{"id":772596,"date":"2026-05-03T12:58:47","date_gmt":"2026-05-03T04:58:47","guid":{"rendered":"https:\/\/www.alphaseek.cn\/?p=772596"},"modified":"2026-05-03T12:58:47","modified_gmt":"2026-05-03T04:58:47","slug":"deepseek-v4%e6%9c%80%e5%a4%a7%e7%9a%84%e9%81%97%e6%86%be","status":"publish","type":"post","link":"https:\/\/www.alphaseek.cn\/?p=772596","title":{"rendered":"DeepSeek V4\u6700\u5927\u7684\u9057\u61be"},"content":{"rendered":"

henry \u53d1\u81ea \u51f9\u975e\u5bfa<\/p>\n

\u91cf\u5b50\u4f4d | \u516c\u4f17\u53f7 QbitAI<\/p>\n

DeepSeekV4\u7684\u6280\u672f\u62a5\u544a\u91cc\u6709mHC\uff0c\u6709CSA\uff0c\u6709HCA\uff0c\u6709Muon\uff0c\u6709FP4\u2026\u2026<\/p>\n

\u552f\u72ec\u6ca1\u6709Engram<\/strong>\u3002<\/p>\n

Engram\u53bb\u54ea\u4e86\uff1f<\/p>\n

\u8fd9\u4e2a\u8bdd\u9898\u4e00\u5ea6\u6210\u4e3a\u7f51\u53cb\u4eec\u8ba8\u8bba\u7684\u70ed\u70b9\u3002<\/p>\n

Engram\u5728\u4eca\u5e741\u6708\u7531DeepSeek\u548c\u5317\u5927\u8054\u5408\u5f00\u6e90\uff0c\u4e3b\u8981\u7814\u7a76\u5927\u6a21\u578b\u7684\u8bb0\u5fc6\u4e0e\u6548\u7387\u95ee\u9898\u3002<\/p>\n

\u81ea\u6302\u4e0aarXiv\u7684\u90a3\u4e00\u523b\u8d77\uff0c\u5708\u5b50\u91cc\u56f4\u7ed5\u5b83\u7684\u63a2\u8ba8\u5c31\u6ca1\u6709\u505c\u6b62\u8fc7\u2026<\/p>\n

\"DeepSeek<\/p>\n

\u4e0d\u4ec5\u4ec5\u56e0\u4e3a\u5b83\u662fV4\u7684\u524d\u594f\uff0c\u800c\u662f\u6709\u4e86Engram\uff0c\u300c\u4f26\u6566\u662f\u82f1\u56fd\u9996\u90fd\u300d\u8fd9\u79cd\u4e8b\u5b9e\uff0c\u6a21\u578b\u4e0d\u7528\u52a8\u7528\u6574\u4e2a\u6df1\u5c42\u7f51\u7edc\u53bb\u91cd\u65b0\u63a8\u4e00\u904d\uff0c\u76f4\u63a5\u67e5\u5c31\u884c\u3002<\/p>\n

\u4e0d\u4ec5\u7701\u663e\u5b58\uff0c\u8fd8\u80fd\u91ca\u653e\u6df1\u5c42\u7f51\u7edc\u5bb9\u91cf\uff0c\u7528\u4e8e\u66f4\u9ad8\u9636\u7684\u63a8\u7406\u3002<\/strong><\/p>\n

\u6b63\u56e0\u5982\u6b64\uff0c\u81ea1\u6708\u521d\u8bba\u6587\u53d1\u8868\u4ee5\u6765\uff0c\u6240\u6709\u4eba\u90fd\u89c9\u5f97\uff0cEngram\u5c31\u662fV4\u7684\u67b6\u6784\u5730\u57fa\uff0c\u6240\u6709\u4eba\u90fd\u5728\u76fc\u3002<\/p>\n

\u4ee5\u81f3\u4e8eV4\u53d1\u8868\u540e\uff0c\u5927\u5bb6\u7b2c\u4e00\u65f6\u95f4\u5c31\u662fcommand+f\u53bb\u8bba\u6587\u91cc\u627eEngram\uff0c\u53ef\u60dc\u5e76\u6ca1\u6709\u3002<\/p>\n

\"DeepSeek<\/p>\n

\u4ee5\u81f3\u4e8e\u4e0d\u5c11\u7f51\u53cb\u751a\u81f3\u89c9\u5f97\uff0c\u6ca1\u6709Engram\uff0cV4\u5c31\u662f\u4e0d\u5b8c\u6574\u7684\u3002<\/p>\n

\"DeepSeek<\/p>\n

\u6ca1\u6709Engram\uff0c\u53ef\u80fd\u662fDeepSeekV4\u6700\u5927\u7684\u9057\u61be\u3002<\/p>\n

\u4e0d\u8fc7\uff0cEngram\u5e76\u6ca1\u6709\u6d88\u5931\u3002\u968f\u540e\u4e09\u7bc7\u503c\u5f97\u6ce8\u610f\u7684\u8bba\u6587\u63a5\u529b\u51fa\u73b0\uff1a<\/p>\n

CXL\u5185\u5b58\u6c60\u5316\u7248\u672c\uff1a\u628aEngram\u653e\u8fdb\u591a\u673a\u5171\u4eab\u7684CXL\u5185\u5b58\u6c60\uff0c\u89e3\u51b3\u5927\u6a21\u578b\u591a\u673a\u90e8\u7f72\u7684\u5b58\u50a8\u95ee\u9898\u3002<\/strong><\/p>\n

\u65e0\u51b2\u7a81\u70ed\u5c42\u5b9e\u9a8c\uff1a\u5bf9Engram\u7684\u591a\u5934\u54c8\u5e0c\u4f18\u5316\u8fdb\u884c\u4e86\u5b9e\u8bc1\u68c0\u9a8c\uff0c\u8bc1\u4f2a\u4e86\u4e00\u4e9b\u76f4\u89c9\u5f0f\u6539\u8fdb\u65b9\u6848\u3002<\/strong><\/p>\n

\u89c6\u89c9Tiny Engram\uff1aAutoArk\u56e2\u961f\u628a\u6587\u672cEngram\u642c\u5230\u89c6\u89c9\u6a21\u6001\uff0c\u6269\u5c55\u4e86\u5b83\u7684\u5e94\u7528\u8fb9\u754c\u3002<\/strong><\/p>\n

\u6240\u4ee5\uff0c\u867d\u7136V4\u6ca1\u6709Engram\uff0c\u4f46\u5b83\u7684\u7406\u5ff5\u3001\u63a2\u7d22\u548c\u540e\u7eed\u5e94\u7528\u5df2\u7ecf\u6084\u7136\u94fa\u5f00\uff0c\u4e3a\u4e0b\u4e00\u4ee3\u6a21\u578b\u6253\u4e0b\u57fa\u7840\u3002<\/p>\n

Engram\u5230\u5e95\u662f\u4ec0\u4e48<\/strong><\/p>\n

\u628a\u65f6\u95f4\u5012\u56de2026\u5e741\u670812\u65e5\u3002<\/p>\n

\u90a3\u4e00\u5929\uff0cDeepSeek\u8054\u5408\u5317\u5927\u653e\u51fa\u4e86\u4e00\u7bc733\u9875\u7684\u8bba\u6587\u300aConditionalMemoryviaScalableLookup\u300b\u3002\u7b2c\u4e00\u4f5c\u8005ChengXin\uff0c\u5317\u5927\u535a\u58eb\u751f\uff0c\u66fe\u7ecf\u7f72\u540d\u8fc7V3\u3002\u6700\u540e\u4e00\u4f4d\u4f5c\u8005\uff0c\u6881\u6587\u950b\u3002<\/p>\n

\"DeepSeek<\/p>\n

\u5148\u6765\u4e00\u53e5\u8bdd\u901f\u901a\u7248\uff0cEngram\u662f\u7ed9Transformer\u52a0\u7684\u4e00\u4e2a\u539f\u751f\u77e5\u8bc6\u67e5\u8868\u6a21\u5757<\/strong>\u3002\u80fd\u67e5\u7684\u522b\u7b97\uff0c\u5148\u67e5\u4e00\u4e0b\u3002<\/p>\n

\u56e2\u961f\u7684\u6838\u5fc3\u89c2\u5bdf\u662f\uff0c\u8bed\u8a00\u5efa\u6a21\u5176\u5b9e\u5305\u542b\u4e24\u79cd\u6027\u8d28\u5b8c\u5168\u4e0d\u540c\u7684\u4efb\u52a1\uff0c\u4e00\u79cd\u662f\u9700\u8981\u6df1\u5ea6\u52a8\u6001\u8ba1\u7b97\u7684\u7ec4\u5408\u63a8\u7406\uff0c\u53e6\u4e00\u79cd\u662f\u68c0\u7d22\u9759\u6001\u77e5\u8bc6\u3002<\/p>\n

\u4e4b\u524d\u7684\u95ee\u9898\u5728\u4e8e\uff0cTransformer\u628a\u8fd9\u4e24\u4ef6\u4e8b\u6df7\u5728\u4e00\u8d77\u505a\u3002\u6a21\u578b\u8bc6\u522b\u4e00\u4e2a\u5b9e\u4f53\u65f6\uff0c\u5f97\u6d88\u8017\u597d\u51e0\u5c42\u6ce8\u610f\u529b\u548c\u524d\u9988\u7f51\u7edc\u9010\u5c42\u62fc\u51d1\u7279\u5f81\u3002<\/p>\n

\u8bba\u6587\u91cc\u4e3e\u4e86\u4e2a\u4f8b\u5b50\uff0c\u300cDiana\uff0cPrincess of Wales\u300d\u3002\u6a21\u578b\u8981\u8d706\u5c42\u624d\u80fd\u628a\u8fd9\u4e2a\u8bc6\u522b\u5b8c\u3002<\/p>\n

\u524d\u51e0\u5c42\u8fd8\u5728\u7ea0\u7ed3\u300cWales\u662f\u82f1\u56fd\u7684\u4e00\u4e2a\u5730\u533a\u300d\u3001\u300cPrincess of Wales\u662f\u67d0\u79cd\u5934\u8854\u300d\u8fd9\u4e9b\u4e2d\u95f4\u72b6\u6001\uff0c\u6700\u540e\u4e00\u5c42\u624d\u53cd\u5e94\u8fc7\u6765\u8fd9\u662f\u6234\u5b89\u5a1c\u738b\u5983\u3002<\/p>\n

\u8fd9\u79cd\u300c\u7528\u6602\u8d35\u7684\u8fd0\u884c\u65f6\u8ba1\u7b97\u91cd\u5efa\u4e00\u4e2a\u9759\u6001\u67e5\u627e\u8868\u300d\u7684\u6d3b\uff0c\u672c\u6765\u53ef\u4ee5\u8ba9\u6df1\u5c42\u7f51\u7edc\u53bb\u5e72\u66f4\u9ad8\u9636\u7684\u63a8\u7406\u3002<\/p>\n

\u5bf9\u6b64\uff0cEngram\u7684\u601d\u8def\u76f8\u5f53\u76f4\u63a5\uff0c\u65e2\u7136\u7ecf\u5178\u7684N-gram\u6a21\u578b\u5c31\u80fd\u7528O(1)\u7684\u65f6\u95f4\u590d\u6742\u5ea6\u6355\u83b7\u8fd9\u4e9b\u5c40\u90e8\u4f9d\u8d56\uff0c\u90a3\u5e72\u8106\u628a\u8fd9\u80fd\u529b\u76f4\u63a5\u5d4c\u8fdbTransformer\u3002<\/p>\n

\u6253\u4e2a\u6bd4\u65b9\uff0c\u5c31\u50cf\u4f60\u505a\u6570\u5b66\u9898\uff0c\u8be5\u7528\u7684\u516c\u5f0f\u4e0d\u5fc5\u6bcf\u6b21\u4ece\u5934\u63a8\u4e00\u904d\uff0c\u7ffb\u8868\u4ee3\u8fdb\u53bb\u5c31\u884c\u3002Transformer\u4e4b\u524d\u6ca1\u8fd9\u5f20\u8868\uff0c\u53ea\u80fd\u6bcf\u9053\u9898\u90fd\u4ece\u516c\u7406\u8d70\u8d77\u3002Engram\u7b49\u4e8e\u628a\u8fd9\u5f20\u8868\u4ea4\u5230\u6a21\u578b\u624b\u91cc\u3002<\/p>\n

\u5177\u4f53\u505a\u6cd5\u662f\uff0c\u5728Transformer\u7684\u7b2c2\u5c42\u548c\u7b2c15\u5c42\u4e4b\u95f4\u5404\u63d2\u5165\u4e00\u4e2aEngram\u6a21\u5757\u3002<\/p>\n

\u6bcf\u4e2a\u4f4d\u7f6e\u7684\u8f93\u5165\u4f1a\u89e6\u53d1\u4e00\u6b21\u54c8\u5e0c\u67e5\u627e\uff0c\u628a\u5f53\u524dtoken\u548c\u524d\u9762\u51e0\u4e2atoken\u7ec4\u6210\u7684N-gram\u6620\u5c04\u5230\u4e00\u4e2a\u5de8\u5927\u7684\u5d4c\u5165\u8868\u91cc\uff0c\u76f4\u63a5\u53d6\u51fa\u5bf9\u5e94\u7684\u5411\u91cf\u3002<\/p>\n

\u95e8\u63a7\u673a\u5236\u4fdd\u8bc1\u67e5\u5230\u7684\u5185\u5bb9\u8ddf\u5f53\u524d\u4e0a\u4e0b\u6587\u4e0d\u5339\u914d\u65f6\u81ea\u52a8\u5c4f\u853d\u3002\u6bd4\u5982\u300c\u5f20\u300d\u662f\u4e2a\u5e38\u89c1\u59d3\u6c0f\uff0c\u4f46\u300c\u5f20\u4ef2\u666f\u300d\u4e09\u4e2a\u5b57\u51d1\u4e00\u8d77\u5c31\u662f\u56fa\u5b9a\u5386\u53f2\u4eba\u7269\u5b9e\u4f53\u4e86\uff0c\u95e8\u63a7\u5c31\u8d1f\u8d23\u8ba4\u51fa\u8fd9\u79cd\u533a\u522b\u3002<\/p>\n

Engram\u7684\u5b9a\u4f4d\u662fMoE\u4e4b\u5916\u7684\u53e6\u4e00\u6761\u7a00\u758f\u8f74\u3002MoE\u662f\u628a\u8ba1\u7b97\u7a00\u758f\u5316\uff0c\u53ea\u6fc0\u6d3b\u4e00\u90e8\u5206\u4e13\u5bb6\u3002Engram\u662f\u628a\u5b58\u50a8\u7a00\u758f\u5316\uff0c\u53ea\u67e5\u4e00\u90e8\u5206\u6761\u76ee\u3002\u4e24\u8005\u4e92\u8865\uff0c\u4e0d\u51b2\u7a81\u3002<\/p>\n

\"DeepSeek<\/p>\n

\u8bba\u6587\u6700\u6838\u5fc3\u7684\u4e00\u6bb5\u5b9e\u9a8c\uff0c\u662f\u56fa\u5b9a\u603b\u53c2\u6570\u548c\u6bcftoken\u6fc0\u6d3b\u53c2\u6570\uff0c\u7136\u540e\u8ba9MoE\u4e13\u5bb6\u548cEngram\u8bb0\u5fc6\u62a2\u9884\u7b97\uff0c\u5f97\u5230\u4e00\u6761U\u5f62\u66f2\u7ebf\u3002<\/p>\n

\"DeepSeek<\/p>\n

\u7eafMoE\u4e0d\u662f\u6700\u4f18\u89e3\u3002\u628a\u5927\u7ea620%-25%\u7684\u7a00\u758f\u53c2\u6570\u5206\u7ed9Engram\uff0c\u6a21\u578bloss\u8fbe\u5230\u6700\u4f4e\u70b9\u3002<\/p>\n

\u6309\u8fd9\u4e2a\u66f2\u7ebf\u6307\u5bfc\uff0c\u56e2\u961f\u628aEngram\u6269\u523027B\u9a8c\u8bc1\u3002\u6fc0\u6d3b\u53c2\u65703.8B\uff0c\u8bad\u7ec3262B tokens\uff0c\u4e25\u683c\u8ddfMoE-27B\u57fa\u7ebf\u5bf9\u9f50\u3002<\/p>\n

\u7ed3\u679c\u77e5\u8bc6\u5bc6\u96c6\u578b\u4efb\u52a1\u7684\u63d0\u5347\u7b26\u5408\u9884\u671f(MMLU +3.4\uff0cCMMLU +4.0)\uff0c\u4f46\u901a\u7528\u63a8\u7406\u548c\u4ee3\u7801\u6570\u5b66\u7684\u63d0\u5347\u8d85\u51fa\u9884\u671f(BBH +5.0\uff0cARC-Challenge +3.7\uff0cHumanEval +3.0\uff0cMATH +2.4)\uff0c\u957f\u4e0a\u4e0b\u6587\u573a\u666f\u66f4\u5938\u5f20\uff0cMulti-Query NIAH\u4ece84.2%\u8dc3\u5347\u523097.0%\u3002<\/p>\n

\"DeepSeek<\/p>\n

\u90a3\u4e48\uff0c\u4e3a\u4ec0\u4e48\u8bb0\u5fc6\u6a21\u5757\u8fd8\u80fd\u53cd\u8fc7\u6765\u63d0\u5347\u63a8\u7406?<\/p>\n

LogitLens\u548cCKA\u7ed9\u51fa\u4e86\u7b54\u6848\uff0cEngram-27B\u7b2c5\u5c42\u7684\u8868\u5f81\uff0c\u8ddfMoE\u57fa\u7ebf\u7b2c12\u5c42\u7684\u8868\u5f81\u6700\u76f8\u4f3c\u3002<\/p>\n

Engram\u628a\u6a21\u578b\u7684\u65e9\u671f\u5c42\u4ece\u300c\u91cd\u5efa\u9759\u6001\u77e5\u8bc6\u300d\u8fd9\u79cd\u82e6\u529b\u6d3b\u91cc\u89e3\u653e\u51fa\u6765\uff0c\u8fd9\u90e8\u5206\u7f51\u7edc\u6df1\u5ea6\u88ab\u817e\u51fa\u6765\u505a\u66f4\u590d\u6742\u7684\u63a8\u7406\u3002Engram\u4e0d\u662f\u65b0\u589e\u4e86\u4e00\u5757\u8bb0\u5fc6\uff0c\u5b83\u8fd8\u53d8\u76f8\u628a\u7f51\u7edc\u52a0\u6df1\u4e86\u3002<\/strong><\/p>\n

\"DeepSeek<\/p>\n

\u5de5\u7a0b\u4e0a\u3002\u8bba\u6587\u628a\u4e00\u4e2a1000\u4ebf\u53c2\u6570\u7684Engram\u8868\u6574\u4e2a\u7529\u5230host DRAM\uff0c\u5728H800\u4e0a\u8dd1\u63a8\u7406\uff0c8B-Dense\u7684\u541e\u5410\u635f\u5931\u53ea\u67092.8%\u3002<\/p>\n

\u9760\u7684\u662fEngram\u7d22\u5f15\u7684\u786e\u5b9a\u6027\uff0c\u53ea\u53d6\u51b3\u4e8e\u8f93\u5165token\u5e8f\u5217\uff0c\u5b8c\u5168\u53ef\u4ee5\u63d0\u524d\u7b97\uff0cCPU\u5f02\u6b65\u9884\u53d6\u8ddfGPU\u8ba1\u7b97\u91cd\u53e0\u3002<\/p>\n

\u53ef\u4ee5\u8bf4\uff0c\u8fd9\u4e2a\u6a21\u5757\u5929\u751f\u5c31\u4e0d\u9760HBM\uff0c\u53ea\u53ef\u60dc\u5982\u4ecaV4\u6765\u4e86\uff0cEngram\u6ca1\u6765\u3002<\/p>\n

\u6ca1\u5728v4\uff0c\u4f46\u5728\u5176\u4ed6\u5730\u65b9<\/strong><\/p>\n

\u53d1\u660e\u8005\u628a\u5b83\u653e\u5728\u90a3\u91cc\u6ca1\u52a8\uff0c\u4f46\u8def\u4e0a\u8fd8\u662f\u6709\u4eba\u3002\u4e09\u4e2a\u6708\u91cc\uff0c\u81f3\u5c11\u51fa\u73b0\u4e86\u4e09\u4e2a\u503c\u5f97\u8bf4\u4e00\u4e0b\u7684\u5de5\u4f5c\u3002<\/p>\n

\u628aEngram\u585e\u8fdbCXL\u5185\u5b58\u6c60<\/strong><\/p>\n

3\u670810\u65e5\uff0c\u5317\u5927\u3001\u963f\u91cc\u4e91\u3001\u5c71\u4e1c\u82f1\u4fe1\u3001\u4eba\u5927\u3001\u6e2f\u5927\u8054\u5408\u53d1\u4e86\u4e00\u7bc7\u7cfb\u7edf\u8bba\u6587\uff0c\u300aPooling Engram Conditional Memory in Large Language Models using CXL\u300b\u3002<\/p>\n

\"DeepSeek<\/p>\n

\u4ed6\u4eec\u6ca1\u6539Engram\u672c\u8eab\uff0c\u800c\u662f\u56de\u7b54\u4e86\u4e00\u4e2a\u66f4\u5de5\u7a0b\u7684\u95ee\u9898\uff0c\u5982\u679cEngram\u771f\u7684\u6210\u4e86\u4e0b\u4e00\u4ee3\u6807\u914d\uff0c\u5185\u5b58\u653e\u54ea\u3002<\/p>\n

\u7b54\u6848\u662fCXL\u5185\u5b58\u6c60\u5316\u3002GPU HBM\u653e\u8ba1\u7b97\u6743\u91cd\uff0c\u672c\u5730DRAM\u505a\u4e8c\u7ea7\u7f13\u5b58\uff0cCXL\u6c60\u505a\u4e09\u7ea7\u30028\u53f0\u670d\u52a1\u5668\u5171\u4eab4TB\u5185\u5b58\u6c60\uff0cXConn XC50256\u4ea4\u6362\u82af\u7247<\/a>\u505a\u62d3\u6251\uff0c512GB\/s\u5e26\u5bbd\u3002<\/p>\n

\u6574\u5957\u96c6\u6210\u8fdbSGLang\uff0c\u505a\u4e86\u9884\u53d6-\u8ba1\u7b97\u91cd\u53e0\uff0c\u8dd1\u4e0b\u6765\u7aef\u5230\u7aef\u541e\u5410\u635f\u5931\u5c0f\u4e8e5%\u3002Engram\u8bba\u6587\u91cc\u90a3\u53e5\u300c1000\u4ebf\u5d4c\u5165\u8868\u5378\u8f7dDRAM\u300d\u7684\u8f7b\u63cf\u6de1\u5199\uff0c\u88ab\u4ed6\u4eec\u505a\u6210\u4e8627B\u548c40B\u4e24\u4e2a\u89c4\u6a21\u7684\u771f\u5b9e\u6d4b\u8bd5\u3002<\/p>\n

\u7ed3\u8bba\u5f88\u6e05\u695a\uff0cEngram\u8fd9\u79cd\u786e\u5b9a\u6027\u5bfb\u5740\u3001\u53ef\u9884\u53d6\u7684\u8d1f\u8f7d\uff0c\u51e0\u4e4e\u662f\u4e3aCXL\u91cf\u8eab\u5b9a\u505a\u7684\u3002<\/p>\n

\u4e00\u4e2a\u53cd\u76f4\u89c9\u7684\u5b9e\u9a8c<\/strong><\/p>\n

Engram\u8bba\u6587\u4e0a\u7ebf\u7b2c\u5341\u4e00\u5929\uff0c1\u670823\u65e5\uff0c\u4e00\u4e2a\u53ebTaoLin\u7684\u7814\u7a76\u8005\uff0c\u5355\u4f5c\u8005\uff0c\u653e\u51fa\u4e86\u300aA Collision-FreeHot-Tier Extension for Engram-Style Conditional Memory\u300b\u3002<\/p>\n

\"DeepSeek<\/p>\n

\u4ed6\u60f3\u9a8c\u8bc1\u4e00\u4e2a\u770b\u4e0a\u53bb\u663e\u7136\u7684\u4f18\u5316\uff0cEngram\u7528\u591a\u5934\u54c8\u5e0c\u67e5\u8868\u4f1a\u6709\u51b2\u7a81\uff0c\u5982\u679c\u628a\u9ad8\u9891N-gram\u7528Minimal Perfect Hash Function\u5b8c\u5168\u6d88\u9664\u51b2\u7a81\uff0c\u6a21\u578b\u4f1a\u4e0d\u4f1a\u66f4\u597d\u3002<\/p>\n

\u4ed6\u8bbe\u8ba1\u4e86Engram-Nine\uff0c\u628a\u8bb0\u5fc6\u5206\u6210\u65e0\u51b2\u7a81\u7684\u300c\u70ed\u5c42\u300d\u548c\u4fdd\u7559\u591a\u5934\u54c8\u5e0c\u7684\u300c\u51b7\u5c42\u300d\u3002<\/p>\n

\u7ed3\u679c\u53cd\u76f4\u89c9\u3002\u5728\u4e25\u683ciso-parameter\u63a7\u5236\u4e0b\uff0c\u65e0\u51b2\u7a81\u8bbe\u8ba1\u6ca1\u6709\u7a33\u5b9a\u63d0\u5347\u9a8c\u8bc1loss\u3002<\/p>\n

route-stratified\u8bc4\u4f30\u8fd8\u53d1\u73b0\uff0c\u8bad\u7ec3\u521d\u671f\u70ed\u8def\u5f84(\u9ad8\u9891)loss\u66f4\u4f4e\uff0c\u4f46\u8bad\u7ec3\u540e\u671f\u51b7\u8def\u5f84\u53cd\u8fc7\u6765\u8d85\u8fc7\u70ed\u8def\u5f84\u3002<\/p>\n

\u4e00\u4e2a\u770b\u4e0a\u53bb\u663e\u7136\u7684\u4f18\u5316\u65b9\u5411\uff0c\u88ab\u4e00\u4e2a\u771f\u505a\u5b9e\u9a8c\u7684\u4eba\u8bc1\u4f2a\u4e86\u3002<\/p>\n

\u628aEngram\u63a8\u5230\u89c6\u89c9(AutoArk\/TinyEngram)<\/strong><\/p>\n

GitHub\u4e0a\u4e00\u4e2a\u53ebAutoArk\u7684\u56e2\u961f\u641e\u4e86Tiny Engram\u3002<\/p>\n

\"DeepSeek<\/p>\n

\u57fa\u4e8eQwen-3\u5b8c\u6574\u590d\u73b0\u6587\u672cEngram\u4e4b\u540e\uff0c\u4ed6\u4eec\u505a\u4e86\u4e00\u4ef6\u8bba\u6587\u91cc\u6ca1\u505a\u7684\u4e8b\uff0c\u628aEngram\u642c\u5230Stable Diffusion\u4e0a\u3002<\/p>\n

\u89c6\u89c9patch\u7ecf\u8fc7\u5206\u5c42\u7f16\u7801\uff0c\u5e95\u5c42\u6293\u7eb9\u7406\uff0c\u4e2d\u5c42\u6293\u90e8\u4ef6\uff0c\u9ad8\u5c42\u6293\u98ce\u683c\uff0c\u7136\u540e\u6574\u5957\u4e22\u8fdb\u54c8\u5e0c\u67e5\u8868\u3002<\/p>\n

\u8ddfLoRA\u6bd4\u4e0b\u6765\uff0c\u8fbe\u5230\u540c\u7b49\u6548\u679c\uff0cEngram\u9700\u8981\u7684\u989d\u5916\u53c2\u6570\u53ea\u6709LoRA\u768415%\u523030%\u3002\u8fde\u7eed\u6ce8\u5165\u591a\u4e2a\u65b0\u6982\u5ff5\u65f6\uff0cLoRA\u4f1a\u51fa\u73b0\u660e\u663e\u7684\u6982\u5ff5\u9000\u5316\uff0cEngram\u4e0d\u4f1a\u3002<\/p>\n

Engram\u539f\u672c\u662f\u4e3a\u6587\u672c\u8bbe\u8ba1\u7684\u3002AutoArk\u7b49\u4e8e\u628a\u8fd9\u6247\u95e8\u649e\u5f00\u4e86\uff0c\u51e1\u662f\u80fd\u79bb\u6563\u5316\u3001\u80fd\u54c8\u5e0c\u7684\u6a21\u6001\uff0cEngram\u90fd\u80fd\u642c\u3002<\/p>\n

\u4e09\u4e2a\u6708\u91cc\uff0cEngram\u8fd9\u6761\u8def\u4e0a\uff0c\u53d1\u660e\u8005\u6700\u6c89\u9ed8\uff0c\u8ddf\u8fdb\u8005\u5404\u81ea\u8d70\u4e86\u4e00\u6b65\u3002<\/p>\n

\u4e00\u4e2a\u56e2\u961f\u66ff\u5b83\u89e3\u51b3\u591a\u673a\u5185\u5b58\u5c42\u7ea7\uff0c\u4e00\u4e2a\u72ec\u7acb\u7814\u7a76\u8005\u8bc1\u4f2a\u4e86\u5b83\u4e00\u4e2a\u770b\u4f3c\u663e\u7136\u7684\u4f18\u5316\u65b9\u5411\uff0c\u4e00\u4e2a\u5f00\u6e90\u56e2\u961f\u628a\u5b83\u63a8\u5230\u4e86\u89c6\u89c9\u3002<\/p>\n

\"DeepSeek<\/p>\n

\u800cdeepseek-ai\/Engram\u8fd9\u4e2a\u4ed3\u5e93\uff0c\u6700\u540e\u4e00\u6b21\u63d0\u4ea4\u8fd8\u505c\u57281\u670814\u65e5\u3002<\/p>\n

One more thing<\/strong><\/p>\n

Engram\u8bba\u6587\u7684\u6458\u8981\u7ed3\u5c3e\u6709\u4e00\u53e5\u8bdd\uff1a<\/p>\n

\u6211\u4eec\u8ba4\u4e3a\u6761\u4ef6\u8bb0\u5fc6\u5c06\u662f\u4e0b\u4e00\u4ee3\u7a00\u758f\u6a21\u578b\u4e0d\u53ef\u6216\u7f3a\u7684\u5efa\u6a21\u539f\u8bed\u3002<\/p>\n<\/p>\n

\u770b\u6765\uff0c\u8fd9\u4e2a\u4e0b\u4e00\u4ee3\u5f97\u662fV5\u4e86\uff0c\u96be\u4e0d\u6210\u4f1a\u662fV4.1?<\/p>\n

(\u6765\u6e90\uff1a\u65b0\u6d6a\u79d1\u6280)<\/p>\n","protected":false},"excerpt":{"rendered":"

henry \u53d1\u81ea \u51f9\u975e\u5bfa \u91cf\u5b50\u4f4d | \u516c\u4f17\u53f7 QbitAI DeepSeekV4\u7684\u6280\u672f\u62a5\u544a\u91cc\u6709mHC\uff0c\u6709CSA […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":[],"categories":[6],"tags":[339],"views":9,"_links":{"self":[{"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=\/wp\/v2\/posts\/772596"}],"collection":[{"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=772596"}],"version-history":[{"count":0,"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=\/wp\/v2\/posts\/772596\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=772596"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=772596"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=772596"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}