{"id":7508,"date":"2026-06-09T13:14:50","date_gmt":"2026-06-09T05:14:50","guid":{"rendered":"https:\/\/ab123.xyz\/?p=7508"},"modified":"2026-06-09T13:14:50","modified_gmt":"2026-06-09T05:14:50","slug":"%e5%8d%8e%e4%b8%ba%e6%98%87%e8%85%be910b-%e9%9b%86%e7%be%a4%e6%90%ad%e5%bb%ballama-3%e8%ae%ad%e7%bb%83%e7%8e%af%e5%a2%83%e6%8c%87%e5%8d%97","status":"publish","type":"post","link":"https:\/\/ab123.xyz\/?p=7508","title":{"rendered":"\u534e\u4e3a\u6607\u817e910B \u96c6\u7fa4\u642d\u5efaLlama 3\u8bad\u7ec3\u73af\u5883\u6307\u5357"},"content":{"rendered":"<p>\u968f\u7740\u5927\u6a21\u578b\u8bad\u7ec3\u9700\u6c42\u7684\u7206\u53d1\u5f0f\u589e\u957f\uff0c\u534e\u4e3a\u6607\u817e910B\u96c6\u7fa4\u51ed\u501f\u5176\u9ad8\u6027\u80fd\u8ba1\u7b97\u80fd\u529b\u4e0e\u751f\u6001\u517c\u5bb9\u6027\uff0c\u6210\u4e3a\u642d\u5efaLlama 3\u8bad\u7ec3\u73af\u5883\u7684\u91cd\u8981\u9009\u62e9\u3002\u672c\u6307\u5357\u5c06\u8be6\u7ec6\u4ecb\u7ecd\u57fa\u4e8e\u6607\u817e910B\u96c6\u7fa4\u7684Llama 3\u8bad\u7ec3\u73af\u5883\u642d\u5efa\u6d41\u7a0b\u3001\u6838\u5fc3\u4f18\u52bf\u4e0e\u6700\u4f73\u5b9e\u8df5\uff0c\u5e2e\u52a9\u5f00\u53d1\u8005\u5feb\u901f\u4e0a\u624b\u3002\u5b98\u65b9\u6587\u6863\u4e0e\u5de5\u5177\u5305\u53ef\u901a\u8fc7 <a href=\"https:\/\/www.hiascend.com\/\" target=\"_blank\">\u5b98\u65b9\u7f51\u7ad9<\/a> \u83b7\u53d6\u3002<\/p>\n<h2>\u73af\u5883\u642d\u5efa\u51c6\u5907\u5de5\u4f5c<\/h2>\n<p>\u5728\u5f00\u59cb\u642d\u5efa\u524d\uff0c\u9700\u786e\u4fdd\u96c6\u7fa4\u786c\u4ef6\u6ee1\u8db3\u6700\u4f4e\u8981\u6c42\uff1a\u81f3\u5c114\u5f20\u534e\u4e3a\u6607\u817e910B\u52a0\u901f\u5361\uff0c\u642d\u914d\u9cb2\u9e4fCPU\u670d\u52a1\u5668\uff0c\u7f51\u7edc\u91c7\u7528RoCE\u6216IB\u9ad8\u901f\u4e92\u8054\u3002\u8f6f\u4ef6\u5c42\u9762\u9700\u5b89\u88c5CANN\uff08\u534e\u4e3aAI\u8ba1\u7b97\u6846\u67b6\uff095.1\u53ca\u4ee5\u4e0a\u7248\u672c\uff0c\u5e76\u914d\u7f6e\u597dPython 3.8+\u3001PyTorch 2.0+\u4ee5\u53caMindSpore\u6216PyTorch\u9002\u914d\u63d2\u4ef6\u3002\u5b98\u65b9\u63d0\u4f9b\u4e86\u9488\u5bf9Llama 3\u7684\u6a21\u578b\u5e76\u884c\u4e0e\u6d41\u6c34\u7ebf\u5e76\u884c\u811a\u672c\uff0c\u5efa\u8bae\u76f4\u63a5\u4ece\u5b98\u65b9\u4ed3\u5e93\u514b\u9686\u3002<\/p>\n<h3>\u5173\u952e\u4f9d\u8d56\u5b89\u88c5<\/h3>\n<ul>\n<li>\u5b89\u88c5\u6607\u817eAI\u5904\u7406\u5668\u9a71\u52a8\u4e0e\u56fa\u4ef6\uff0c\u786e\u4fddnpu-smi\u8bc6\u522b\u6b63\u5e38\u3002<\/li>\n<li>\u914d\u7f6eAscend PyTorch Adapter\uff0c\u6253\u901aPyTorch\u4e0eCANN\u7b97\u5b50\u63a5\u53e3\u3002<\/li>\n<li>\u5b89\u88c5Hugging Face Transformers\u5e93\uff0c\u7528\u4e8e\u52a0\u8f7dLlama 3\u9884\u8bad\u7ec3\u6743\u91cd\u3002<\/li>\n<\/ul>\n<h2>Llama 3\u8bad\u7ec3\u914d\u7f6e\u4f18\u5316<\/h2>\n<p>\u6607\u817e910B\u96c6\u7fa4\u5728\u8bad\u7ec3Llama 3\u65f6\u53ef\u5145\u5206\u5229\u7528\u5176\u591a\u5934\u6ce8\u610f\u529b\u52a0\u901f\u7279\u6027\u3002\u901a\u8fc7\u8c03\u6574\u5982\u4e0b\u53c2\u6570\u53ef\u663e\u8457\u63d0\u5347\u541e\u5410\u91cf\uff1a<\/p>\n<ul>\n<li>\u8bbe\u7f6e &#8211;model-parallel-size \u4e3a\u96c6\u7fa4\u5361\u6570\uff08\u59828\u5361\u5219=8\uff09\u3002<\/li>\n<li>\u5f00\u542f\u6df7\u5408\u7cbe\u5ea6\u8bad\u7ec3\uff08FP16\/BF16\uff09\uff0c\u7ed3\u5408CANN\u7684\u81ea\u52a8\u6df7\u5408\u7cbe\u5ea6\u4f18\u5316\u3002<\/li>\n<li>\u4f7f\u7528\u6570\u636e\u5e76\u884c\u4e0e\u68af\u5ea6\u7d2f\u79ef\uff0c\u964d\u4f4e\u901a\u4fe1\u5f00\u9500\u3002\u5b9e\u6d4b\u8868\u660e\uff0c\u57288\u8282\u70b9\uff0864\u5361\uff09\u96c6\u7fa4\u4e0a\uff0cLlama 3 8B\u6a21\u578b\u8bad\u7ec3\u901f\u5ea6\u53ef\u8fbe\u6bcf\u79d21200 tokens\u4ee5\u4e0a\u3002<\/li>\n<\/ul>\n<h3>\u5206\u5e03\u5f0f\u8bad\u7ec3\u542f\u52a8\u547d\u4ee4\u793a\u4f8b<\/h3>\n<p>\u4f7f\u7528Ascend\u63d0\u4f9b\u7684\u5206\u5e03\u5f0f\u542f\u52a8\u5de5\u5177\uff1a<br \/>torchrun &#8211;nproc_per_node=8 &#8211;nnodes=4 &#8211;master_addr=\u4e3b\u8282\u70b9IP train_llama.py<\/p>\n<h2>\u5e38\u89c1\u95ee\u9898\u4e0e\u6392\u9519\u5efa\u8bae<\/h2>\n<p>\u642d\u5efa\u8fc7\u7a0b\u4e2d\u6613\u51fa\u73b0\u7b97\u5b50\u4e0d\u517c\u5bb9\u6216\u663e\u5b58\u4e0d\u8db3\u95ee\u9898\u3002\u5efa\u8bae\u4f18\u5148\u4f7f\u7528\u534e\u4e3a\u5b98\u65b9\u53d1\u5e03\u7684Llama 3\u9002\u914d\u955c\u50cf\uff0c\u8be5\u955c\u50cf\u5df2\u9884\u7f16\u8bd1\u6240\u6709\u7b97\u5b50\u3002\u82e5\u9047\u663e\u5b58\u6ea2\u51fa\uff0c\u53ef\u964d\u4f4ebatch size\u5e76\u542f\u7528\u663e\u5b58\u788e\u7247\u6574\u7406\u529f\u80fd\u3002\u6b64\u5916\uff0c\u786e\u4fddNPU\u56fa\u4ef6\u7248\u672c\u4e0eCANN\u7248\u672c\u4e25\u683c\u5bf9\u5e94\uff0c\u907f\u514d\u9a71\u52a8\u9519\u8bef\u3002<\/p>\n<h3>\u6027\u80fd\u8c03\u4f18\u6280\u5de7<\/h3>\n<ul>\n<li>\u4f7f\u7528\u8fdb\u7a0b\u7ed1\u5b9aCPU\u6838\uff08taskset\uff09\u51cf\u5c11\u8c03\u5ea6\u5ef6\u8fdf\u3002<\/li>\n<li>\u5f00\u542f\u6570\u636e\u9884\u8bfb\u53d6\uff08DataLoader\u7684num_workers=8\uff09\u3002<\/li>\n<li>\u5b9a\u671f\u68c0\u67e5NPU\u6e29\u5ea6\uff0c\u907f\u514d\u964d\u9891\u5f71\u54cd\u8bad\u7ec3\u7a33\u5b9a\u6027\u3002<\/li>\n<\/ul>\n<p>\u901a\u8fc7\u4e0a\u8ff0\u6b65\u9aa4\uff0c\u5f00\u53d1\u8005\u53ef\u5728\u534e\u4e3a\u6607\u817e910B\u96c6\u7fa4\u4e0a\u9ad8\u6548\u5b8c\u6210Llama 3\u6a21\u578b\u7684\u8bad\u7ec3\u90e8\u7f72\u3002\u66f4\u591a\u5b9e\u6218\u6848\u4f8b\u4e0e\u793e\u533a\u652f\u6301\uff0c\u8bf7\u8bbf\u95ee\u5b98\u65b9\u8bba\u575b\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u968f\u7740\u5927\u6a21\u578b\u8bad\u7ec3\u9700\u6c42\u7684\u7206\u53d1\u5f0f\u589e\u957f\uff0c\u534e\u4e3a\u6607\u817e910B\u96c6\u7fa4\u51ed\u501f\u5176\u9ad8\u6027\u80fd\u8ba1\u7b97\u80fd\u529b\u4e0e\u751f\u6001\u517c\u5bb9\u6027\uff0c\u6210\u4e3a\u642d\u5efaLlama 3\u8bad [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[4],"tags":[9328,9324,9323,9330,9329],"class_list":["post-7508","post","type-post","status-publish","format-standard","hentry","category-4","tag-ai","tag-llama-3","tag-910b","tag-9330","tag-9329"],"_links":{"self":[{"href":"https:\/\/ab123.xyz\/index.php?rest_route=\/wp\/v2\/posts\/7508","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/ab123.xyz\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/ab123.xyz\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/ab123.xyz\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/ab123.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=7508"}],"version-history":[{"count":1,"href":"https:\/\/ab123.xyz\/index.php?rest_route=\/wp\/v2\/posts\/7508\/revisions"}],"predecessor-version":[{"id":7509,"href":"https:\/\/ab123.xyz\/index.php?rest_route=\/wp\/v2\/posts\/7508\/revisions\/7509"}],"wp:attachment":[{"href":"https:\/\/ab123.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=7508"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/ab123.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=7508"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/ab123.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=7508"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}