{"id":5099,"date":"2026-04-28T16:20:29","date_gmt":"2026-04-28T08:20:29","guid":{"rendered":"https:\/\/www.ihub1.com\/?p=5099"},"modified":"2026-04-28T16:22:43","modified_gmt":"2026-04-28T08:22:43","slug":"deepseek%e4%b8%8d%e6%83%9c%e4%bb%a3%e4%bb%b7%e4%bf%9d%e4%bd%8f%e5%ae%83%ef%bc%81v4%e5%85%b3%e9%94%ae%e7%89%b9%e6%80%a7%e8%a2%ab%e6%8c%96%e5%87%ba%e6%9d%a5%e4%ba%86","status":"publish","type":"post","link":"https:\/\/www.ihub1.com\/index.php\/2026\/04\/28\/deepseek%e4%b8%8d%e6%83%9c%e4%bb%a3%e4%bb%b7%e4%bf%9d%e4%bd%8f%e5%ae%83%ef%bc%81v4%e5%85%b3%e9%94%ae%e7%89%b9%e6%80%a7%e8%a2%ab%e6%8c%96%e5%87%ba%e6%9d%a5%e4%ba%86\/","title":{"rendered":"DeepSeek\u4e0d\u60dc\u4ee3\u4ef7\u4fdd\u4f4f\u5b83\uff01V4\u5173\u952e\u7279\u6027\u88ab\u6316\u51fa\u6765\u4e86"},"content":{"rendered":"\t\t<div data-elementor-type=\"wp-post\" data-elementor-id=\"5099\" class=\"elementor elementor-5099\">\n\t\t\t\t<div class=\"elementor-element elementor-element-854f81e e-flex e-con-boxed e-con e-parent\" data-id=\"854f81e\" data-element_type=\"container\" data-e-type=\"container\">\n\t\t\t\t\t<div class=\"e-con-inner\">\n\t\t\t\t<div class=\"elementor-element elementor-element-5623808 elementor-widget elementor-widget-text-editor\" data-id=\"5623808\" data-element_type=\"widget\" data-e-type=\"widget\" data-widget_type=\"text-editor.default\">\n\t\t\t\t<div class=\"elementor-widget-container\">\n\t\t\t\t\t\t\t\t\t<p>DeepSeek V4\u6280\u672f\u62a5\u544a\uff0c\u8fd8\u5728\u88ab\u6df1\u6316\u3002<\/p><p>\u5e38\u770b\u5e38\u65b0\u7684\u90a3\u79cd\u2014\u2014\u4e00\u4e2a\u5f15\u8d77\u5173\u6ce8\u7684\u65b0\u53d1\u73b0\u662f\uff0cV4\u5728\u5de5\u7a0b\u4e0a\u4e3a\u4e86\u4fdd\u7559\u6838\u5fc3\u8bbe\u8ba1\u300c<strong>batch invariance<\/strong>\u300d\uff0c\u90fd\u6709\u70b9\u4e0d\u8ba1\u4ee3\u4ef7\u4e86\u3002<\/p><p>\u600e\u4e48\u8bf4\uff1f<\/p><p>DeepSeek V4\u540c\u65f6\u505a\u5230\u4e86\u300c\u8d85\u957f\u4e0a\u4e0b\u6587\u300d\u300c\u590d\u6742\u540e\u8bad\u7ec3\/\u63a8\u7406\u7ba1\u7ebf\u300d\u300c\u81ea\u7814\u9ad8\u6027\u80fdkernel\u6808\u300d\u8fd9\u51e0\u4ef6\u5f88\u5bb9\u6613\u6253\u67b6\u7684\u4e8b\uff0c\u800c\u80cc\u540e\u7684\u5173\u952e\uff0c\u6b63\u662fbatch invariance\uff08\u6279\u6b21\u4e0d\u53d8\u6027\uff09\u3002<\/p><p>\u4f46batch invariance\u5e76\u975e\u6ca1\u6709\u4ee3\u4ef7\uff0c\u751a\u81f3\u4ee3\u4ef7\u8fd8\u633a\u5927\uff1aGPU\u5229\u7528\u7387\u3001\u63a8\u7406\u901f\u5ea6\u4e0b\u964d\uff0c\u5de5\u7a0b\u590d\u6742\u5ea6\u8fd8\u66f4\u9ad8\u4e86\u2026\u2026<\/p><p>\u6240\u4ee5\uff0c\u4e3a\u4ec0\u4e48DeepSeek V4\u4f1a\u5bf9\u8fd9\u4e00\u8bbe\u8ba1\u5982\u6b64\u6267\u7740\uff1f\u6211\u4eec\u5177\u4f53\u6765\u62c6\u6790\u4e00\u4e0b\u3002<\/p><p>\u4ec0\u4e48\u662fbatch invariance<\/p><p>\u5148\u6765\u770b\u770bbatch invariance\u7684\u5b9a\u4e49\uff1a<\/p><p>\u5bf9\u4e8e\u540c\u4e00\u4e2atoken\uff0c\u65e0\u8bba\u5b83\u5728\u6279\u6b21\u91cc\u6392\u7b2c\u51e0\u3001\u65e0\u8bba\u6279\u6b21\u591a\u5927\u3001\u65e0\u8bba\u548c\u8c01\u4e00\u8d77\u6279\u5904\u7406\uff0c\u8f93\u51fa\u90fd\u80fd\u4fdd\u6301\u9010\u6bd4\u7279\u5b8c\u5168\u4e00\u81f4\u3002<\/p><p>\u8bba\u6587\u63d0\u5230\uff0c\u5176\u6838\u5fc3\u8bbe\u8ba1\u76ee\u7684\uff0c\u662f\u786e\u4fdd\u9884\u8bad\u7ec3\u3001\u540e\u8bad\u7ec3\u548c\u63a8\u7406\u5168\u6d41\u7a0b\u7684\u53ef\u590d\u73b0\u6027\uff0c\u4fdd\u8bc1\u5404\u4e2a\u73af\u8282\u4e4b\u95f4\u7684\u5bf9\u9f50\u3002<\/p><p>\u8fd9\u6837\u505a\u7684\u597d\u5904\u662f\uff0c\u9996\u5148\uff0c\u80fd<strong>\u4fdd\u8bc1\u7ebf\u4e0a\u63a8\u7406\u7ed3\u679c\u7a33\u5b9a<\/strong>\u3002<\/p><p>\u7ebf\u4e0a\u670d\u52a1\u4f1a\u52a8\u6001batching\uff0c\u540c\u4e00\u4e2a\u7528\u6237\u8bf7\u6c42\uff0c\u4eca\u5929\u53ef\u80fd\u548cA\u3001B\u8bf7\u6c42\u62fc\u5728\u4e00\u8d77\uff0c\u660e\u5929\u53ef\u80fd\u548cC\u3001D\u8bf7\u6c42\u62fc\u5728\u4e00\u8d77\u3002<\/p><p>\u5982\u679c\u6ca1\u6709batch invariance\uff0c\u540c\u6837\u7684\u63d0\u793a\u8bcd\u5c31\u53ef\u80fd\u56e0\u4e3abatch\u7ec4\u5408\u4e0d\u540c\u3001\u5e95\u5c42kernel\u5f52\u7ea6\u987a\u5e8f\u4e0d\u540c\u7b49\u56e0\u7d20\uff0c\u88ab\u653e\u5927\u6210\u5b8c\u5168\u4e0d\u540c\u7684\u7b54\u6848\u3002<\/p><p>\u4e5f\u5c31\u662f\u8bf4\uff0cbatch invariance\u80fd\u8ba9<strong>\u540c\u4e00\u4e2a\u8f93\u5165\uff0c\u5c3d\u91cf\u5f97\u5230\u4e25\u683c\u4e00\u81f4\u7684\u8f93\u51fa<\/strong>\u3002<\/p><p><img fetchpriority=\"high\" decoding=\"async\" src=\"https:\/\/q2.itc.cn\/q_70\/images03\/20260428\/8c15815a3d0245399aaef9c2919d945b.png\" width=\"640\" height=\"640\" \/><\/p><p>\u5176\u6b21\uff0c<strong>\u4fdd\u8bc1\u9884\u8bad\u7ec3\u3001\u540e\u8bad\u7ec3\u3001\u63a8\u7406\u4e4b\u95f4\u7684\u5bf9\u9f50<\/strong>\u3002<\/p><p>DeepSeek V4\u6709\u9884\u8bad\u7ec3\u3001SFT\u3001RL\u3001on-policy distillation\u3001\u63a8\u7406\u670d\u52a1\u7b49\u591a\u6761\u94fe\u8def\u3002<\/p><p>\u8fd9\u5c31\u5bfc\u81f4\u4e86\u4e00\u4e2a\u95ee\u9898\uff1a\u6a21\u578b\u884c\u4e3a\u53d8\u5316\u5230\u5e95\u662f\u6765\u81ea\u6570\u636e\u3001RL\u3001\u84b8\u998f\u3001\u91cf\u5316\uff0c\u8fd8\u662f\u6765\u81eabatch shape\/ kernel\u8def\u5f84\u53d8\u5316\uff1f<\/p><p>\u6709\u4e86batch invariance\uff0c\u5de5\u7a0b\u56e2\u961f\u66f4\u5bb9\u6613\u5224\u65ad\uff0c\u662f\u4e0d\u662fbatch\u7ec4\u7ec7\u65b9\u5f0f\u6539\u53d8\u4e86\u6570\u503c\u7ed3\u679c\u3002<\/p><p>\u5f53\u95ee\u9898\u80fd\u591f\u88ab\u51c6\u786e\u5b9a\u4f4d\uff0c\u5f02\u5e38\u4e5f\u5c31\u66f4\u5bb9\u6613\u590d\u73b0\uff0c\u63d0\u9ad8\u4e86\u53ef\u590d\u73b0\u6027\u548c\u53ef\u8c03\u8bd5\u6027\u3002<\/p><p>\u53e6\u5916\uff0c<strong>batch invariance\u4e5f\u662f\u590d\u6742\u4e0a\u4e0b\u6587\u7cfb\u7edf\u7684\u5e95\u5ea7\u4e4b\u4e00<\/strong>\u3002<\/p><p>V4\u540c\u65f6\u7528\u4e86\u957f\u4e0a\u4e0b\u6587attention\u3001\u538b\u7f29KV\u3001\u7a00\u758f\u6ce8\u610f\u529b\u3001MoE\u3001FP4\/FP8\u3001Muon\u3001mHC\u3001\u81ea\u7814 kernel\u7b49\u5f88\u591a\u590d\u6742\u7ec4\u4ef6\u3002<\/p><p>\u7ec4\u4ef6\u8d8a\u591a\uff0c\u6570\u503c\u4e0d\u786e\u5b9a\u6027\u7684\u6765\u6e90\u8d8a\u591a\u3002batch invariance\u76f8\u5f53\u4e8e\u7ed9\u5e95\u5c42\u6267\u884c\u7cfb\u7edf\u52a0\u4e86\u4e00\u6761\u786c\u7ea6\u675f\uff1a\u53ef\u4ee5\u4f18\u5316\u6027\u80fd\uff0c\u4f46\u4e0d\u80fd\u56e0\u4e3abatch\u53d8\u4e86\uff0c\u5c31\u8ba9\u540c\u4e00\u4e2atoken\u7684\u7ed3\u679c\u53d8\u4e86\u3002<\/p><p><strong>batch invariance<\/strong>\u8fd8\u80fd\u8ba9\u540e\u8bad\u7ec3\u66f4\u7a33\u5b9a\u3002<\/p><p>RL\u3001\u84b8\u998f\u3001\u957f\u94fe\u63a8\u7406\u5bf9\u7ec6\u5fae\u5dee\u5f02\u5f88\u654f\u611f\u3002\u4e00\u70b9\u70b9\u6570\u503c\u5dee\u5f02\uff0c\u53ef\u80fd\u6539\u53d8\u91c7\u6837\u8def\u5f84\uff1b\u91c7\u6837\u8def\u5f84\u4e00\u53d8\uff0creward\u3001teacher-student\u5bf9\u9f50\u3001\u8bad\u7ec3\u4fe1\u53f7\u90fd\u4f1a\u53d8\u3002<\/p><p>batch invariance\u80fd\u591f\u51cf\u5c11\u968f\u673a\u6570\u503c\u6270\u52a8\uff0c\u8ba9\u6a21\u578b\u884c\u4e3a\u66f4\u53ef\u63a7\u3002<\/p><p>\u603b\u7ed3\u4e00\u4e0b\uff0c\u5c31\u662fbatch invariance\u662fDeepSeek V4\u7684\u5e95\u5c42\u5de5\u7a0b\u7a33\u5b9a\u5668\uff0c\u53ef\u4ee5\u5728\u5728\u6781\u590d\u6742\u7684\u957f\u4e0a\u4e0b\u6587\u8bad\u7ec3\u3001\u540e\u8bad\u7ec3\u548c\u63a8\u7406\u7cfb\u7edf\u91cc\uff0c\u4fdd\u8bc1\u540c\u4e00\u8f93\u5165\u7684\u6570\u503c\u884c\u4e3a\u4e0d\u88abbatch\u7ec4\u7ec7\u3001kernel\u8c03\u5ea6\u548c\u5f52\u7ea6\u987a\u5e8f\u6c61\u67d3\uff0c\u5b9e\u73b0\u53ef\u590d\u73b0\u3001\u53ef\u8c03\u8bd5\u3001\u53ef\u5bf9\u9f50\u3001\u53ef\u7a33\u5b9a\u90e8\u7f72\u7684\u5de5\u7a0b\u786e\u5b9a\u6027\u3002<\/p><p>\u727a\u7272\u4e86\u4ec0\u4e48<\/p><p>\u8fd9\u6837\u770b\u6765\uff0cbatch invariance\u786e\u5b9e\u5f88\u91cd\u8981\u3002<\/p><p>\u4f46\uff0c\u5c31\u50cf\u524d\u9762\u8bf4\u7684\uff0c\u4ee3\u4ef7\u4e0d\u5c11\u3002<\/p><p>\u4e3a\u4e86batch invariance\uff0cV4\u4e0d\u80fd\u968f\u4fbf\u4f7f\u7528\u4e00\u4e9b\u5e38\u89c1\u6027\u80fd\u4f18\u5316\u4e86\uff0c\u6bd4\u5982split-KV\u3001split-K\u3002<\/p><p>\u5728attention\u91cc\uff0csplit-KV\u4f1a\u628a\u5355\u6761\u5e8f\u5217\u7684\u6ce8\u610f\u529b\u8ba1\u7b97\u5206\u644a\u5230\u591a\u4e2aSM\u4e0a\uff0c\u4ee5\u63d0\u9ad8\u8d1f\u8f7d\u5747\u8861\u548cGPU\u5229\u7528\u7387\u3002\u4f46\u8fd9\u79cd\u505a\u6cd5\u4f1a\u6539\u53d8\u5e76\u884c\u5f52\u7ea6\u8def\u5f84\uff0c\u96be\u4ee5\u4fdd\u8bc1\u540c\u4e00\u4e2atoken\u5728\u4e0d\u540cbatch\u7ec4\u7ec7\u65b9\u5f0f\u4e0b\u7684\u9010\u6bd4\u7279\u4e00\u81f4\u6027\u3002<\/p><p>\u5728GEMM\u91cc\uff0csplit-K\u7684\u505a\u6cd5\u662f\u628a\u77e9\u9635\u4e58\u6cd5\u7684\u5f52\u7ea6\u7ef4\u5ea6K\u5207\u5f00\u5e76\u884c\u8ba1\u7b97\u3002\u591a\u8def\u5e76\u884c\u6c42\u548c\u4e4b\u540e\u8fd8\u8981\u518d\u5f52\u7ea6\uff0c\u800c\u6d6e\u70b9\u52a0\u6cd5\u7684\u5f52\u7ea6\u987a\u5e8f\u4e00\u53d8\uff0c\u6700\u7ec8\u7ed3\u679c\u7684bi \u5c31\u53ef\u80fd\u4e0d\u540c\uff0c\u56e0\u6b64\u4e5f\u548cbatch invariance\u5b58\u5728\u51b2\u7a81\u3002<\/p><p>\u4e3a\u6b64\uff0cDeepSeek\u5728attention\u4fa7\u63d0\u51fa\u4e86\u00a0dual-kernel\uff1a\u4e3a\u540c\u4e00\u4e2a\u6ce8\u610f\u529b\u89e3\u7801\u4efb\u52a1\u51c6\u5907\u4e24\u5957\u8ba1\u7b97\u7a0b\u5e8f\uff0c\u5206\u522b\u5904\u7406\u201cGPU\u5403\u5f97\u6ee1\u201d\u548c\u201cGPU\u5403\u4e0d\u6ee1\u201d\u7684\u60c5\u51b5\uff0c\u540c\u65f6\u4fdd\u8bc1\u4e24\u5957\u7a0b\u5e8f\u7b97\u51fa\u6765\u7684\u7ed3\u679c\u9010\u6bd4\u7279\u4e00\u81f4\u3002<\/p><p>\u77e9\u9635\u4e58\u6cd5\u65b9\u9762\uff0cV4\u5728\u5927\u591a\u6570\u573a\u666f\u4e2d\u653e\u5f03split-K\uff0c\u8f6c\u800c\u505a\u66f4\u53d7\u7ea6\u675f\u7684 batch-invariant GEMM\u3002\u4ed6\u4eec\u7528\u81ea\u7814\u00a0DeepGEMM\u66ff\u4ee3\u4e86\u901a\u7528\u7684cuBLAS\u3002<\/p><p>\u8fd9\u4e9b\u90fd\u5bfc\u81f4\u4e86\u5de5\u7a0b\u590d\u6742\u5ea6\u7684\u660e\u663e\u4e0a\u5347\uff1a\u5f88\u591a\u539f\u672c\u53ef\u4ee5\u4ea4\u7ed9\u901a\u7528\u5e93\u6216\u5e38\u89c4\u4f18\u5316\u7b56\u7565\u7684\u5de5\u4f5c\uff0c\u90fd\u5fc5\u987b\u7531\u81ea\u7814kernel\u548c\u66f4\u4e25\u683c\u7684\u8ba1\u7b97\u8def\u5f84\u6765\u627f\u62c5\u3002<\/p><p>\u5982\u6b64\u79cd\u79cd\uff0c\u7b80\u5355\u603b\u7ed3\u8d77\u6765\u610f\u5473\u7740DeepSeek V4\u5728\u4ee5\u4e0b\u51e0\u4e2a\u65b9\u9762\u505a\u4e86\u727a\u7272\uff1a<\/p><ul><li><p>GPU \u5229\u7528\u7387\uff08\u6ce2\u524d\u91cf\u5316\u95ee\u9898\uff09<\/p><\/li><li><p>\u5c0f\u6279\u91cf\/\u77ed\u5e8f\u5217\u901f\u5ea6<\/p><\/li><li><p>\u539f\u751f\u7b97\u5b50\u517c\u5bb9\u6027<\/p><\/li><li><p>\u90e8\u5206\u7a00\u758f\u52a0\u901f\u7684\u81ea\u7531\u5ea6<\/p><\/li><\/ul><p>\u4ee5\u6362\u53d6\uff1a<\/p><ul><li><p>\u8bad\u7ec3\/\u63a8\u7406\/RL\u4e09\u9636\u6bb5\u9010\u6bd4\u7279\u53ef\u590d\u73b0<\/p><\/li><li><p>\u957f\u4e0a\u4e0b\u6587\u3001Agent\u3001RL\u8bad\u7ec3\u7a33\u5b9a\u5ea6<\/p><\/li><li><p>\u591a\u673a\u591a\u5361\u8dd1\u51fa\u6765\u7684\u7ed3\u679c\u5b8c\u5168\u5bf9\u9f50<\/p><\/li><\/ul><p>One More Thing<\/p><p>DeepSeek V4\u53d1\u5e03\u4e86\u8fd9\u4e48\u4e9b\u5929\uff0c\u4f46\u8fd9\u4e2a\u6280\u672f\u62a5\u544a\u5c5e\u5b9e\u662f\u8d8a\u6316\u8d8a\u6709\u3002<\/p><p>\u770b\u70b9\u8fd8\u5305\u62ec\uff0c\u628a10\u4e2a\u4ee5\u4e0a\u4e13\u5bb6\u6559\u5e08\u6a21\u578b\u84b8\u998f\u6210\u4e00\u4e2a\u5b66\u751f\u6a21\u578b\u7b49\u7b49\u3002<\/p><p>\u5e76\u4e14\u6bcf\u4e2a\u95ee\u9898\u80cc\u540e\u90fd\u6709\u575a\u5b9e\u7684\u6570\u5b66\u89e3\u91ca\u3002<\/p><p>Hugging Face\u7684Transformers\u8d1f\u8d23\u4ebaArthur Zucker\u5c31\u611f\u6168\uff1a<\/p><blockquote><p>\u628a\u6570\u6708\u4e43\u81f3\u6570\u5e74\u7684\u52aa\u529b\u5168\u90e8\u514d\u8d39\u516c\u5f00\uff0c\u8ba9\u4efb\u4f55\u4eba\u90fd\u80fd\u53d7\u76ca\uff0c\u8fd9\u662f\u771f\u6b63\u7684GOAT\uff08Greatest of All Times\uff09\u3002<\/p><\/blockquote><p><img decoding=\"async\" src=\"https:\/\/q5.itc.cn\/q_70\/images03\/20260428\/962d8b46d5814e21a3464d1d26a98324.png\" width=\"640\" height=\"328\" \/><\/p><p>\u00a0<\/p>\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t","protected":false},"excerpt":{"rendered":"<p>\u9c7c\u7f8a \u53d1\u81ea \u51f9\u975e\u5bfa<\/p>\n","protected":false},"author":1,"featured_media":5101,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-5099","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-uncategorized"],"aioseo_notices":[],"_links":{"self":[{"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/posts\/5099","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/comments?post=5099"}],"version-history":[{"count":7,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/posts\/5099\/revisions"}],"predecessor-version":[{"id":5107,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/posts\/5099\/revisions\/5107"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/media\/5101"}],"wp:attachment":[{"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/media?parent=5099"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/categories?post=5099"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.ihub1.com\/index.php\/wp-json\/wp\/v2\/tags?post=5099"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}