGPT\u3001New Bing\u3001GPT-4 \u7b49\u65b0\u4ea7\u54c1\u548c\u65b0\u6280\u672f\u9646\u7eed\u53d1\u5e03\uff0c\u57fa\u7840\u5927\u6a21\u578b\u5728\u8bf8\u591a\u5e94\u7528\u4e2d\u5c06\u53d1\u6325\u65e5\u76ca\u91cd\u8981\u7684\u4f5c\u7528\u3002\u76ee\u524d\u7684\u5927\u8bed\u8a00\u6a21\u578b\u5927\u591a\u662f\u81ea\u56de\u5f52\u6a21\u578b\u3002\u81ea\u56de\u5f52\u662f\u6307\u6a21\u578b\u5728\u8f93\u51fa\u65f6\u5f80\u5f80\u91c7\u7528\u9010\u8bcd\u8f93\u51fa\u7684\u65b9\u5f0f\uff0c\u5373\u5728\u8f93\u51fa\u6bcf\u4e2a\u8bcd\u65f6\uff0c\u6a21\u578b\u9700\u8981\u5c06\u4e4b\u524d\u8f93\u51fa\u7684\u8bcd\u4f5c\u4e3a\u8f93\u5165\u3002\u800c\u8fd9\u79cd\u81ea\u56de\u5f52\u6a21\u5f0f\u901a\u5e38\u5728\u8f93\u51fa\u65f6\u5236\u7ea6\u7740\u5e76\u884c\u52a0\u901f\u5668\u7684\u5145\u5206\u5229\u7528\u3002 \u5728\u8bb8\u591a\u5e94\u7528\u573a\u666f\u4e2d\uff0c\u5927\u6a21\u578b\u7684\u8f93\u51fa\u5e38\u5e38\u4e0e\u4e00\u4e9b\u53c2\u8003\u6587\u672c\u6709\u5f88\u5927\u7684\u76f8\u4f3c\u6027\uff0c\u4f8b\u5982\u5728\u4ee5\u4e0b\u4e09\u4e2a\u5e38\u89c1\u7684\u573a\u666f\u4e2d\uff1a 1. \u68c0\u7d22\u589e\u5f3a\u7684\u751f\u6210\u3002New Bing \u7b49\u68c0\u7d22\u5e94\u7528\u5728\u54cd\u5e94\u7528\u6237\u8f93\u5165\u7684\u5185\u5bb9\u65f6\uff0c\u4f1a\u5148\u8fd4\u56de\u4e00\u4e9b\u4e0e\u7528\u6237\u8f93\u5165\u76f8\u5173\u7684\u4fe1\u606f\uff0c\u7136\u540e\u7528\u8bed\u8a00\u6a21\u578b\u603b\u7ed3\u68c0\u7d22\u51fa\u7684\u4fe1\u606f\uff0c\u518d\u56de\u7b54\u7528\u6237\u8f93\u5165\u7684\u5185\u5bb9\u3002\u5728\u8fd9\u79cd\u573a\u666f\u4e2d\uff0c\u6a21\u578b\u7684\u8f93\u51fa\u5f80\u5f80\u5305\u542b\u5927\u91cf\u68c0\u7d22\u7ed3\u679c\u4e2d\u7684\u6587\u672c\u7247\u6bb5\u3002 2. \u4f7f\u7528\u7f13\u5b58\u7684\u751f\u6210\u3002\u5927\u89c4\u6a21\u90e8\u7f72\u8bed\u8a00\u6a21\u578b\u7684\u8fc7\u7a0b\u4e2d\uff0c\u5386\u53f2\u7684\u8f93\u5165\u8f93\u51fa\u4f1a\u88ab\u7f13\u5b58\u3002\u5728\u5904\u7406\u65b0\u7684\u8f93\u5165\u65f6\uff0c\u68c0\u7d22\u5e94\u7528\u4f1a\u5728\u7f13\u5b58\u4e2d\u5bfb\u627e\u76f8\u4f3c\u7684\u8f93\u5165\u3002\u56e0\u6b64\uff0c\u6a21\u578b\u7684\u8f93\u51fa\u5f80\u5f80\u548c\u7f13\u5b58\u4e2d\u5bf9\u5e94\u7684\u8f93\u51fa\u6709\u5f88\u5927\u7684\u76f8\u4f3c\u6027\u3002 3. \u591a\u8f6e\u5bf9\u8bdd\u4e2d\u7684\u751f\u6210\u3002\u5728\u4f7f\u7528 ChatGPT \u7b49\u5e94\u7528\u65f6\uff0c\u7528\u6237\u5f80\u5f80\u4f1a\u6839\u636e\u6a21\u578b\u7684\u8f93\u51fa\u53cd\u590d\u63d0\u51fa\u4fee\u6539\u8981\u6c42\u3002\u5728\u8fd9\u79cd\u591a\u8f6e\u5bf9\u8bdd\u7684\u573a\u666f\u4e0b\uff0c\u6a21\u578b\u7684\u591a\u6b21\u8f93\u51fa\u5f80\u5f80\u53ea\u6709\u5c11\u91cf\u7684\u53d8\u5316\uff0c\u91cd\u590d\u5ea6\u8f83\u9ad8\u3002 \u57fa\u4e8e\u4ee5\u4e0a\u89c2\u5bdf\uff0c\u7814\u7a76\u5458\u4eec\u4ee5\u53c2\u8003\u6587\u672c\u4e0e\u6a21\u578b\u8f93\u51fa\u7684\u91cd\u590d\u6027\u4f5c\u4e3a\u7a81\u7834\u81ea\u56de\u5f52\u74f6\u9888\u7684\u7740\u529b\u70b9\uff0c\u5e0c\u671b\u53ef\u4ee5\u63d0\u9ad8\u5e76\u884c\u52a0\u901f\u5668\u5229\u7528\u7387\uff0c\u52a0\u901f\u5927\u8bed\u8a00\u6a21\u578b\u63a8\u7406\uff0c\u8fdb\u800c\u63d0\u51fa\u4e86\u4e00\u79cd\u5229\u7528\u8f93\u51fa\u4e0e\u53c2\u8003\u6587\u672c\u7684\u91cd\u590d\u6027\u6765\u5b9e\u73b0\u4e00\u6b65\u8f93\u51fa\u591a\u4e2a\u8bcd\u7684\u65b9\u6cd5 LLM Accelerator\u3002 \u5177\u4f53\u6765\u8bf4\uff0c\u5728\u6bcf\u4e00\u6b65\u89e3\u7801\u65f6\uff0c\u8ba9\u6a21\u578b\u5148\u5339\u914d\u5df2\u6709\u7684\u8f93\u51fa\u7ed3\u679c\u4e0e\u53c2\u8003\u6587\u672c\uff0c\u5982\u679c\u53d1\u73b0\u67d0\u4e2a\u53c2\u8003\u6587\u672c\u4e0e\u5df2\u6709\u7684\u8f93\u51fa\u76f8\u7b26\uff0c\u90a3\u4e48\u6a21\u578b\u5f88\u53ef\u80fd\u987a\u5ef6\u5df2\u6709\u7684\u53c2\u8003\u6587\u672c\u7ee7\u7eed\u8f93\u51fa\u3002\u56e0\u6b64\uff0c\u7814\u7a76\u5458\u4eec\u5c06\u53c2\u8003\u6587\u672c\u7684\u540e\u7eed\u8bcd\u4e5f\u4f5c\u4e3a\u8f93\u5165\u52a0\u5165\u5230\u6a21\u578b\u4e2d\uff0c\u4ece\u800c\u4f7f\u5f97\u4e00\u4e2a\u89e3\u7801\u6b65\u9aa4\u53ef\u4ee5\u8f93\u51fa\u591a\u4e2a\u8bcd\u3002 \u4e3a\u4e86\u4fdd\u8bc1\u8f93\u5165\u8f93\u51fa\u51c6\u786e\uff0c\u7814\u7a76\u5458\u4eec\u8fdb\u4e00\u6b65\u5bf9\u6bd4\u4e86\u6a21\u578b\u8f93\u51fa\u7684\u8bcd\u4e0e\u4ece\u53c2\u8003\u6587\u6863\u8f93\u5165\u7684\u8bcd\u3002\u5982\u679c\u4e24\u8005\u4e0d\u4e00\u81f4\uff0c\u90a3\u4e48\u4e0d\u6b63\u786e\u7684\u8f93\u5165\u8f93\u51fa\u7ed3\u679c\u5c06\u88ab\u820d\u5f03\u3002\u4ee5\u4e0a\u65b9\u6cd5\u80fd\u591f\u4fdd\u8bc1\u89e3\u7801\u7ed3\u679c\u4e0e\u57fa\u51c6\u65b9\u6cd5\u5b8c\u5168\u4e00\u81f4\uff0c\u5e76\u53ef\u4ee5\u63d0\u9ad8\u6bcf\u4e2a\u89e3\u7801\u6b65\u9aa4\u7684\u8f93\u51fa\u8bcd\u6570\uff0c\u4ece\u800c\u5b9e\u73b0\u5927\u6a21\u578b\u63a8\u7406\u7684\u65e0\u635f\u52a0\u901f\u3002LLM Accelerator \u65e0\u9700\u989d\u5916\u8f85\u52a9\u6a21\u578b\uff0c\u7b80\u5355\u6613\u7528\uff0c\u53ef\u4ee5\u65b9\u4fbf\u5730\u90e8\u7f72\u5230\u5404\u79cd\u5e94\u7528\u573a\u666f\u4e2d\u3002 \u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2304.04487.pdf \u9879\u76ee\u94fe\u63a5\uff1ahttps:\/\/github.com\/microsoft\/LMOps \u4f7f\u7528 LLM Accelerator\uff0c\u6709\u4e24\u4e2a\u8d85\u53c2\u6570\u9700\u8981\u8c03\u6574\u3002\u4e00\u662f\u89e6\u53d1\u5339\u914d\u673a\u5236\u6240\u9700\u7684\u8f93\u51fa\u4e0e\u53c2\u8003\u6587\u672c\u7684\u5339\u914d\u8bcd\u6570\uff1a\u5339\u914d\u8bcd\u6570\u8d8a\u957f\u5f80\u5f80\u8d8a\u51c6\u786e\uff0c\u53ef\u4ee5\u66f4\u597d\u5730\u4fdd\u8bc1\u4ece\u53c2\u8003\u6587\u672c\u62f7\u8d1d\u7684\u8bcd\u662f\u6b63\u786e\u7684\u8f93\u51fa\uff0c\u51cf\u5c11\u4e0d\u5fc5\u8981\u7684\u89e6\u53d1\u548c\u8ba1\u7b97\uff1b\u66f4\u77ed\u7684\u5339\u914d\uff0c\u89e3\u7801\u6b65\u9aa4\u66f4\u5c11\uff0c\u6f5c\u5728\u52a0\u901f\u66f4\u5feb\u3002\u4e8c\u662f\u6bcf\u6b21\u62f7\u8d1d\u8bcd\u7684\u6570\u91cf\uff1a\u62f7\u8d1d\u8bcd\u6570\u8d8a\u591a\uff0c\u52a0\u901f\u6f5c\u529b\u8d8a\u5927\uff0c\u4f46\u4e5f\u53ef\u80fd\u9020\u6210\u66f4\u591a\u4e0d\u6b63\u786e\u7684\u8f93\u51fa\u88ab\u820d\u5f03\uff0c\u6d6a\u8d39\u8ba1\u7b97\u8d44\u6e90\u3002\u7814\u7a76\u5458\u4eec\u901a\u8fc7\u5b9e\u9a8c\u53d1\u73b0\uff0c\u66f4\u52a0\u6fc0\u8fdb\u7684\u7b56\u7565\uff08\u5339\u914d\u5355\u4e2a\u8bcd\u89e6\u53d1\uff0c\u4e00\u6b21\u62f7\u8d1d15\u523020\u4e2a\u8bcd\uff09\u5f80\u5f80\u80fd\u591f\u53d6\u5f97\u66f4\u597d\u7684\u52a0\u901f\u6bd4\u3002 \u4e3a\u4e86\u9a8c\u8bc1 LLM Accelerator \u7684\u6709\u6548\u6027\uff0c\u7814\u7a76\u5458\u4eec\u5728\u68c0\u7d22\u589e\u5f3a\u548c\u7f13\u5b58\u8f85\u52a9\u751f\u6210\u65b9\u9762\u8fdb\u884c\u4e86\u5b9e\u9a8c\uff0c\u5229\u7528 MS-MARCO \u6bb5\u843d\u68c0\u7d22\u6570\u636e\u96c6\u6784\u9020\u4e86\u5b9e\u9a8c\u6837\u672c\u3002\u5728\u68c0\u7d22\u589e\u5f3a\u5b9e\u9a8c\u4e2d\uff0c\u7814\u7a76\u5458\u4eec\u4f7f\u7528\u68c0\u7d22\u6a21\u578b\u5bf9\u6bcf\u4e2a\u67e5\u8be2\u8fd4\u56de10\u4e2a\u6700\u76f8\u5173\u7684\u6587\u6863\uff0c\u7136\u540e\u62fc\u63a5\u5230\u67e5\u8be2\u540e\u4f5c\u4e3a\u6a21\u578b\u8f93\u5165\uff0c\u5c06\u8fd910\u4e2a\u6587\u6863\u4f5c\u4e3a\u53c2\u8003\u6587\u672c\u3002\u5728\u7f13\u5b58\u8f85\u52a9\u751f\u6210\u5b9e\u9a8c\u4e2d\uff0c\u6bcf\u4e2a\u67e5\u8be2\u751f\u6210\u56db\u4e2a\u76f8\u4f3c\u7684\u67e5\u8be2\uff0c\u7136\u540e\u7528\u6a21\u578b\u8f93\u51fa\u5bf9\u5e94\u7684\u67e5\u8be2\u4f5c\u4e3a\u53c2\u8003\u6587\u672c\u3002 \u7814\u7a76\u5458\u4eec\u4f7f\u7528\u901a\u8fc7 OpenAI \u63a5\u53e3\u5f97\u5230\u7684 Davinci-003 \u6a21\u578b\u7684\u8f93\u51fa\u7ed3\u679c\u4f5c\u4e3a\u76ee\u6807\u8f93\u51fa\uff0c\u4ee5\u83b7\u5f97\u9ad8\u8d28\u91cf\u7684\u8f93\u51fa\u3002\u5f97\u5230\u6240\u9700\u8f93\u5165\u3001\u8f93\u51fa\u548c\u53c2\u8003\u6587\u672c\u540e\uff0c\u7814\u7a76\u5458\u4eec\u5728\u5f00\u6e90\u7684 LLaMA \u8bed\u8a00\u6a21\u578b\u4e0a\u8fdb\u884c\u4e86\u5b9e\u9a8c\u3002\u7531\u4e8e LLaMA \u6a21\u578b\u7684\u8f93\u51fa\u4e0e Davinci-003 \u8f93\u51fa\u4e0d\u4e00\u81f4\uff0c\u6240\u4ee5\u7814\u7a76\u5458\u4eec\u91c7\u7528\u4e86\u76ee\u6807\u5bfc\u5411\u7684\u89e3\u7801\u65b9\u6cd5\u6765\u6d4b\u8bd5\u7406\u60f3\u8f93\u51fa\uff08Davinci-003 \u6a21\u578b\u7ed3\u679c\uff09\u7ed3\u679c\u4e0b\u7684\u52a0\u901f\u6bd4\u3002\u7814\u7a76\u5458\u4eec\u5229\u7528\u7b97\u6cd52\u5f97\u5230\u4e86\u8d2a\u5a6a\u89e3\u7801\u65f6\u751f\u6210\u76ee\u6807\u8f93\u51fa\u6240\u9700\u7684\u89e3\u7801\u6b65\u9aa4\uff0c\u5e76\u5f3a\u5236 LLaMA \u6a21\u578b\u6309\u7167\u5f97\u5230\u7684\u89e3\u7801\u6b65\u9aa4\u8fdb\u884c\u89e3\u7801\u3002 \u5bf9\u4e8e\u53c2\u6570\u91cf\u4e3a 7B \u548c 13B \u7684\u6a21\u578b\uff0c\u7814\u7a76\u5458\u4eec\u5728\u5355\u4e2a [&hellip;]<\/p>\n","protected":false},"author":34512,"featured_media":1072257,"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-1072113","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\/1072113","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\/34512"}],"version-history":[{"count":4,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1072113\/revisions"}],"predecessor-version":[{"id":1079973,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1072113\/revisions\/1079973"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/media\/1072257"}],"wp:attachment":[{"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/media?parent=1072113"}],"wp:term":[{"taxonomy":"msr-research-area","embeddable":true,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/research-area?post=1072113"},{"taxonomy":"msr-locale","embeddable":true,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-locale?post=1072113"},{"taxonomy":"msr-post-option","embeddable":true,"href":"https:\/\/newed.any0.dpdns.org\/en-us\/research\/wp-json\/wp\/v2\/msr-post-option?post=1072113"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}