{"id":7472,"date":"2026-08-30T09:00:21","date_gmt":"2026-08-30T01:00:21","guid":{"rendered":"https:\/\/1ow.top\/index.php\/2026\/08\/30\/ai%e6%9c%ac%e5%9c%b0%e9%83%a8%e7%bd%b2%e4%b8%8d%e5%a6%82%e5%ae%98%e6%96%b9%e7%89%88%e7%9a%84%e5%85%83%e5%87%b6%e6%89%be%e5%88%b0%e4%ba%86%ef%bc%9a734%e4%b8%aa%e4%be%9d%e8%b5%96%e5%8c%85%ef%bc%8c\/"},"modified":"2026-08-30T09:00:21","modified_gmt":"2026-08-30T01:00:21","slug":"ai%e6%9c%ac%e5%9c%b0%e9%83%a8%e7%bd%b2%e4%b8%8d%e5%a6%82%e5%ae%98%e6%96%b9%e7%89%88%e7%9a%84%e5%85%83%e5%87%b6%e6%89%be%e5%88%b0%e4%ba%86%ef%bc%9a734%e4%b8%aa%e4%be%9d%e8%b5%96%e5%8c%85%ef%bc%8c","status":"publish","type":"post","link":"https:\/\/1ow.top\/index.php\/2026\/08\/30\/ai%e6%9c%ac%e5%9c%b0%e9%83%a8%e7%bd%b2%e4%b8%8d%e5%a6%82%e5%ae%98%e6%96%b9%e7%89%88%e7%9a%84%e5%85%83%e5%87%b6%e6%89%be%e5%88%b0%e4%ba%86%ef%bc%9a734%e4%b8%aa%e4%be%9d%e8%b5%96%e5%8c%85%ef%bc%8c\/","title":{"rendered":"AI\u672c\u5730\u90e8\u7f72\u4e0d\u5982\u5b98\u65b9\u7248\u7684\u5143\u51f6\u627e\u5230\u4e86\uff1a734\u4e2a\u4f9d\u8d56\u5305\uff0c\u6bcf\u4e00\u4e2a\u90fd\u53ef\u80fd\u5751"},"content":{"rendered":"<figure class=\"wp-block-image\"><div class='fancybox-wrapper lazyload-container-unload' data-fancybox='post-images' href='https:\/\/1ow.top\/wp-content\/uploads\/2026\/08\/9fa80a7c34502598a43bf3fdb47db4f2.webp'><img class=\"lazyload lazyload-style-1\" src=\"data:image\/svg+xml;base64,PCEtLUFyZ29uTG9hZGluZy0tPgo8c3ZnIHdpZHRoPSIxIiBoZWlnaHQ9IjEiIHhtbG5zPSJodHRwOi8vd3d3LnczLm9yZy8yMDAwL3N2ZyIgc3Ryb2tlPSIjZmZmZmZmMDAiPjxnPjwvZz4KPC9zdmc+\"  decoding=\"async\" data-original=\"https:\/\/1ow.top\/wp-content\/uploads\/2026\/08\/9fa80a7c34502598a43bf3fdb47db4f2.webp\" src=\"data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAADsQAAA7EAZUrDhsAAAANSURBVBhXYzh8+PB\/AAffA0nNPuCLAAAAAElFTkSuQmCC\" alt=\"AI\u672c\u5730\u90e8\u7f72\u4e0d\u5982\u5b98\u65b9\u7248\u7684\u5143\u51f6\u627e\u5230\u4e86\uff1a734\u4e2a\u4f9d\u8d56\u5305\uff0c\u6bcf\u4e00\u4e2a\u90fd\u53ef\u80fd\u5751\"\/><\/div><\/figure>\n<figure class=\"wp-block-image\"><div class='fancybox-wrapper lazyload-container-unload' data-fancybox='post-images' href='https:\/\/1ow.top\/wp-content\/uploads\/2026\/08\/cbe88fd369bd8727e99f1e88caf991a0.webp'><img class=\"lazyload lazyload-style-1\" src=\"data:image\/svg+xml;base64,PCEtLUFyZ29uTG9hZGluZy0tPgo8c3ZnIHdpZHRoPSIxIiBoZWlnaHQ9IjEiIHhtbG5zPSJodHRwOi8vd3d3LnczLm9yZy8yMDAwL3N2ZyIgc3Ryb2tlPSIjZmZmZmZmMDAiPjxnPjwvZz4KPC9zdmc+\"  decoding=\"async\" data-original=\"https:\/\/1ow.top\/wp-content\/uploads\/2026\/08\/cbe88fd369bd8727e99f1e88caf991a0.webp\" src=\"data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAADsQAAA7EAZUrDhsAAAANSURBVBhXYzh8+PB\/AAffA0nNPuCLAAAAAElFTkSuQmCC\" alt=\"AI\u672c\u5730\u90e8\u7f72\u4e0d\u5982\u5b98\u65b9\u7248\u7684\u5143\u51f6\u627e\u5230\u4e86\uff1a734\u4e2a\u4f9d\u8d56\u5305\uff0c\u6bcf\u4e00\u4e2a\u90fd\u53ef\u80fd\u5751\"\/><\/div><\/figure>\n<figure class=\"wp-block-image\"><div class='fancybox-wrapper lazyload-container-unload' data-fancybox='post-images' href='https:\/\/1ow.top\/wp-content\/uploads\/2026\/08\/052b0b3815cfbcad3052bdbb7e8e2155.webp'><img class=\"lazyload lazyload-style-1\" src=\"data:image\/svg+xml;base64,PCEtLUFyZ29uTG9hZGluZy0tPgo8c3ZnIHdpZHRoPSIxIiBoZWlnaHQ9IjEiIHhtbG5zPSJodHRwOi8vd3d3LnczLm9yZy8yMDAwL3N2ZyIgc3Ryb2tlPSIjZmZmZmZmMDAiPjxnPjwvZz4KPC9zdmc+\"  decoding=\"async\" data-original=\"https:\/\/1ow.top\/wp-content\/uploads\/2026\/08\/052b0b3815cfbcad3052bdbb7e8e2155.webp\" src=\"data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAADsQAAA7EAZUrDhsAAAANSURBVBhXYzh8+PB\/AAffA0nNPuCLAAAAAElFTkSuQmCC\" alt=\"AI\u672c\u5730\u90e8\u7f72\u4e0d\u5982\u5b98\u65b9\u7248\u7684\u5143\u51f6\u627e\u5230\u4e86\uff1a734\u4e2a\u4f9d\u8d56\u5305\uff0c\u6bcf\u4e00\u4e2a\u90fd\u53ef\u80fd\u5751\"\/><\/div><\/figure>\n<figure class=\"wp-block-image\"><div class='fancybox-wrapper lazyload-container-unload' data-fancybox='post-images' href='https:\/\/1ow.top\/wp-content\/uploads\/2026\/08\/2a2fbf8fc384d0088d9738a96c3fa8d6.webp'><img class=\"lazyload lazyload-style-1\" src=\"data:image\/svg+xml;base64,PCEtLUFyZ29uTG9hZGluZy0tPgo8c3ZnIHdpZHRoPSIxIiBoZWlnaHQ9IjEiIHhtbG5zPSJodHRwOi8vd3d3LnczLm9yZy8yMDAwL3N2ZyIgc3Ryb2tlPSIjZmZmZmZmMDAiPjxnPjwvZz4KPC9zdmc+\"  decoding=\"async\" data-original=\"https:\/\/1ow.top\/wp-content\/uploads\/2026\/08\/2a2fbf8fc384d0088d9738a96c3fa8d6.webp\" src=\"data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAADsQAAA7EAZUrDhsAAAANSURBVBhXYzh8+PB\/AAffA0nNPuCLAAAAAElFTkSuQmCC\" alt=\"AI\u672c\u5730\u90e8\u7f72\u4e0d\u5982\u5b98\u65b9\u7248\u7684\u5143\u51f6\u627e\u5230\u4e86\uff1a734\u4e2a\u4f9d\u8d56\u5305\uff0c\u6bcf\u4e00\u4e2a\u90fd\u53ef\u80fd\u5751\"\/><\/div><\/figure>\n<p>\u63a8\u7406\u8f6f\u4ef6\u6808\u7684\u5fae\u5c0f\u5dee\u5f02\uff0c\u5c31\u80fd\u6539\u53d8\u8f93\u51fatoken<\/p>\n<p>\u4e38\u8fa3\uff01\u8f9b\u8f9b\u82e6\u82e6\u672c\u5730\u90e8\u7f72\u7684\u5927\u6a21\u578b\uff0c\u600e\u4e48\u5c31\u662f\u6bd4\u5b98\u65b9\u7248\u66f4\u7b28\uff1f\uff1f<\/p>\n<p>\u5373\u4f7f\u662f\u540c\u4e00\u5f20\u663e\u5361\uff0c\u4e00\u6bdb\u4e00\u6837\u7684\u6743\u91cd\uff0c\u63a8\u7406\u8f6f\u4ef6\u6808\u7684\u5fae\u5c0f\u5dee\u5f02\u5c31\u8ba9\u6a21\u578b\u5728\u5173\u952e\u4f4d\u7f6e\u8f93\u51fa\u5b8c\u5168\u4e0d\u540c\u7684token\uff0c\u751a\u81f3\u5bfc\u81f4\u5de5\u5177\u8c03\u7528\u5f7b\u5e95\u5931\u8d25\u3002<\/p>\n<p>\u8fd9\u79cd\u9677\u9631\u5f88\u5bb9\u6613\u89e6\u53d1\uff0c\u4f46\u5f88\u96be\u6392\u67e5\uff0c\u8fd8\u4ee5\u4e3a\u4e09\u4f53\u4eba\u6765\u667a\u5b50\u5c01\u9501\u4e86\u5462\u3002<\/p>\n<p>Level1Techs\u8bba\u575b\u7528\u6237thr3e\u505a\u4e86\u4e00\u7cfb\u5217\u5b9e\u9a8c\uff0c\u7528Qwen3.6-27B\u5728RTX PRO 6000 Blackwell\u663e\u5361\u4e0a\u5b8c\u6210\u4e86\u8d85\u8fc710\u4e07token\u7684\u5168\u91cflogit\u6355\u83b7\u6d4b\u8bd5\u3002<\/p>\n<p>Logits\u662f\u4e3a\u6240\u6709\u5019\u9009token\u8ba1\u7b97\u7684\u4e00\u7ec4\u539f\u59cb\u5206\u6570\uff0c\u518d\u7ecf\u8fc7\u91c7\u6837\u5668\uff08sampler\uff09\u8f93\u51fa\u4e0b\u4e00\u4e2atoken\u3002<\/p>\n<p>\u5173\u952e\u5728\u4e8eLogits\u662f\u7eaf\u7cb9\u7684\u6570\u5b66\u4ea7\u7269\uff0c\u77e9\u9635\u4e58\u6cd5\u3001\u6ce8\u610f\u529b\u8ba1\u7b97\u3001\u6fc0\u6d3b\u51fd\u6570\u5c42\u5c42\u53e0\u52a0\u540e\u7684\u6d6e\u70b9\u6570\u7ed3\u679c\u3002\u5982\u679c\u4e24\u5957\u7cfb\u7edf\u8dd1\u540c\u4e00\u4efd\u6743\u91cd\u548c\u540c\u4e00\u6bb5\u8f93\u5165\uff0c\u7406\u8bba\u4e0a\u5e94\u8be5\u7b97\u51fa\u5b8c\u5168\u76f8\u540c\u7684logits\u3002<\/p>\n<p>\u4f46\u73b0\u5b9e\u4e2d\uff0c\u6d6e\u70b9\u8fd0\u7b97\u7684\u7cbe\u5ea6\u3001\u7d2f\u52a0\u987a\u5e8f\u3001\u786c\u4ef6\u6307\u4ee4\u96c6\u7684\u5dee\u5f02\uff0c\u90fd\u4f1a\u8ba9\u6700\u7ec8\u6570\u503c\u4ea7\u751f\u5fae\u5c0f\u504f\u79fb\u3002<\/p>\n<p>\u5f53\u8fd9\u4e2a\u504f\u79fb\u5927\u5230\u8db3\u4ee5\u6539\u53d8\u6982\u7387\u6700\u9ad8\u7684\u90a3\u4e2atoken\u65f6\uff0c\u6a21\u578b\u5c31\u4f1a\u8868\u73b0\u51fa\u5dee\u5f02\u3002<\/p>\n<p>\u867d\u7136\u4f60\u7684\u672c\u5730\u90e8\u7f72\u5f88\u70c2\uff0c\u4f46\u522b\u4f24\u5fc3\uff0c\u5176\u4ed6\u4eba\u90e8\u7f72\u7684\u5404\u6709\u4e0d\u4e00\u6837\u7684\u70c2\u6cd5\u3002<\/p>\n<p>\u4e00\u4e2a\u5173\u952e\u5728\u63a8\u7406\u6d41\u7a0b\u4e2d\u201c\u6ce8\u610f\u529b\u540e\u7aef\u201d\u8fd9\u4e2a\u73af\u8282\u3002<\/p>\n<p>vLLM\u4e3aQwen3.6-27B\u63d0\u4f9b\u4e86\u4e09\u79cd\u53ef\u9009\u7684\u5168\u6ce8\u610f\u529b\u540e\u7aef\uff1aFlashAttention 2\u3001Flash Inference\u548cTriton Attention\u3002<\/p>\n<p>\u9664\u4e86\u5207\u6362\u8fd9\u4e00\u9879\u914d\u7f6e\uff0c\u5176\u4f59\u786c\u4ef6\u3001\u8f6f\u4ef6\u3001\u6743\u91cd\u3001KV\u7f13\u5b58\u7cbe\u5ea6\u5168\u90e8\u4fdd\u6301\u4e0d\u53d8\u3002<\/p>\n<p>\u5b9e\u9a8c\u4f7f\u7528\u7684\u8f93\u5165\u662f\u4e00\u6bb5\u7ea610\u4e07token\u7684\u771f\u5b9e\u5de5\u4f5c\u573a\u666f\uff0c\u6765\u81ea\u4e00\u4e2a\u5305\u542b\u591a\u6b21\u5de5\u5177\u8c03\u7528\u7684Agent\u5de5\u4f5c\u6d41\u3002<\/p>\n<p>thr3e\u7279\u522b\u5f3a\u8c03\uff0c\u8fd9\u6bb5\u6570\u636e\u4e0d\u51fa\u73b0\u5728\u4efb\u4f55\u516c\u5f00\u57fa\u51c6\u6216\u8bad\u7ec3\u96c6\u4e2d\uff0c\u6ca1\u6709\u4eba\u80fd\u9488\u5bf9\u5b83\u505a\u8fc7benchmark\u4f18\u5316\u6216\u91cf\u5316\u6821\u51c6\u3002<\/p>\n<p>\u6d4b\u8bd5\u65b9\u6cd5\u662f\u6bcf\u969432\u4e2atoken\u91c7\u6837\u4e00\u6b21\u5168\u8bcd\u8868logit\uff0c\u4e8b\u540e\u7528FP64\u7cbe\u5ea6\u8ba1\u7b97KL\u6563\u5ea6\u548cTop-1\u4e00\u81f4\u6027\u3002<\/p>\n<p>\u7ed3\u679c\u89c2\u5bdf\u5230\u201cTop-1\u7ffb\u8f6c\u201d\u73b0\u8c61\uff0c\u4e5f\u5c31\u662f\u5207\u6362\u540e\u7aef\u5c31\u4f1a\u9009\u51fa\u4e0e\u57fa\u7ebf\u4e0d\u540c\u7684\u8d2a\u5fc3\u89e3\u7801token\u3002<\/p>\n<p>\u6bd4\u5982\u5177\u4f53\u8ffd\u8e2a\u4e00\u4e2a\u51fa\u9519\u6848\u4f8b\uff0c\u6a21\u578b\u6267\u884c\u4e00\u4e2a\u5de5\u5177\u8c03\u7528\uff0c\u76ee\u6807\u662fCisco \u8def\u7531\u5668\u4e0a\u7684\u4e00\u4e2a\u63a5\u53e3GigabitEthernet0\/0\/1.201\u3002<\/p>\n<p>FlashAttention 2\u641e\u9519\u4e86\uff0c\u5bfc\u81f4\u63a5\u53e3\u53d8\u6210\u4e86GigabitEthernet0\/1\/4\uff0c\u968f\u540e\u6a21\u578b\u53c8\u5728\u4e24\u6b21\u540e\u7eed\u5de5\u5177\u8c03\u7528\u4e2d\u6267\u884c\u4e86\u9519\u8bef\u7684\u547d\u4ee4\u3002<\/p>\n<p>\u4e3a\u4fdd\u6301\u6570\u5b66\u53ef\u6bd4\u6027\uff0c\u6240\u6709\u540e\u7aef\u5171\u4eab\u540c\u4e00\u4efd\u5f3a\u5236token\u5386\u53f2\uff0c\u7ffb\u8f6c\u53ea\u8bb0\u5f55\u201c\u672c\u6765\u4f1a\u9009\u9519\u201d\u7684\u60c5\u51b5\uff0c\u4e0d\u8ba9\u9519\u8bef\u4f20\u64ad\u3002<\/p>\n<p>\u7ed3\u679c\u524d\u51e0\u5343\u4e2atoken\uff0c\u4e09\u4e2a\u540e\u7aef\u7684\u8f93\u51fa\u5b8c\u5168\u4e00\u81f4\u3002\u4f46\u968f\u7740\u4e0a\u4e0b\u6587\u589e\u957f\uff0c\u5206\u6b67\u5f00\u59cb\u51fa\u73b0\uff0c\u800c\u4e14\u5206\u5e03\u4e0d\u5747\u5300\u3002<\/p>\n<p>thr3e\u540c\u65f6\u505a\u4e86\u540c\u540e\u7aef\u591a\u6b21\u8fd0\u884c\u7684\u91cd\u590d\u6027\u5bf9\u7167\uff0c\u4ec5\u4ec5\u662f\u5207\u6362\u4e86vLLM\u7684\u4e00\u4e2a\u914d\u7f6e\u9879\uff0c\u540c\u4e00\u5757GPU\u3001\u540c\u4e00\u4efd\u64cd\u4f5c\u7cfb\u7edf\u548c\u9a71\u52a8\u3001\u540c\u4e00\u4efd\u6743\u91cd\u3001\u540c\u4e00\u6bb5prompt\uff0c\u6bcf\u4e00\u4e2a\u9690\u85cf\u72b6\u6001\u7684logit\u5728\u591a\u6b21\u8fd0\u884c\u95f4\u90fd\u662f\u9010\u4f4d\u76f8\u540c\u7684\u3002<\/p>\n<p>\u8fd9\u610f\u5473\u7740\u89c2\u5bdf\u5230\u7684\u5206\u6b67\u5b8c\u5168\u6765\u81ea\u4e0d\u540cCUDA\u6838\u51fd\u6570\u5728prefill\u9636\u6bb5\u6267\u884c\u77e9\u9635\u4e58\u6cd5\u548c\u7d2f\u52a0\u65f6\u7684\u6570\u503c\u5dee\u5f02\u3002<\/p>\n<p>\u63a5\u4e0b\u6765\u7684\u5b9e\u9a8c\u8bbe\u7f6e\u6210\u4fdd\u6301\u6743\u91cd\u4e3aBF16\u4e0d\u52a8\uff0c\u6ce8\u610f\u529b\u540e\u7aef\u56fa\u5b9a\u4e3aTriton\uff0c\u4ec5\u6539\u53d8KV\u7f13\u5b58\u7684\u91cf\u5316\u7cbe\u5ea6\uff0c\u5206\u522b\u6d4b\u8bd5BF16\u3001INT8\u548cINT4\u4e09\u79cd\u914d\u7f6e\u3002<\/p>\n<p>\u7ed3\u679cINT4 KV\u7f13\u5b58\u5728\u957f\u4e0a\u4e0b\u6587\u4e2d\u7684Top-1\u7ffb\u8f6c\u7387\u6025\u5267\u6500\u5347\uff0c\u6700\u7ec8\u5bfc\u81f4\u5de5\u5177\u8c03\u7528\u65e0\u6cd5\u6062\u590d\u3002INT8 KV\u7f13\u5b58\u867d\u7136\u4e5f\u51fa\u73b0\u4e86\u7ffb\u8f6c\uff0c\u4f46\u6a21\u578b\u6700\u7ec8\u8bbe\u6cd5\u56de\u5230\u4e86\u6b63\u786e\u8f68\u9053\u3002\u53ea\u6709BF16 KV\u7f13\u5b58\u5168\u7a0b\u4fdd\u6301\u7a33\u5b9a\u3002<\/p>\n<p>thr3e\u8ba9\u7ffb\u8f6c\u540e\u7684\u751f\u6210\u81ea\u7531\u8fd0\u884c\u800c\u975e\u62c9\u56de\u57fa\u7ebf\uff0c\u89c2\u5bdf\u5b9e\u9645\u540e\u679c\u3002<\/p>\n<p>BF16\u6b63\u5e38\u5b8c\u6210\u6240\u6709\u8c03\u7528\uff0cINT8\u5728\u51fa\u9519\u540e\u201c\u6700\u7ec8\u6323\u624e\u7740\u6062\u590d\u4e86\u201d\uff0c\u800cINT4\u5219\u5f7b\u5e95\u504f\u79bb\u4e86\uff0c\u5de5\u5177\u8c03\u7528\u5931\u8d25\u4e14\u65e0\u6cd5\u81ea\u6211\u7ea0\u6b63\u3002<\/p>\n<p>\u8fd9\u4e2a\u95ee\u9898\u4e13\u5751\u90a3\u4e9b\u4e3a\u4e86\u7701\u663e\u5b58\u800c\u5c06KV\u7f13\u5b58\u538b\u5230INT4\u7684\u672c\u5730\u7528\u6237\u3002<\/p>\n<p>\u5728\u77ed\u4e0a\u4e0b\u6587\u4e2d\u4f60\u53ef\u80fd\u611f\u77e5\u4e0d\u5230\u5dee\u5f02\uff0c\u4f46\u4e00\u65e6\u5bf9\u8bdd\u6216agent\u5de5\u4f5c\u6d41\u62c9\u957f\u5230\u51e0\u4e07token\uff0c\u7d2f\u79ef\u7684\u6570\u503c\u6f02\u79fb\u8db3\u4ee5\u8ba9\u6a21\u578b\u505a\u51fa\u81f4\u547d\u9519\u8bef\u51b3\u7b56\u3002<\/p>\n<p>\u6700\u540e\u4e00\u7ec4\u5b9e\u9a8c\u5c06KV\u7f13\u5b58\u7edf\u4e00\u4e3aBF16\uff0c\u8f6c\u800c\u6bd4\u8f83\u4e94\u79cd\u4e0d\u540c\u7684\u6743\u91cd\u91cf\u5316\u65b9\u6848\u3002<\/p>\n<p>\u53c2\u8d5b\u9009\u624b\u5305\u62ec\uff1aQwen\u5b98\u65b9BF16\u57fa\u7ebf\u3001Qwen\u5b98\u65b9FP8\uff08W8A8\uff09\u3001TheHouseOfTheDude\u53d1\u5e03\u7684INT8\uff08W8A16\uff0c\u65e0\u6821\u51c6\u6570\u636e\u96c6\u7684\u4e00\u6b21\u6027\u91cf\u5316\uff09\u3001\u82f1\u4f1f\u8fbe\u5b98\u65b9NVFP4\u3001\u4ee5\u53cacyankiwi\u53d1\u5e03\u7684AWQ INT4\uff08W4A16\uff0c\u4f7f\u7528STEM\u548cAgentic\u6570\u636e\u96c6\u6821\u51c6\uff09\u3002<\/p>\n<p>\u4e94\u79cd\u65b9\u6848\u5404\u81ea\u8c03\u7528\u4e0d\u540c\u7684CUDA\u6838\u51fd\u6570\u5b8c\u6210\u77e9\u9635\u8fd0\u7b97\u3002BF16\u8d70\u6807\u51c6torch\u7ebf\u6027\u5c42\uff0cFP8\u8d70CUTLASS\u7684FP8\u5206\u5757\u7f29\u653e\u6838\uff0cINT8\u548cAWQ\u90fd\u8d70Marlin\u6838\u3002<\/p>\n<p>NVFP4\u5219\u662f\u6df7\u5408\u8def\u5f84\uff0c208\u4e2a\u76ee\u6807\u8d70FlashInfer\u7684FP8\u7f29\u653e\u6838\uff0c193\u4e2aMLP\u6295\u5f71\u8d70Marlin\u7684NvFp4\u6838\u3002<\/p>\n<p>\u5728\u8fd9\u6b21\u6d4b\u8bd5\u7684vLLM nightly\u7248\u672c\u4e2d\uff0cGPU\u8def\u5f84\u88ab\u5224\u5b9a\u4e3a\u4e0d\u652f\u6301\u539f\u751fFP4\u8fd0\u7b97\uff0cNVFP4\u5b9e\u9645\u6267\u884c\u7684\u662f\u901a\u8fc7Marlin\u5185\u6838\u7684\u4ec5\u6743\u91cdFP4\u89e3\u538b\u7f29\uff0c\u800c\u975e\u771f\u6b63\u7684FP4\u8fd0\u7b97\u3002<\/p>\n<p>\u7ed3\u679c\u4e2d\u6700\u4eae\u773c\u7684\u662fTheHouseOfTheDude\u7684INT8\uff08W8A16\uff09\uff0c\u8fd9\u662f\u4e00\u4e2a\u6ca1\u6709\u4f7f\u7528\u4efb\u4f55\u6821\u51c6\u6570\u636e\u96c6\u3001\u4ec5\u505a\u4e86\u901a\u9053\u7ea7\u5bf9\u79f0\u91cf\u5316\u7684\u793e\u533a\u7248\u672c\uff0c\u5176Top-1\u4e00\u81f4\u6027\u78be\u538b\u4e86Qwen\u5b98\u65b9FP8\u548c\u82f1\u4f1f\u8fbe\u5b98\u65b9NVFP4\u3002<\/p>\n<p>\u7ecf\u5206\u6790\uff0c\u8fd9\u5f52\u529f\u4e8eW8A16\u4fdd\u7559\u4e86BF16\u6fc0\u6d3b\u7cbe\u5ea6\uff0c\u52a0\u4e0a\u8be5\u91cf\u5316\u6392\u9664\u4e86Gated DeltaNet\u6295\u5f71\u548clm_head\u5c42\u3002<\/p>\n<p>\u82f1\u4f1f\u8fbeNVFP4\u5728\u8fd9\u7ec4\u6d4b\u8bd5\u6c47\u603b\u8868\u73b0\u6700\u5dee\uff0c\u523088k\u4e0a\u4e0b\u6587\u65f6Top-1\u7ffb\u8f6c\u7387\u903c\u8fd150%\uff0c\u76f8\u5f53\u4e8e\u6709\u4e00\u534a\u7684\u4f4d\u7f6e\u6a21\u578b\u4f1a\u9009\u51fa\u4e0d\u540c\u7684token\u3002<\/p>\n<p>\u5728\u5b9e\u9645\u5de5\u5177\u8c03\u7528\u6d4b\u8bd5\u4e2d\uff0cNVFP4\u548cAWQ W4A16\u90fd\u672a\u80fd\u6b63\u786e\u5173\u95ed\u5de5\u5177\u8c03\u7528\uff0c\u5e76\u4e14\u641e\u9519\u4e86Cisco\u547d\u4ee4\u884c\u8bed\u6cd5\uff08\u6267\u884c\u4e86\u201dshow run\u201d\u800c\u975e\u6b63\u786e\u7684\u201dshow arp\u201d\uff09\uff0c\u800cFP8\u548cINT8\u5747\u987a\u5229\u5b8c\u6210\u3002<\/p>\n<p>thr3e\u8fd8\u5c55\u793a\u4e86\u5f20\u91cf\u5e76\u884c\u7684\u8be1\u5f02\u73b0\u8c61\u3002\u540c\u4e00\u4efdBF16\u6743\u91cd\uff0cTP1\u5355\u5361\u80fd\u6b63\u786e\u5b8c\u6210\u5de5\u5177\u8c03\u7528\uff0c\u5207\u5230TP2\u53cc\u5361\u53cd\u800c\u5931\u8d25\uff0c\u518d\u5207\u5230TP4\u56db\u5361\u53c8\u6210\u529f\u4e86\u3002<\/p>\n<p>\u7ecf\u8fc7\u8fdb\u4e00\u6b65\u8c03\u8bd5\u548cNCCL\u901a\u4fe1\u56fe\u6355\u83b7\uff0c\u8fd9\u901a\u5e38\u662fNCCL\u8de8\u5361\u5f52\u7ea6\u64cd\u4f5c\u4e2d\u7684\u6570\u503c\u5dee\u5f02\u5bfc\u81f4\u7684\u3002<\/p>\n<p>thr3e\u8bf4\u4ed6\u968f\u624b\u4e0b\u8f7d\u7684vLLM nightly\u5bb9\u5668\u955c\u50cf\u91cc\u5305\u542b734\u4e2a\u8f6f\u4ef6\u5305\uff0c\u5176\u4e2d252\u4e2a\u662fPython\u7684uv\/pip\u5305\u3002<\/p>\n<p>\u8fd9734\u4e2a\u4ee3\u7801\u5e93\u5404\u6709\u5404\u7684bug\u548c\u672a\u8bb0\u5f55\u7684\u884c\u4e3a\u7279\u6027\uff0c\u4f60\u7684\u7279\u5b9a\u786c\u4ef6\u548c\u6a21\u578b\u914d\u7f6e\u5728\u8fd9\u5ea7\u4ee3\u7801\u5c71\u4e2d\u8d70\u51fa\u7684\u8def\u5f84\u662f\u72ec\u4e00\u65e0\u4e8c\u7684\u3002<\/p>\n<p>\u8fd9\u4e5f\u662f\u4e3a\u4ec0\u4e48HuggingFace\u6a21\u578b\u5361\u4e0a\u6807\u7684\u6781\u4f4e\u7684KL\u6563\u5ea6\u6570\u5b57\u4e0d\u80fd\u8f7b\u4fe1\u3002<\/p>\n<p>\u9664\u975e\u4f5c\u8005\u5b8c\u6574\u62ab\u9732\u4e86\u53c2\u8003\u68c0\u67e5\u70b9\u3001\u5b8c\u6574\u8fd0\u884c\u65f6\u73af\u5883\u3001\u8bc4\u4f30\u6587\u672c\u3001\u6821\u51c6\u6570\u636e\u3001\u4e0a\u4e0b\u6587\u957f\u5ea6\u3001\u91c7\u6837\u4f4d\u7f6e\u3001KL\u65b9\u5411\u3001\u8bcd\u8868\u622a\u65ad\u65b9\u5f0f\u4ee5\u53ca\u805a\u5408\u65b9\u6cd5\uff0c\u5426\u5219\u90a3\u4e2a\u6570\u5b57\u6839\u672c\u65e0\u6cd5\u89e3\u8bfb\u3002<\/p>\n<p>thr3e\u76ee\u524d\u5df2\u7ecf\u5b8c\u6210\u4e86\u4e0d\u540c\u6743\u91cd\u3001\u4e0d\u540c\u6a21\u578b\uff08\u5305\u62ecQwen3.6 vs 3.8\u7684\u8de8\u6a21\u578b\u5bf9\u6bd4\uff09\u3001\u4e0d\u540cKV\u7f13\u5b58\u91cf\u5316\u3001\u4e0d\u540c\u5f20\u91cf\u5e76\u884c\u5ea6\u3001\u4e0d\u540cNCCL\u914d\u7f6e\u3001\u4e0d\u540c\u6ce8\u610f\u529b\u540e\u7aef\u3001\u4e0d\u540c\u663e\u5361\uff08RTX PRO 6000 vs RTX 5090\uff0c\u540c\u5c5eSM120\u67b6\u6784\uff09\u7b49\u7ef4\u5ea6\u7684\u5168\u91cflogit\u6355\u83b7\u4e0e\u5206\u53c9\u8ffd\u8e2a\uff0c\u6b63\u5728\u5c06\u6d4b\u8bd5\u5de5\u5177\u548c\u6570\u636e\u96c6\u6253\u5305\u6210\u53ef\u5206\u53d1\u7248\u672c\uff0c\u4f9b\u5176\u4ed6\u7528\u6237\u5728\u81ea\u5df1\u7684\u8bbe\u5907\u4e0a\u8fd0\u884c\u5e76\u4e0a\u62a5\u7ed3\u679c\u3002<\/p>\n<p>\u5982\u679c\u4f60\u542c\u8bf4\u67d0\u4e2a\u6a21\u578b\u70b8\u88c2\u9707\u64bc\u65e0\u654c\u4e86\uff0c\u4e0b\u8f7d\u5230\u672c\u5730\u5374\u89c9\u5f97\u5b83\u8822\uff0c\u53ef\u80fd\u662f\u4f60\u7684\u63a8\u7406\u6808\u91cc\u4ece\u6ce8\u610f\u529b\u6838\u51fd\u6570\u3001\u5230KV\u7f13\u5b58\u7cbe\u5ea6\u3001\u5230\u6743\u91cd\u91cf\u5316\u65b9\u6848\u3001\u5230\u591a\u5361\u901a\u4fe1\u534f\u8bae\uff0c\u6bcf\u4e00\u5c42\u90fd\u5728\u5236\u9020\u4e0e\u539f\u59cb\u57fa\u51c6\u4e0d\u540c\u7684\u6570\u5b66\u7ed3\u679c\u3002<\/p>\n<p>\u800c\u8fd9\u4e9b\u5dee\u5f02\u5728\u957f\u4e0a\u4e0b\u6587\u4e2d\u4f1a\u50cf\u6eda\u96ea\u7403\u4e00\u6837\u7d2f\u79ef\uff0c\u76f4\u5230\u6a21\u578b\u5728\u5173\u952e\u65f6\u523b\u505a\u51fa\u5b8c\u5168\u9519\u8bef\u7684\u51b3\u5b9a\u3002<\/p>\n<p>\u53c2\u8003\u94fe\u63a5\uff1a[1]https:\/\/forum.level1techs.com\/t\/why-your-local-llm-feels-dumber-than-it-is\/253917\/4<\/p>\n<blockquote>\n<p><strong>\u7248\u6743\u58f0\u660e\uff1a<\/strong><\/p>\n<p>\u672c\u535a\u5ba2\u90e8\u5206\u5185\u5bb9\u4e3a\u8f6c\u8f7d\u6587\u7ae0\uff0c\u65e8\u5728\u5206\u4eab\u6709\u4ef7\u503c\u7684\u4fe1\u606f\uff0c\u7248\u6743\u5f52\u539f\u4f5c\u8005\u6240\u6709\u3002<\/p>\n<p>\u8f6c\u8f7d\u4ec5\u4e3a\u4e2a\u4eba\u5b66\u4e60\u4e0e\u4ea4\u6d41\u76ee\u7684\uff0c\u4e0d\u5bf9\u6587\u7ae0\u89c2\u70b9\u8d1f\u8d23\uff0c\u4ea6\u4e0d\u7528\u4e8e\u4efb\u4f55\u5546\u4e1a\u7528\u9014\u3002<\/p>\n<p>\u5982\u6d89\u53ca\u7248\u6743\u95ee\u9898\uff0c\u8bf7\u8054\u7cfb\u672c\u4eba\u5220\u9664\u3002<\/p>\n<\/blockquote>\n","protected":false},"excerpt":{"rendered":"<p>\u63a8\u7406\u8f6f\u4ef6\u6808\u7684\u5fae\u5c0f\u5dee\u5f02\uff0c\u5c31\u80fd\u6539\u53d8\u8f93\u51fatoken<\/p>\n<p>\u4e38\u8fa3\uff01\u8f9b\u8f9b\u82e6\u82e6\u672c\u5730\u90e8\u7f72\u7684\u5927\u6a21\u578b\uff0c\u600e\u4e48\u5c31\u662f\u6bd4\u5b98\u65b9\u7248\u66f4\u7b28\uff1f\uff1f<\/p>\n<p>\u5373\u4f7f\u662f\u540c\u4e00\u5f20\u663e\u5361\uff0c\u4e00\u6bdb\u4e00\u6837\u7684\u6743\u91cd\uff0c\u63a8\u7406\u8f6f\u4ef6\u6808\u7684\u5fae\u5c0f\u5dee\u5f02\u5c31\u8ba9\u6a21\u578b\u5728\u5173\u952e\u4f4d\u7f6e\u8f93\u51fa\u5b8c\u5168\u4e0d\u540c\u7684token\uff0c\u751a\u81f3\u5bfc\u81f4\u5de5\u5177\u8c03\u7528\u5f7b\u5e95\u5931\u8d25\u3002<\/p>\n<p>\u8fd9\u79cd\u9677\u9631\u5f88\u5bb9\u6613\u89e6\u53d1\uff0c\u4f46\u5f88\u96be\u6392\u67e5\uff0c\u8fd8\u4ee5\u4e3a\u4e09\u4f53\u4eba\u6765\u667a\u5b50\u5c01\u9501\u4e86\u5462\u3002<\/p>\n<p>Level1Techs\u8bba\u575b\u7528\u6237thr3e\u505a\u4e86\u4e00\u7cfb\u5217\u5b9e\u9a8c\uff0c\u7528Qwen3.6-27B\u5728RTX PRO 6<\/p>\n","protected":false},"author":1,"featured_media":7468,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[2],"class_list":["post-7472","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-xwzx","tag-ai"],"_links":{"self":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/posts\/7472","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/comments?post=7472"}],"version-history":[{"count":0,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/posts\/7472\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/media\/7468"}],"wp:attachment":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/media?parent=7472"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/categories?post=7472"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/tags?post=7472"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}