- \n
- \u4ece\u4e00\u6761\u76f4\u7ebf\u5230\u5927\u6a21\u578b\u8f93\u51fa\u4e00\u4e2atoken（\u5341）：\u9996 token \u8bde\u751f\u4e0e KV-Cache<\/li>\n
- \n
- \n
- 1. \u53d6\u6700\u540e\u4e00\u884c：\u4e3a\u4ec0\u4e48\u662f\u300c？\u300d<\/li>\n
- 2. lm_head：\u6700\u540e\u4e00\u6b21\u300c\u4e00\u884c\u4e58\u4e00\u5217\u300d<\/li>\n
- 3. Softmax：\u5206\u6570\u53d8\u6982\u7387<\/li>\n
- 4. \u9009 token：\u4e09\u79cd\u7b56\u7565<\/li>\n
- 5. \u9996 token \u8bde\u751f<\/li>\n
- 6. \u81ea\u56de\u5f52：\u7b2c\u4e8c\u4e2a\u5b57\u600e\u4e48\u6765<\/li>\n
- 7. KV-Cache：\u65e7 token \u7684 K\/V \u4e0d\u7528\u91cd\u7b97<\/li>\n
- \n
- \n
- 7.1 \u6d6a\u8d39\u5728\u54ea<\/li>\n
- 7.2 \u7f13\u5b58\u65b9\u6848<\/li>\n
- 7.3 \u6536\u76ca\u4e0e\u4ee3\u4ef7<\/li>\n
- 7.4 GQA \u4e0e MLA：\u7ed9 KV-Cache \u7626\u8eab\u7684\u4e24\u628a\u5200<\/li>\n<\/ul>\n<\/li>\n
- 8. \u6536\u5b98：\u4ece\u4e00\u6761\u76f4\u7ebf\u5230\u4e00\u4e2a token<\/li>\n
- \n
- \n
- 8.1 \u5341\u7bc7\u65c5\u7a0b<\/li>\n
- 8.2 230B \u7684\u6700\u540e\u4e00\u73af<\/li>\n
- 8.3 \u5199\u5728\u7cfb\u5217\u5b8c\u7ed3<\/li>\n<\/ul>\n<\/li>\n
- 9. \u4e0b\u4e00\u7cfb\u5217\u9884\u544a：17 \u4e2a\u95ee\u9898<\/li>\n
- \u5c0f\u7ed3<\/li>\n
- \u7cfb\u5217\u5b8c\u7ed3<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n
)<\/p>\n
1. \u53d6\u6700\u540e\u4e00\u884c：\u4e3a\u4ec0\u4e48\u662f\u300c？\u300d<\/h3>\n
\u7bc79 \u7ed3\u5c3e\u8bf4\u8fc7：\u8f93\u51fa\u77e9\u9635\u7684\u6700\u540e\u4e00\u884c\u2014\u2014\u300c？\u300d\u884c\u2014\u2014\u662f\u9884\u6d4b\u9996 token \u7684\u5168\u90e8\u539f\u6599\u3002\u4e3a\u4ec0\u4e48\u504f\u504f\u662f\u6700\u540e\u4e00\u884c？<\/p>\n
\u56e0\u4e3a\u56e0\u679c\u8bed\u8a00\u6a21\u578b\u4ece\u5de6\u5230\u53f3\u9884\u6d4b\u4e0b\u4e00\u4e2a token：\u8bad\u7ec3\u65f6\u6bcf\u4e2a token \u7684\u4efb\u52a1\u90fd\u662f"\u6839\u636e\u5de6\u8fb9\u6240\u6709 token，\u9884\u6d4b\u6211\u53f3\u8fb9\u662f\u8c01"\u3002\u300c\u4eca\u5929\u300d\u9884\u6d4b\u300c\u897f\u5b89\u300d，\u300c\u897f\u5b89\u300d\u9884\u6d4b\u300c\u7684\u300d\u2026\u2026\u800c\u300c？\u300d\u662f\u53e5\u5b50\u7684\u6700\u540e\u4e00\u4e2a token，\u53ea\u6709\u5b83\u89c1\u8fc7\u5168\u90e8\u524d\u6587\u2014\u2014\u4eca\u5929\u3001\u897f\u5b89\u3001\u7684\u3001\u5929\u6c14\u3001\u600e\u4e48\u6837，\u4e00\u4e2a\u4e0d\u843d\u3002<\/p>\n
\u6240\u4ee5\u9884\u6d4b"？\u201c\u540e\u9762\u662f\u4ec0\u4e48，\u5c31\u7528\u201d？"\u7684\u884c\u5411\u91cf\u3002\u628a 6\u00d74096 \u7684\u8f93\u51fa\u77e9\u9635\u53ea\u53d6\u6700\u540e\u4e00\u884c，\u5f97\u5230 v\u2014\u20141\u00d74096 \u7684\u884c\u5411\u91cf（\u6f14\u793a 1\u00d74），\u5982\u56fe 10-1 \u6240\u793a\u3002<\/p>\n
![\"\u53d6\u6700\u540e\u4e00\u884c\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260828233341-6a921ad54728a.png\")
<\/p>\n\u56fe10-1 \u53ea\u53d6\u6700\u540e\u4e00\u884c：\u300c？\u300d\u884c\u662f\u5168\u53e5\u4fe1\u606f\u7684\u6c47\u96c6\u70b9<\/p>\n
\u6ce8\u610f\u56fe 10-1 \u91cc\u7684\u7ec6\u8282：v = [2.85, -0.35, 1.90, -1.32]，d1 = 2.85 \u662f\u884c\u5185\u6700\u5927\u503c\u2014\u2014\u8fd9\u6b63\u662f\u7bc79 \u53cd\u590d\u94fa\u57ab\u7684\u90a3\u6761\u7ebf：\u300c？\u300d\u884c d1 \u5728\u6df1\u5c42\u5168\u9762\u63a5\u7ba1，"\u95ee\u5b8c\u4e86，\u8be5\u4f5c\u7b54"\u7684\u5f00\u5173\u4fe1\u53f7\u5df2\u7ecf\u5c31\u4f4d\u3002<\/p>\n
2. lm_head：\u6700\u540e\u4e00\u6b21\u300c\u4e00\u884c\u4e58\u4e00\u5217\u300d<\/h3>\n
\u73b0\u5728\u628a v \u7ffb\u8bd1\u6210\u8bcd\u8868\u4e0a\u6bcf\u4e2a\u5019\u9009 token \u7684\u5206\u6570\u3002\u7bc79 \u5df2\u7ecf\u9884\u544a\u8fc7\u5de5\u5177：lm_head\u2014\u2014\u4e00\u4e2a 4096\u00d7129,280 \u7684\u6743\u91cd\u77e9\u9635（5.3 \u4ebf\u53c2\u6570）\u3002\u8ba1\u7b97\u8fd8\u662f\u90a3\u4e2a\u8d2f\u7a7f\u5168\u7cfb\u5217\u7684\u52a8\u4f5c：\u4e00\u884c\u4e58\u4e00\u5217\u3002<\/p>\n
\u6f14\u793a\u7528 8 \u4e2a\u5019\u9009 token \u7684\u5c0f\u8bcd\u8868（\u4eca\u3001\u6674\u3001\u96e8\u3001\u591a\u3001\u9634\u3001\u96ea\u3001\u98ce\u3001\u5f88），lm_head \u5c31\u662f 4\u00d78 \u7684\u5c0f\u77e9\u9635\u3002v \u4e58\u5b83，\u5f97\u5230 8 \u4e2a\u5206\u6570（logits），\u5982\u56fe 10-2 \u6240\u793a\u3002<\/p>\n
![\"lm_head](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260828233343-6a921ad7a09c5.png\")
<\/p>\n\u56fe10-2 lm_head：\u6700\u540e\u4e00\u6b21\u300c\u4e00\u884c\u4e58\u4e00\u5217\u300d\u2014\u20144096 \u7ef4\u7ffb\u8bd1\u6210\u8bcd\u8868\u5206\u6570<\/p>\n
\u624b\u7b97\u300c\u6674\u300d\u90a3\u4e00\u5217（\u56fe 10-2 \u9ec4\u8272\u9ad8\u4eae\u5217）：<\/p>\n
<\/p>\n
2.85<\/p>\n
\u00d7<\/p>\n
0.70<\/p>\n
+<\/p>\n
(<\/p>\n
\u2212<\/p>\n
0.35<\/p>\n
)<\/p>\n
\u00d7<\/p>\n
(<\/p>\n
\u2212<\/p>\n
0.30<\/p>\n
)<\/p>\n
+<\/p>\n
1.90<\/p>\n
\u00d7<\/p>\n
0.40<\/p>\n
+<\/p>\n
(<\/p>\n
\u2212<\/p>\n
1.32<\/p>\n
)<\/p>\n
\u00d7<\/p>\n
0.50<\/p>\n
=<\/p>\n
2.20<\/p>\n
2.85 \\\\times 0.70 + (-0.35) \\\\times (-0.30) + 1.90 \\\\times 0.40 + (-1.32) \\\\times 0.50 = 2.20<\/p>\n
<\/span><\/span>2.85<\/span><\/span>\u00d7<\/span><\/span><\/span><\/span>0.70<\/span><\/span>+<\/span><\/span><\/span><\/span>(<\/span>\u2212<\/span>0.35<\/span>)<\/span><\/span>\u00d7<\/span><\/span><\/span><\/span>(<\/span>\u2212<\/span>0.30<\/span>)<\/span><\/span>+<\/span><\/span><\/span><\/span>1.90<\/span><\/span>\u00d7<\/span><\/span><\/span><\/span>0.40<\/span><\/span>+<\/span><\/span><\/span><\/span>(<\/span>\u2212<\/span>1.32<\/span>)<\/span><\/span>\u00d7<\/span><\/span><\/span><\/span>0.50<\/span><\/span>=<\/span><\/span><\/span><\/span>2.20<\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n
8 \u4e2a logits \u5168\u7b97\u51fa\u6765：<\/p>\n
\n
\u5019\u9009\u4eca\u6674\u96e8\u591a\u9634\u96ea\u98ce\u5f88<\/tr>\n \n\n \u5206\u6570<\/td>\n 0.18<\/td>\n 2.20<\/td>\n 0.69<\/td>\n 0.94<\/td>\n 0.73<\/td>\n 0.54<\/td>\n 0.41<\/td>\n -0.14<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n \u886810-1 8 \u4e2a\u5019\u9009 token \u7684 logits（\u6f14\u793a\u8bcd\u8868，\u771f\u5b9e\u662f 129,280 \u4e2a\u5206\u6570）<\/p>\n
\u300c\u6674\u300d\u4ee5 2.20 \u4e00\u9a91\u7edd\u5c18\u3002\u4f46\u5206\u6570\u8fd8\u4e0d\u662f\u6982\u7387\u2014\u2014\u5b83\u4eec\u751a\u81f3\u53ef\u4ee5\u662f\u4efb\u610f\u5b9e\u6570（\u6709\u8d1f\u6570，\u4e5f\u6ca1\u5f52\u4e00\u5316）\u3002\u4ece\u5206\u6570\u5230\u6982\u7387，\u5dee\u4e00\u4e2a Softmax\u3002<\/p>\n
3. Softmax：\u5206\u6570\u53d8\u6982\u7387<\/h3>\n
Softmax \u7684\u516c\u5f0f：<\/p>\n
<\/p>\n
P<\/p>\n
i<\/p>\n
=<\/p>\n
e<\/p>\n
z<\/p>\n
i<\/p>\n
\u2211<\/p>\n
j<\/p>\n
e<\/p>\n
z<\/p>\n
j<\/p>\n
P_i = \\\\frac{e^{z_i}}{\\\\sum_j e^{z_j}}<\/p>\n
<\/span><\/span>P<\/span><\/span>i<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span><\/span><\/span>\u2211<\/span><\/span>j<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>e<\/span><\/span>z<\/span><\/span>j<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>e<\/span><\/span>z<\/span><\/span>i<\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n
\u9010\u4e2a\u5019\u9009\u7b97\u6307\u6570\u3001\u518d\u9664\u4ee5\u603b\u548c\u3002\u624b\u7b97\u4e00\u904d，\u53ea\u6709\u4e09\u6b65：<\/p>\n
\u7b2c\u4e00\u6b65：\u9010\u683c\u53d6\u6307\u6570 e^z\u3002 \u300c\u6674\u300de^2.20 = 9.03，\u300c\u591a\u300de^0.94 = 2.55，\u300c\u96e8\u300de^0.69 = 1.99\u2026\u2026\u300c\u5f88\u300de^-0.14 = 0.87\u3002\u6307\u6570\u5e72\u7684\u4e8b：\u628a\u5dee\u8ddd\u653e\u5927\u2014\u2014\u5206\u6570\u4e0a\u6674\u53ea\u662f\u591a\u7684 2.3 \u500d，\u6307\u6570\u540e\u62c9\u5f00\u5230 3.5 \u500d；\u5206\u6570\u4e0a\u300c\u5f88\u300d\u662f\u8d1f\u7684，\u6307\u6570\u540e\u53d8\u6210 0.87，\u57ab\u5e95\u4f46\u6ca1\u5f52\u96f6\u3002<\/p>\n
\u7b2c\u4e8c\u6b65：\u5168\u90e8\u52a0\u8d77\u6765\u3002 1.19 + 9.03 + 1.99 + 2.55 + 2.08 + 1.71 + 1.50 + 0.87 = 20.91\u3002<\/p>\n
\u7b2c\u4e09\u6b65：\u6bcf\u683c\u9664\u4ee5\u603b\u548c\u3002 \u300c\u6674\u300d= 9.03 \/ 20.91 = 43.2%；\u300c\u591a\u300d= 2.55 \/ 20.91 = 12.2%\u2026\u20268 \u4e2a\u6982\u7387\u52a0\u8d77\u6765\u6070\u597d 100%\u3002<\/p>\n
\u6574\u4e2a\u8fc7\u7a0b\u52a8\u56fe\u6f14\u793a\u5982\u56fe 10-3 \u6240\u793a（4 \u5e27\u5faa\u73af，\u6bcf\u5e27\u8ffd\u52a0\u4e00\u6b65\u63a8\u5bfc）\u3002<\/p>\n
![\"Softmax](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260828233346-6a921ada41042.gif\")
<\/p>\n\u56fe10-3 Softmax \u624b\u7b97\u5168\u8fc7\u7a0b：logits \u2192 e^z \u2192 \u6c42\u548c \u2192 \u6982\u7387（4 \u5e27\u8ffd\u52a0\u5f0f\u52a8\u56fe）<\/p>\n
8 \u4e2a\u5019\u9009\u7684\u6700\u7ec8\u6982\u7387（\u70ed\u529b\u56fe）\u5982\u56fe 10-4 \u6240\u793a\u3002<\/p>\n
![\"Softmax](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260828233347-6a921adb1e153.png\")
<\/p>\n\u56fe10-4 Softmax \u540e\u7684\u6982\u7387：\u989c\u8272\u8d8a\u6df1\u6982\u7387\u8d8a\u9ad8（\u5de6：\u70ed\u529b\u56fe；\u53f3：\u6e29\u5ea6\u5bf9\u6bd4）<\/p>\n
- \n
- \u6674 43.2%\u2014\u2014\u4e00\u5bb6\u72ec\u5927<\/li>\n
- \u591a 12.2%\u3001\u9634 9.9%\u3001\u96e8 9.5%\u2014\u2014\u7b2c\u4e8c\u68af\u961f，\u90fd\u662f\u5408\u7406\u5929\u6c14\u63cf\u8ff0<\/li>\n
- \u96ea 8.2%\u3001\u98ce 7.2%\u2014\u2014\u8fb9\u7f18\u5019\u9009<\/li>\n
- \u4eca 5.7%\u3001\u5f88 4.1%\u2014\u2014\u966a\u8dd1<\/li>\n<\/ul>\n
\u4e3a\u4ec0\u4e48\u300c\u4eca\u300d\u8fd9\u4e48\u4f4e？\u56e0\u4e3a**\u300c\u4eca\u5929\u300d\u5df2\u7ecf\u5728\u8f93\u5165\u91cc\u4e86**\u2014\u2014\u6ce8\u610f\u529b\u673a\u5236\u8ba9\u300c？\u300d\u5438\u6536\u8fc7\u300c\u4eca\u5929\u300d\u7684\u4fe1\u606f，\u6a21\u578b\u77e5\u9053\u65f6\u95f4\u9650\u5b9a\u5df2\u7ecf\u4ea4\u4ee3\u8fc7，\u7b54\u6848\u4e0d\u9700\u8981\u518d\u8bf4\u4e00\u904d"\u4eca\u5929"\u3002\u4e3a\u4ec0\u4e48\u300c\u5f88\u300d\u57ab\u5e95？\u56e0\u4e3a\u300c\u600e\u4e48\u6837\u300d\u89c4\u5b9a\u7684\u662f\u63cf\u8ff0\u6027\u8bcd\u6027，\u526f\u8bcd\u5f00\u5934（\u201c\u5f88\u6674\u201d？）\u8bed\u6cd5\u4e0a\u52c9\u5f3a\u3001\u8bed\u4e49\u4e0a\u522b\u626d\u3002\u7bc79 \u7684 6 \u6761\u7ea6\u675f\u94fe，\u5728\u6982\u7387\u5206\u5e03\u4e0a\u5168\u90e8\u5151\u73b0：<\/p>\n
\u65f6\u95f4\u9650\u5b9a（\u4e0d\u8bf4"\u4eca\u5929"\u4e86）\u2713 \u4e3b\u4f53\u9501\u5b9a（\u8bf4\u7684\u662f\u897f\u5b89\u4e0d\u662f\u522b\u5904）\u2713 \u7ed3\u6784\u9ecf\u5408（"\u897f\u5b89\u7684\u5929\u6c14"\u662f\u5b8c\u6574\u4e3b\u8bed）\u2713 \u9886\u57df\u9501\u5b9a（\u5019\u9009\u5168\u662f\u6c14\u8c61\u8bcd）\u2713 \u8bcd\u6027\u9501\u5b9a（\u63cf\u8ff0\u6027\u5f00\u5934）\u2713 \u5f00\u59cb\u5f00\u5173（d1 = 2.85 \u5df2\u7ecf\u6309\u4e0b）<\/p>\n
4. \u9009 token：\u4e09\u79cd\u7b56\u7565<\/h3>\n
\u6982\u7387\u6709\u4e86，\u600e\u4e48\u9009\u51fa\u90a3\u4e2a token？\u4e09\u79cd\u4e3b\u6d41\u7b56\u7565，\u5982\u56fe 10-5 \u6240\u793a\u3002<\/p>\n
![\"\u9009token\u7b56\u7565\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260828233349-6a921add239d8.png\")
<\/p>\n\u56fe10-5 \u62ff\u5230\u6982\u7387\u4e4b\u540e\u600e\u4e48\u9009：\u4e09\u79cd\u7b56\u7565<\/p>\n
- \n
- \u8d2a\u5fc3\u89e3\u7801：\u6c38\u8fdc\u9009\u6982\u7387\u6700\u9ad8\u7684\u2014\u2014\u5fc5\u9009\u300c\u6674\u300d\u3002\u6bcf\u6b21\u8f93\u51fa\u5b8c\u5168\u4e00\u6837，\u786e\u5b9a\u4f46\u5446\u677f，\u8fde\u7eed\u751f\u6210\u65f6\u5bb9\u6613\u9677\u5165\u91cd\u590d\u5faa\u73af\u3002\u9002\u5408\u4ee3\u7801\u3001\u6570\u5b66\u3002<\/li>\n
- Top-k \u91c7\u6837：\u53ea\u5728\u524d k \u4e2a\u5019\u9009\u91cc\u968f\u673a\u62bd\u3002\u6bd4\u5982 k=2：\u53d6\u6674\u3001\u591a，\u91cd\u65b0\u5f52\u4e00\u5316\u6210 78% \/ 22%\u2014\u2014\u591a\u6570\u65f6\u5019\u9009\u6674，\u5076\u5c14\u9009\u591a\u3002k \u662f\u786c\u95e8\u69db，\u5206\u5e03\u5f88\u5e73\u65f6\u5019\u9009\u4f1a\u663e\u5f97\u592a\u5c11\u3002<\/li>\n
- Top-p \u91c7\u6837（nucleus）：\u6309\u6982\u7387\u964d\u5e8f\u7d2f\u52a0，\u7d2f\u79ef\u5230 p（\u6bd4\u5982 0.9）\u5c31\u505c，\u5728\u622a\u65ad\u7684\u96c6\u5408\u91cc\u62bd\u3002\u5019\u9009\u6570\u968f\u5206\u5e03\u81ea\u9002\u5e94：\u5206\u5e03\u5c16\u65f6\u5019\u9009\u5c11，\u5e73\u65f6\u5019\u9009\u591a\u3002\u5f53\u524d\u5404\u5bb6 API \u7684\u9ed8\u8ba4\u9009\u9879\u3002<\/li>\n<\/ul>\n
**\u6e29\u5ea6（Temperature）**\u662f\u91c7\u6837\u524d\u7684\u5206\u5e03\u8c03\u8282\u65cb\u94ae：\u628a logits \u9664\u4ee5 T \u518d Softmax\u3002\u56fe 10-4 \u53f3\u4fa7\u662f\u4e09\u79cd\u6e29\u5ea6\u7684\u5bf9\u6bd4\u2014\u2014T=0.5 \u65f6\u300c\u6674\u300d\u653e\u5927\u5230 78.6%（\u4fdd\u5b88），T=2.0 \u65f6\u538b\u5e73\u5230 25.0%（\u653e\u98de）\u3002\u6e29\u5ea6\u4e0d\u6539\u5019\u9009\u6392\u5e8f，\u53ea\u6539"\u96c6\u4e2d\u5ea6"\u3002<\/p>\n
\u672c\u7cfb\u5217\u6f14\u793a\u7528\u8d2a\u5fc3：\u9996 token = \u300c\u6674\u300d\u3002<\/p>\n
5. \u9996 token \u8bde\u751f<\/h3>\n
\u628a\u5168\u94fe\u8def\u4e32\u4e00\u904d，\u8fd9\u662f\u6574\u4e2a\u7cfb\u5217\u7b49\u4e86 10 \u7bc7\u7684\u65f6\u523b：<\/p>\n
- \u7bc72：\u300c\u4eca\u5929\u897f\u5b89\u7684\u5929\u6c14\u600e\u4e48\u6837？\u300d\u2192 6 \u4e2a token [5237, 23872, 301, 16652, 19602, 1148]<\/li>\n
- \u7bc73：\u67e5\u5d4c\u5165\u8868 \u2192 6\u00d74096<\/li>\n
- \u7bc74：RoPE \u628a\u4f4d\u7f6e\u65cb\u8fdb\u5411\u91cf<\/li>\n
- \u7bc75~8：\u8fc7\u7b2c 1 \u4e2a Block\u2014\u2014\u5f52\u4e00\u5316\u3001\u6ce8\u610f\u529b\u4ea4\u6362\u4fe1\u606f\u3001\u6b8b\u5dee\u4fdd\u5e95\u3001FFN \u6df1\u52a0\u5de5<\/li>\n
- \u7bc79：\u518d\u8fc7 31 \u5c42 \u2192 \u8f93\u51fa\u77e9\u9635，\u300c？\u300d\u884c d1 \u5168\u9762\u63a5\u7ba1<\/li>\n
- \u672c\u7bc7：\u53d6\u300c？\u300d\u884c \u2192 \u4e58 lm_head \u5f97 129,280 \u4e2a\u5206\u6570 \u2192 Softmax \u2192 \u300c\u6674\u300d\u4ee5 43.2% \u80dc\u51fa<\/li>\n
\u4ece\u300c\u4eca\u5929\u897f\u5b89\u7684\u5929\u6c14\u600e\u4e48\u6837？\u300d\u5230\u300c\u6674\u300d\u2014\u2014\u4e00\u4e2a token \u8bde\u751f\u4e86\u3002\u8fd9\u4e0d\u662f\u68c0\u7d22\u3001\u4e0d\u662f\u67e5\u8868，\u800c\u662f 80.4 \u4ebf\u4e2a\u53c2\u6570\u63a5\u529b\u52a0\u5de5\u51fa\u6765\u7684\u6982\u7387\u9009\u62e9\u3002\u56de\u7b54\u7684\u5f00\u5934，\u5c31\u6b64\u843d\u5730\u3002<\/p>\n
6. \u81ea\u56de\u5f52：\u7b2c\u4e8c\u4e2a\u5b57\u600e\u4e48\u6765<\/h3>\n
\u4e00\u4e2a token \u663e\u7136\u4e0d\u591f\u2014\u2014\u7528\u6237\u8981\u7684\u662f\u5b8c\u6574\u7684\u5929\u6c14\u64ad\u62a5\u3002\u81ea\u56de\u5f52（autoregressive）：\u628a\u521a\u751f\u6210\u7684\u300c\u6674\u300d\u62fc\u56de\u8f93\u5165\u53e5\u5c3e，\u8f93\u5165\u4ece 6 \u4e2a token \u53d8\u6210 7 \u4e2a，\u518d\u5b8c\u6574\u8d70\u4e00\u904d\u524d\u5411（32 \u5c42 \u2192 \u53d6\u6700\u540e\u4e00\u884c \u2192 lm_head \u2192 Softmax \u2192 \u9009 token），\u5f97\u5230\u7b2c\u4e8c\u4e2a token\u300c\u5929\u300d\u3002\u5faa\u73af\u5f80\u590d，\u50cf\u591a\u7c73\u8bfa\u9aa8\u724c，\u5982\u56fe 10-6 \u6240\u793a\u3002<\/p>\n
![\"\u81ea\u56de\u5f52\u5faa\u73af\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260828233351-6a921adf663c6.png\")
<\/p>\n\u56fe10-6 \u81ea\u56de\u5f52：\u9996 token \u62fc\u56de\u53e5\u5c3e，\u751f\u6210\u7b2c\u4e8c\u4e2a\u5b57<\/p>\n
\u7528\u4f2a\u4ee3\u7801\u8868\u8fbe：<\/p>\n
```python
\ntokens = \u5206\u8bcd("\u4eca\u5929\u897f\u5b89\u7684\u5929\u6c14\u600e\u4e48\u6837？") # 6 \u4e2a
\nwhile \u672a\u9047\u5230 EOS and \u957f\u5ea6\u672a\u8d85\u9650:
\n out_matrix = transformer_32\u5c42(\u5d4c\u5165(tokens))
\n v = out_matrix[-1] # \u53d6\u6700\u540e\u4e00\u884c
\n logits = v @ lm_head # \u8bcd\u8868\u5206\u6570
\n probs = softmax(logits \/ T) # \u6982\u7387
\n next_tok = \u91c7\u6837(probs) # \u8d2a\u5fc3 \/ top-k \/ top-p
\n tokens.append(next_tok) # \u62fc\u56de\u53e5\u5c3e
\n\u56de\u7b54 = \u53cd\u67e5\u8bcd\u8868(tokens[6:]) # \u6674\u5929，\u2026\u2026
\n```<\/p>\n\u6ce8\u610f\u5faa\u73af\u4f53\u7684\u6bcf\u4e00\u8f6e\u90fd\u8981\u8fc7\u5b8c\u6574\u7684 32 \u5c42\u2014\u2014\u751f\u6210\u662f\u9010 token \u7684，\u6bcf\u4e2a\u5b57\u90fd\u662f\u4e00\u6b21\u5168\u91cf\u524d\u5411\u3002\u8fd9\u5e26\u6765\u4e86\u4e00\u4e2a\u660e\u663e\u7684\u6d6a\u8d39，\u4e0b\u4e00\u8282\u89e3\u51b3\u3002<\/p>\n
7. KV-Cache：\u65e7 token \u7684 K\/V \u4e0d\u7528\u91cd\u7b97<\/h3>\n
7.1 \u6d6a\u8d39\u5728\u54ea<\/h4>\n
\u751f\u6210\u300c\u5929\u300d\u65f6\u8f93\u5165\u662f 7 \u4e2a token，\u6734\u7d20\u505a\u6cd5 7 \u4e2a\u5168\u90e8\u8fc7 32 \u5c42\u3002\u4f46\u4ed4\u7ec6\u60f3：\u524d 6 \u4e2a token \u7684 K\u3001V \u5411\u91cf，\u548c\u4e0a\u4e00\u8f6e\u7b97\u51fa\u6765\u7684\u4e00\u4e2a\u6570\u90fd\u4e0d\u5dee\u3002<\/p>\n
\u4e3a\u4ec0\u4e48？\u7bc77 \u8bb2\u8fc7\u56e0\u679c\u63a9\u7801：\u6bcf\u4e2a token \u53ea\u80fd\u770b\u89c1\u81ea\u5df1\u548c\u5de6\u8fb9\u7684 token\u3002\u300c\u6674\u300d\u62fc\u5728\u53e5\u5c3e，\u6539\u53d8\u4e0d\u4e86\u65e7 token \u7684\u89c6\u91ce\u2014\u2014\u300c\u897f\u5b89\u300d\u8be5\u770b\u8c01\u8fd8\u662f\u770b\u8c01，\u5b83\u7684 K\/V \u548c\u6ce8\u610f\u529b\u7ed3\u679c\u7eb9\u4e1d\u4e0d\u52a8\u3002\u8fd9\u662f\u6570\u5b66\u4e0a\u7684\u7b49\u4ef7，\u4e0d\u662f\u8fd1\u4f3c\u3002<\/p>\n
7.2 \u7f13\u5b58\u65b9\u6848<\/h4>\n
KV-Cache \u7684\u505a\u6cd5：\u628a\u6bcf\u4e00\u5c42\u7b97\u51fa\u7684 K\u3001V \u7f13\u5b58\u4e0b\u6765；\u4e0b\u4e00\u6b65\u53ea\u7b97\u65b0 token \u4e00\u4e2a\u7684 Q\u3001K\u3001V\u2014\u2014\u65b0 K\/V \u8ffd\u52a0\u8fdb\u7f13\u5b58，\u65b0 Q \u548c\u5168\u90e8\u7f13\u5b58 K \u7b97\u6ce8\u610f\u529b\u3002\u5bf9\u6bd4\u6548\u679c\u5982\u56fe 10-7 \u6240\u793a\u3002<\/p>\n
![\"KV-Cache\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260828233353-6a921ae1b75fa.png\")
<\/p>\n\u56fe10-7 KV-Cache：\u7f13\u5b58\u6bcf\u5c42\u7684 K\/V，\u65b0 token \u53ea\u7b97\u81ea\u5df1\u7684 Q\/K\/V<\/p>\n
\u4ece\u77e9\u9635\u89c6\u89d2\u770b\u7f13\u5b58\u6269\u5145：\u751f\u6210\u300c\u6674\u300d\u4e4b\u524d，K \u7f13\u5b58\u662f 6\u00d74；\u300c\u6674\u300d\u62fc\u8fdb\u6765\u540e，\u53ea\u65b0\u7b97\u5b83\u90a3\u4e00\u884c，\u8ffd\u52a0\u6210 7\u00d74\u2014\u2014\u524d 6 \u884c\u539f\u5c01\u4e0d\u52a8，\u5982\u56fe 10-8 \u6240\u793a\u3002<\/p>\n
![\"KV\u7f13\u5b58\u6269\u5145\u70ed\u529b\u56fe\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260828233356-6a921ae48a715.png\")
<\/p>\n\u56fe10-8 KV \u7f13\u5b58\u7684\u6269\u5145：\u4ece 6\u00d74 \u5230 7\u00d74\u2014\u2014\u65e7\u7684\u9010\u683c\u4e0d\u52a8，\u65b0\u7684\u6574\u884c\u8ffd\u52a0（\u989c\u8272\u6df1\u6d45 = \u6570\u503c\u5927\u5c0f）<\/p>\n
\u4f2a\u4ee3\u7801：<\/p>\n
```python
\nkv_cache = [] # \u6bcf\u5c42\u4e00\u4efd (K, V)
\n# \u9996 token：\u5168\u91cf\u8ba1\u7b97，K\/V \u5168\u90e8\u5165\u7f13\u5b58
\nout, kv_cache = forward_32\u5c42(\u5d4c\u5165(tokens), cache=None)
\n# \u540e\u7eed token：\u53ea\u7b97\u65b0\u7684 1 \u4e2a
\nfor \u6bcf\u4e00\u6b65:
\n out, kv_cache = forward_32\u5c42(\u5d4c\u5165(\u65b0token), cache=kv_cache)
\n # \u5185\u90e8：\u65b0 Q \u00d7 \u7f13\u5b58 K \u2192 \u6ce8\u610f\u529b；\u65b0 K\/V \u8ffd\u52a0\u8fdb\u7f13\u5b58
\n```<\/p>\n\u628a\u56fe 10-6 \u548c\u56fe 10-7 \u7684\u673a\u5236\u5408\u8d77\u6765，\u5982\u56fe 10-9 \u6240\u793a（4 \u5e27\u5faa\u73af\u52a8\u56fe）：<\/p>\n
![\"\u81ea\u56de\u5f52+KV-Cache](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260828233358-6a921ae6c4580.gif\")
<\/p>\n\u56fe10-9 \u81ea\u56de\u5f52\u751f\u6210 + KV-Cache \u5168\u8fc7\u7a0b\u52a8\u56fe<\/p>\n
7.3 \u6536\u76ca\u4e0e\u4ee3\u4ef7<\/h4>\n
\u6536\u76ca：\u751f\u6210\u7b2c N \u4e2a token \u7684\u6bcf\u6b65\u8ba1\u7b97\u91cf，\u4ece"\u5168\u90e8 N \u4e2a token \u8fc7 32 \u5c42"\u964d\u5230"1 \u4e2a\u65b0 token \u8fc7 32 \u5c42"\u3002\u957f\u56de\u7b54\u7684\u8ba1\u7b97\u91cf\u4ece\u5e73\u65b9\u7ea7\u964d\u56de\u8fd1\u7ebf\u6027\u2014\u2014\u8fd9\u662f\u6240\u6709\u5927\u6a21\u578b\u63a8\u7406\u5f15\u64ce\u7684\u6807\u914d\u4f18\u5316，\u6ca1\u6709\u5b83，\u9010 token \u751f\u6210\u7684\u5ef6\u8fdf\u4f1a\u968f\u56de\u7b54\u53d8\u957f\u6025\u5267\u6076\u5316\u3002<\/p>\n
\u4ee3\u4ef7：\u7f13\u5b58\u8981\u5360\u663e\u5b58\u3002\u7c97\u7565\u516c\u5f0f：2 \u00d7 \u5c42\u6570 \u00d7 \u5e8f\u5217\u957f\u5ea6 \u00d7 \u5bbd\u5ea6（K \u548c V \u4e24\u4efd）\u3002\u5bf9 LLaMA-3-8B：2 \u00d7 32 \u5c42 \u00d7 \u957f\u5ea6 \u00d7 4096 \u7ef4\u2014\u201432K \u4e0a\u4e0b\u6587\u65f6 KV-Cache \u5c31\u8981\u5360\u7ea6 8GB（fp16），\u957f\u4e0a\u4e0b\u6587\u7684\u663e\u5b58\u5927\u5934\u3002\u7bc76 \u57cb\u7684\u4f0f\u7b14\u5728\u6b64\u56de\u6536\u2014\u2014\u600e\u4e48\u7ed9 KV-Cache \u7626\u8eab？\u4e1a\u754c\u6709\u4e24\u628a\u5200\u3002<\/p>\n
7.4 GQA \u4e0e MLA：\u7ed9 KV-Cache \u7626\u8eab\u7684\u4e24\u628a\u5200<\/h4>\n
\u7b2c\u4e00\u628a\u5200：GQA（\u5206\u7ec4\u67e5\u8be2\u6ce8\u610f\u529b，Grouped-Query Attention）\u2014\u2014\u4ece"\u5934\u6570"\u4e0b\u624b\u3002<\/p>\n
\u7bc76 \u8bb2\u8fc7：\u6807\u51c6 MHA（\u591a\u5934\u6ce8\u610f\u529b）\u91cc 32 \u4e2a Q \u5934\u914d 32 \u7ec4 K\/V \u5934\u3002GQA \u7684\u601d\u8def：Q \u5934\u4fdd\u6301 32 \u4e2a\u4e0d\u52a8，K\/V \u5934\u780d\u5230 8 \u7ec4，\u6bcf 4 \u4e2a Q \u5934\u5171\u4eab 1 \u7ec4 K\/V\u3002\u6ce8\u610f\u529b\u7684\u8ba1\u7b97\u7ed3\u6784\u51e0\u4e4e\u4e0d\u53d8（Q \u548c K \u7684\u4e58\u6cd5\u7167\u505a），\u4f46\u7f13\u5b58\u91cc\u6bcf\u5c42\u53ea\u9700\u5b58 8 \u7ec4 K\/V\u2014\u2014\u7f13\u5b58\u76f4\u63a5\u7701 4 \u500d\u3002<\/p>\n
\u4e3a\u4ec0\u4e48\u6562\u5171\u4eab？\u56e0\u4e3a\u5b9e\u9a8c\u53d1\u73b0：\u4e0d\u540c Q \u5934\u5173\u6ce8\u7684 K\/V \u6a21\u5f0f\u6709\u5927\u91cf\u91cd\u53e0\u2014\u2014\u4f60\u67e5"\u4eca\u5929\u662f\u65f6\u95f4"\u3001\u4ed6\u67e5"\u4eca\u5929\u662f\u65e5\u5b50"，\u7528\u540c\u4e00\u4efd"\u4eca\u5929"\u7684 K\/V \u5b8c\u5168\u591f\u7528\u3002GQA \u7531 Ainslie et al. 2023 \u63d0\u51fa（\u300aGQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints\u300b），LLaMA-2\/3\u3001Mistral\u3001Qwen \u5168\u7cfb\u91c7\u7528（\u7bc76 \u51fa\u73b0\u7684"W_K \u662f 4096\u00d71024 \u800c\u4e0d\u662f 4096\u00d74096"，\u5c31\u662f GQA \u7684\u75d5\u8ff9\u2014\u2014\u53c2\u6570\u91cf\u4e5f\u968f\u4e4b\u7701\u4e86 4 \u500d）\u3002<\/p>\n
\u66f4\u6fc0\u8fdb\u7684 MQA（Multi-Query，Shazeer 2019）\u53ea\u7559 1 \u7ec4 K\/V（\u7701 32 \u500d），\u4f46\u7cbe\u5ea6\u635f\u5931\u504f\u5927，\u73b0\u5728\u5c11\u7528\u2014\u2014GQA \u662f\u7cbe\u5ea6\u4e0e\u7701\u5b58\u7684\u5e73\u8861\u70b9\u3002<\/p>\n
\u7b2c\u4e8c\u628a\u5200：MLA（\u591a\u5934\u6f5c\u5728\u6ce8\u610f\u529b，Multi-head Latent Attention）\u2014\u2014\u4ece"\u7ef4\u5ea6"\u4e0b\u624b，DeepSeek-V2\/V3 \u7684\u62db\u724c\u8bbe\u8ba1\u3002<\/p>\n
GQA \u53ea\u662f\u628a K\/V \u4ece 32 \u7ec4\u780d\u5230 8 \u7ec4，\u6bcf\u7ec4\u8fd8\u662f 128 \u7ef4\u3002MLA \u66f4\u72e0：\u6839\u672c\u4e0d\u7f13\u5b58 K\/V \u672c\u8eab，\u7f13\u5b58\u4e00\u4e2a 576 \u7ef4\u7684"\u6f5c\u5728\u5411\u91cf"\u3002<\/p>\n
\u539f\u7406\u4e00\u53e5\u8bdd：\u8bad\u7ec3\u65f6\u8ba9\u6a21\u578b\u5b66\u4f1a\u628a 4096 \u7ef4\u7684\u4e0a\u4e0b\u6587\u538b\u7f29\u6210 512 \u7ef4\u6f5c\u5728\u8868\u793a（\u5916\u52a0 64 \u7ef4 RoPE \u4f4d\u7f6e\u4fe1\u606f），\u7528\u65f6\u518d\u89e3\u538b\u56de K\/V\u3002\u7c7b\u6bd4：\u4e0d\u5b58\u6574\u672c\u4e66，\u5b58\u4e00\u672c"\u6458\u8981"\u2014\u2014\u63a8\u7406\u65f6\u6309\u6458\u8981\u91cd\u5efa\u9700\u8981\u7684\u5185\u5bb9\u3002\u4ee3\u4ef7\u662f\u8ba1\u7b97\u591a\u4e00\u6b65"\u89e3\u538b"，\u6362\u6765\u7684\u662f\u6bcf token \u6bcf\u5c42\u7f13\u5b58\u4ece 8192 \u4e2a\u6570\u964d\u5230\u7ea6 576 \u4e2a，\u51cf\u5c11\u7ea6 93%（DeepSeek-V2 \u8bba\u6587\u300aDeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model\u300b\u7684\u5b9e\u6d4b\u53e3\u5f84）\u2014\u2014\u8fd9\u5c31\u662f DeepSeek \u6562\u505a\u8d85\u957f\u4e0a\u4e0b\u6587\u7684\u5e95\u6c14\u4e4b\u4e00\u3002<\/p>\n
\u56db\u79cd\u65b9\u6848\u5bf9\u8d26（\u4ee5 4096 \u5bbd\u300132 \u5934\u3001\u6bcf\u5934 128 \u7ef4\u4e3a\u4f8b，\u6bcf token \u6bcf\u5c42\u7684 KV \u7f13\u5b58\u91cf）：<\/p>\n
\n
\u65b9\u6848K\/V \u5934\u6570\u7f13\u5b58\u7ef4\u5ea6（\u00d72 \u4efd）\u76f8\u5bf9 MHA\u4ee3\u8868\u6a21\u578b<\/tr>\n \n\n MHA<\/td>\n 32 \u7ec4<\/td>\n 2 \u00d7 4096 = 8192<\/td>\n 1\u00d7<\/td>\n \u539f\u59cb Transformer \/ BERT<\/td>\n<\/tr>\n \n MQA<\/td>\n 1 \u7ec4<\/td>\n 2 \u00d7 128 = 256<\/td>\n 1\/32<\/td>\n PaLM \u7b49（\u5c11\u7528）<\/td>\n<\/tr>\n \n GQA-8<\/td>\n 8 \u7ec4<\/td>\n 2 \u00d7 1024 = 2048<\/td>\n 1\/4<\/td>\n LLaMA-3 \/ Mistral \/ Qwen<\/td>\n<\/tr>\n \n MLA<\/td>\n \u6f5c\u5728\u538b\u7f29<\/td>\n \u2248 576 + 64<\/td>\n \u22481\/14（\u51cf 93%）<\/td>\n DeepSeek-V3 \/ R1<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n \u886810-2 \u56db\u79cd\u6ce8\u610f\u529b\u7684 KV-Cache \u5bf9\u6bd4：\u90a3\u4e9b"\u770b\u4e0d\u61c2\u7684\u538b\u7f29\u8bbe\u8ba1"，\u672c\u8d28\u4e0a\u90fd\u662f\u5728\u7ed9 KV-Cache \u7626\u8eab<\/p>\n
8. \u6536\u5b98：\u4ece\u4e00\u6761\u76f4\u7ebf\u5230\u4e00\u4e2a token<\/h3>\n
8.1 \u5341\u7bc7\u65c5\u7a0b<\/h4>\n
\u628a 10 \u7bc7\u4e32\u6210\u4e00\u6761\u7ebf，\u5982\u56fe 10-10 \u6240\u793a\u3002<\/p>\n
![\"\u7cfb\u5217\u6536\u5b98\u5168\u666f\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260828233359-6a921ae73c0ac.png\")
<\/p>\n\u56fe10-10 \u4ece\u4e00\u6761\u76f4\u7ebf\u5230\u4e00\u4e2a token：10 \u7bc7\u7684\u5b8c\u6574\u65c5\u7a0b<\/p>\n
- \n
- \u7bc71 \u7528\u4e00\u6761\u8fc7\u539f\u70b9\u7684\u76f4\u7ebf（1 \u4e2a\u53c2\u6570）\u51fa\u53d1，\u7bc71 \u672b\u52a0\u622a\u8ddd（2 \u4e2a）<\/li>\n
- \u7bc72 \u5206\u8bcd\u2014\u2014\u53e5\u5b50\u53d8\u6210 token \u5e8f\u5217；\u7bc73 \u5d4c\u5165\u67e5\u8868（5.3 \u4ebf）<\/li>\n
- \u7bc74 RoPE；\u7bc75 Block \u516d\u6b65；\u7bc76 QKV；\u7bc77 \u6ce8\u610f\u529b；\u7bc78 FFN\u2014\u2014\u7ec4\u4ef6\u9010\u4e2a\u4e0a\u9f50<\/li>\n
- \u7bc79 \u5806 32 \u5c42 + \u8f93\u51fa\u77e9\u9635（75.1 \u4ebf）；\u7bc710 lm_head + Softmax（80.4 \u4ebf）\u2192 \u300c\u6674\u300d<\/li>\n<\/ul>\n
\u4e00\u53e5\u8bdd\u603b\u7ed3\u7cfb\u5217：\u5927\u6a21\u578b = \u628a\u300c\u4e00\u884c\u4e58\u4e00\u5217\u300d\u8fd9\u4e2a\u52a8\u4f5c，\u5728 80 \u4ebf\u4e2a\u6570\u5b57\u4e0a\u91cd\u590d\u51e0\u4e07\u4ebf\u6b21，\u518d\u8ba9\u4e0b\u4e00\u4e2a token \u4ece\u6982\u7387\u91cc\u8bde\u751f\u3002<\/p>\n
8.2 230B \u7684\u6700\u540e\u4e00\u73af<\/h4>\n
\u7bc79 \u627f\u8bfa\u8fc7：\u7bc710 \u7ed3\u5c3e\u56de\u8865 230B \u7684\u95ed\u73af\u3002\u73b0\u5728\u6240\u6709\u7ec4\u4ef6\u90fd\u5728\u624b\u4e0a\u4e86，\u53ef\u4ee5\u628a\u5b83\u8bf4\u900f：<\/p>\n
\u8c46\u5305 Seed-1.6 \u7684 230B \u548c\u672c\u7cfb\u5217\u7684 80.4 \u4ebf\u5171\u4eab\u5168\u90e8\u539f\u7406\u2014\u2014\u5206\u8bcd\u3001\u5d4c\u5165\u3001RoPE\u3001\u6ce8\u610f\u529b\u3001FFN\u3001\u6b8b\u5dee\u3001\u5806\u53e0\u3001lm_head\u3001Softmax\u3001\u81ea\u56de\u5f52\u3001KV-Cache，\u4e00\u4e2a\u7ec4\u4ef6\u90fd\u4e0d\u591a\u3002\u533a\u522b\u53ea\u5728\u4e09\u4e2a\u65cb\u94ae：<\/p>\n
\n
LLaMA-3-8B（\u672c\u7cfb\u5217\u5bf9\u8d26\u57fa\u51c6）Seed-1.6（\u7bc71 \u89c4\u6a21\u611f\u8d77\u70b9）<\/tr>\n \n\n \u5bbd\u5ea6<\/td>\n 4096<\/td>\n \u7ea6 7168<\/td>\n<\/tr>\n \n \u6df1\u5ea6<\/td>\n 32 \u5c42<\/td>\n \u6570\u5341\u5c42<\/td>\n<\/tr>\n \n FFN<\/td>\n \u6bcf\u5c42 1 \u4e2a<\/td>\n \u6bcf\u5c42\u82e5\u5e72\u4e13\u5bb6，\u6bcf token \u6311\u7740\u6fc0\u6d3b<\/td>\n<\/tr>\n \n \u603b\u53c2\u6570<\/td>\n 8.03B<\/td>\n 230B<\/td>\n<\/tr>\n \n \u5355 token \u6fc0\u6d3b<\/td>\n 8.03B（\u5168\u6fc0\u6d3b）<\/td>\n 23B（\u7a00\u758f\u6fc0\u6d3b）<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n \u886810-3 \u7a20\u5bc6 8B \u4e0e MoE 230B：\u540c\u4e00\u5957\u539f\u7406\u7684\u4e24\u79cd\u914d\u6bd4<\/p>\n
MoE \u7684"230B \u603b\u53c2\u6570\u300123B \u6fc0\u6d3b"，\u7ffb\u8bd1\u8fc7\u6765\u5c31\u662f：\u6a21\u578b\u8bb0\u5f97\u591a（230B \u7684\u77e5\u8bc6\u5b58\u91cf），\u4f46\u6bcf\u6b21\u53ea\u60f3\u5c11（23B \u7684\u8ba1\u7b97\u91cf）\u2014\u2014\u50cf\u4e00\u4e2a\u85cf\u4e66 230 \u4e07\u518c\u7684\u56fe\u4e66\u9986，\u6bcf\u6b21\u67e5\u8be2\u53ea\u52a8 23 \u4e07\u518c\u76f8\u5173\u7684\u4e66\u67b6\u3002KV-Cache \u4f18\u5316\u7684\u4e5f\u662f\u540c\u4e00\u4ef6\u4e8b\u7684\u4e24\u9762：\u77e5\u8bc6\u8981\u591a\u5b58\u3001\u8ba1\u7b97\u8981\u5c11\u505a\u3002<\/p>\n
\u7bc71 \u90a3\u4e2a"\u628a 230B \u780d\u5230 1 \u4e2a\u53c2\u6570"\u7684\u601d\u60f3\u5b9e\u9a8c，\u81f3\u6b64\u5b8c\u5168\u95ed\u5408：\u780d\u6389\u7684\u662f\u89c4\u6a21，\u7559\u4e0b\u7684\u662f\u539f\u7406；\u52a0\u56de\u6765\u7684\u662f\u8d26\u672c，\u770b\u6e05\u7684\u662f\u9aa8\u67b6\u3002<\/p>\n
8.3 \u5199\u5728\u7cfb\u5217\u5b8c\u7ed3<\/h4>\n
\u5199\u8fd9\u4e2a\u7cfb\u5217\u7684\u521d\u8877\u5f88\u7b80\u5355：\u5e02\u9762\u4e0a\u7684 Transformer \u6559\u7a0b，\u8981\u4e48\u4ece\u8bba\u6587\u516c\u5f0f\u5f00\u59cb\u529d\u9000，\u8981\u4e48\u505c\u5728"\u6ce8\u610f\u529b\u5c31\u50cf\u67e5\u5b57\u5178"\u7684\u6bd4\u55bb\u5c42\u9762\u3002\u6211\u60f3\u8bd5\u8bd5\u7b2c\u4e09\u6761\u8def\u2014\u2014\u5047\u8bbe\u8bfb\u8005\u53ea\u5b66\u8fc7\u9ad8\u7b49\u6570\u5b66，\u80fd\u4e0d\u80fd\u628a"\u5927\u6a21\u578b\u8f93\u51fa\u4e00\u4e2a token"\u8fd9\u4ef6\u4e8b，\u4ece y=wx \u4e00\u8def\u63a8\u5230\u5e95\u3002<\/p>\n
10 \u7bc7\u5199\u4e0b\u6765，\u6700\u5927\u7684\u611f\u53d7\u662f：\u5927\u6a21\u578b\u6ca1\u6709\u9b54\u6cd5，\u53ea\u6709\u5de5\u7a0b\u3002\u6240\u8c13\u7684"\u667a\u80fd\u6d8c\u73b0"，\u62c6\u5f00\u770b\u662f\u5d4c\u5165\u67e5\u8868\u3001\u4e00\u884c\u4e58\u4e00\u5217\u3001\u6307\u6570\u5f52\u4e00\u5316\u8fd9\u4e9b\u672c\u79d1\u751f\u90fd\u80fd\u770b\u61c2\u7684\u52a8\u4f5c；\u590d\u6742\u5ea6\u4e0d\u5728\u4efb\u4f55\u5355\u4e2a\u7ec4\u4ef6，\u800c\u5728"\u7b80\u5355\u52a8\u4f5c \u00d7 \u5929\u6587\u6570\u5b57\u7684\u91cd\u590d"\u3002<\/p>\n
\u5982\u679c\u8fd9\u4e2a\u7cfb\u5217\u8ba9\u4f60\u4e0b\u6b21\u518d\u770b\u5230"\u6ce8\u610f\u529b\u5934"\u201cKV-Cache\u201d"MoE \u7a00\u758f\u6fc0\u6d3b"\u8fd9\u4e9b\u8bcd\u65f6，\u8111\u5b50\u91cc\u6d6e\u73b0\u7684\u662f\u5177\u4f53\u7684\u5c0f\u77e9\u9635\u548c\u5b83\u7684\u8ba1\u7b97\u8fc7\u7a0b\u2014\u2014\u8fd9\u4e2a\u7cfb\u5217\u7684\u4efb\u52a1\u5c31\u5b8c\u6210\u4e86\u3002<\/p>\n
\u4ece\u4e00\u6761\u76f4\u7ebf，\u5230\u300c\u6674\u300d\u3002\u8c22\u8c22\u4e00\u8def\u770b\u5230\u8fd9\u91cc\u3002<\/p>\n
9. \u4e0b\u4e00\u7cfb\u5217\u9884\u544a：17 \u4e2a\u95ee\u9898<\/h3>\n
\u5199\u5b8c\u8fd9\u4e2a\u7cfb\u5217，\u6700\u81ea\u7136\u7684\u611f\u53d7\u662f："\u4e00\u4e2a token \u5982\u4f55\u8bde\u751f"\u8bb2\u6e05\u695a\u4e86，\u4f46\u56f4\u7ed5\u5927\u6a21\u578b\u8fd8\u6709\u4e00\u5927\u7247\u6ca1\u8bb2\u7684\u5730\u3002\u987a\u7740\u672c\u7cfb\u5217\u7684\u601d\u8def\u5f80\u4e0b\u95ee，\u81f3\u5c11\u8fd8\u6709 17 \u4e2a\u597d\u95ee\u9898\u3002<\/p>\n
\u5148\u629b\u4e09\u4e2a\u6700\u624e\u5fc3\u7684：<\/p>\n
\u95ee\u9898\u4e00：\u5927\u6a21\u578b\u5176\u5b9e\u4e0d\u77e5\u9053\u4eca\u5929\u662f\u51e0\u53f7\u3002 \u5b83\u7684\u77e5\u8bc6\u51bb\u7ed3\u5728\u8bad\u7ec3\u622a\u6b62\u90a3\u5929\u2014\u2014\u4f60\u95ee"\u4eca\u5929\u897f\u5b89\u7684\u5929\u6c14"，\u5b83\u7b54\u7684"\u6674"\u53ea\u662f\u6982\u7387\u4e0a\u6700\u50cf\u7b54\u6848\u7684\u8bcd，\u4e0d\u662f\u771f\u5929\u6c14\u3002\u600e\u4e48\u8ba9\u5b83\u611f\u77e5"\u4eca\u5929"？\u6700\u76f4\u63a5\u7684\u65b9\u6848\u662f Prompt：\u628a\u65e5\u671f\u3001\u5929\u6c14\u6570\u636e\u585e\u8fdb\u4e0a\u4e0b\u6587（\u672c\u7cfb\u5217\u7684\u8f93\u5165\u53e5\u5c31\u662f\u8fd9\u4e48\u6765\u7684）；\u66f4\u8fdb\u4e00\u6b65\u662f MCP \u5de5\u5177\u8c03\u7528：\u8ba9\u6a21\u578b\u81ea\u5df1\u51b3\u5b9a"\u5148\u67e5\u5929\u6c14 API，\u518d\u56de\u7b54"\u3002<\/p>\n
\u95ee\u9898\u4e8c：\u56de\u7b54\u603b\u662f\u300c\u6674\u300d，\u600e\u4e48\u7834？ \u5c31\u7b97\u63a5\u4e86\u5de5\u5177，\u6a21\u578b\u4e5f\u53ef\u80fd\u5077\u61d2\u6bcf\u6b21\u90fd\u8fd4\u56de\u540c\u4e00\u4e2a\u7b54\u6848\u3002RAG（\u68c0\u7d22\u589e\u5f3a\u751f\u6210）\u4ece\u4f60\u7684\u79c1\u6709\u5927\u6587\u6863\u91cc\u68c0\u7d22\u771f\u5b9e\u6750\u6599\u518d\u56de\u7b54；AI-Memory \u8ba9\u6a21\u578b\u8bb0\u4f4f\u4f60\u662f\u8c01\u3001\u4f60\u4e4b\u524d\u8bf4\u8fc7\u4ec0\u4e48\u2014\u2014\u53d8\u5e9f\u4e3a\u5b9d，\u4ece"\u5783\u573e\u6570\u636e"\u91cc\u63d0\u53d6"\u5b9d\u8d1d\u77e5\u8bc6"\u3002<\/p>\n
\u95ee\u9898\u4e09：\u672c\u7cfb\u5217\u53ea\u7ed9\u4e86 6 \u4e2a token \u7684\u4e0a\u4e0b\u6587，\u771f\u5b9e\u5bf9\u8bdd\u52a8\u8f84\u51e0\u4e07 token\u2014\u2014\u4e0a\u4e0b\u6587\u7206\u70b8\u4e86\u600e\u4e48\u529e？ SubAgent（\u5b50\u4ee3\u7406）\u7684\u601d\u8def：\u4e3b Agent \u53ea\u7559\u6458\u8981，\u7ec6\u8282\u6d3e\u7ed9\u5b50 Agent \u5904\u7406\u3002<\/p>\n
\u628a\u8fd9 17 \u4e2a\u95ee\u9898\u5168\u90e8\u5217\u51fa，\u6309"\u61c2\u4f60 \/ \u6709\u636e \/ \u770b\u89c1\u66f4\u591a \/ \u53d8\u5f3a"\u5206\u6210\u56db\u7ec4，\u5982\u56fe 10-11 \u6240\u793a\u3002<\/p>\n
![\"\u4e0b\u4e00\u7cfb\u5217\u4e3b\u9898\u5730\u56fe\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260828233401-6a921ae9d0637.png\")
<\/p>\n\u56fe10-11 \u4e0b\u4e00\u7cfb\u5217\u5019\u9009\u4e3b\u9898：17 \u4e2a\u95ee\u9898\u5206\u56db\u7ec4（\u987a\u5e8f\u548c\u5185\u5bb9\u90fd\u53ef\u80fd\u8c03\u6574，\u6b22\u8fce\u7559\u8a00\u70b9\u5355）<\/p>\n
- \n
- A \u7ec4\u00b7\u8ba9\u6a21\u578b\u66f4\u61c2\u4f60：01 \u5927\u6a21\u578b\u5982\u4f55\u611f\u77e5\u3010\u4eca\u5929\u3011（Prompt）\u300103 \u5982\u4f55\u8ba9\u5bf9\u8bdd\u66f4\u62df\u4eba（\u89d2\u8272\u626e\u6f14）\u300106 \u8ba9\u5927\u6a21\u578b\u4e86\u89e3\u7ec6\u8282（Skill）\u300108 \u8ba9\u5927\u6a21\u578b\u9002\u914d\u4f60（AI-Memory）<\/li>\n
- B \u7ec4\u00b7\u8ba9\u6a21\u578b\u6709\u636e\u53ef\u4f9d：02 \u56de\u7b54\u603b\u662f\u3010\u6674\u3011\u600e\u4e48\u529e（MCP）\u300109 \u4ece\u5783\u573e\u91cc\u63d0\u53d6\u5b9d\u8d1d（RAG）\u300107 \u4e0a\u4e0b\u6587\u7206\u70b8\u4e86\u600e\u4e48\u529e（SubAgent）<\/li>\n
- C \u7ec4\u00b7\u8ba9\u6a21\u578b\u770b\u89c1\u66f4\u591a：04 \u5927\u6a21\u578b\u5982\u4f55\u770b\u56fe（\u591a\u6a21\u6001-\u56fe\u7247）\u300105 \u5927\u6a21\u578b\u5982\u4f55\u611f\u77e5\u65f6\u95f4（\u591a\u6a21\u6001-\u89c6\u9891）\u300116 \u64cd\u4f5c\u4f60\u7684\u7535\u8111（GUI\/Computer Use）\u300115 \u591a\u4e2a\u5927\u6a21\u578b\u534f\u4f5c（Multi-Agent）\u300117 Agent \u793e\u4f1a（\u65af\u5766\u798f AI \u5c0f\u9547）<\/li>\n
- D \u7ec4\u00b7\u8ba9\u6a21\u578b\u53d8\u5f3a：10 \u77e5\u8bc6\u4ece\u54ea\u6765（\u9884\u8bad\u7ec3）\u300111 \u5b66\u4f1a\u542c\u6307\u4ee4（SFT）\u300112 \u5b66\u4f1a\u504f\u597d（RL）\u300113 \u81ea\u5df1\u53d8\u5f3a（\u81ea\u8fdb\u5316）\u300114 \u5feb\u601d\u8003\u4e0e\u6162\u601d\u8003（\u63a8\u7406\u6a21\u578b）<\/li>\n<\/ul>\n
A\/B \u7ec4\u56de\u7b54"\u600e\u4e48\u7528\u597d\u4e00\u4e2a\u5927\u6a21\u578b"（\u5e94\u7528\u5c42），C \u7ec4\u56de\u7b54"\u600e\u4e48\u6269\u5c55\u5b83\u7684\u611f\u5b98\u4e0e\u624b\u8db3"（\u591a\u6a21\u6001\u4e0e Agent），D \u7ec4\u56de\u7b54"\u600e\u4e48\u70bc\u6210\u4e00\u4e2a\u5927\u6a21\u578b"（\u8bad\u7ec3\u5c42）\u3002\u672c\u7cfb\u5217（\u539f\u7406\u7bc7）\u56de\u7b54\u7684\u662f"\u5b83\u600e\u4e48\u5de5\u4f5c"\u2014\u2014\u56db\u4e2a\u7cfb\u5217\u8fde\u8d77\u6765，\u5c31\u662f\u4ece\u539f\u7406\u5230\u5e94\u7528\u7684\u5b8c\u6574\u5730\u56fe\u3002<\/p>\n
\u987a\u5e8f\u548c\u5185\u5bb9\u90fd\u53ef\u80fd\u8c03\u6574\u3002\u60f3\u5148\u770b\u54ea\u4e2a？\u8bc4\u8bba\u533a\u70b9\u5355\u3002<\/p>\n
\u5c0f\u7ed3<\/h3>\n
\u8fd9\u4e00\u7bc7\u8ba9\u9996 token \u771f\u6b63\u8bde\u751f，\u7cfb\u5217\u6536\u5b98：<\/p>\n
- \n
- \u53d6\u6700\u540e\u4e00\u884c：\u56e0\u679c\u8bed\u8a00\u6a21\u578b\u53ea\u6709\u6700\u540e\u4e00\u4e2a token（\u300c？\u300d）\u89c1\u8fc7\u5168\u90e8\u524d\u6587，\u5b83\u7684\u884c\u5411\u91cf v（1\u00d74096）\u662f\u9884\u6d4b\u9996 token \u7684\u552f\u4e00\u539f\u6599；d1=2.85 \u6b63\u662f\u7bc79 \u7684"\u8be5\u4f5c\u7b54"\u5f00\u5173<\/li>\n
- lm_head：v \u00d7 W_L（4096\u00d7129,280 = 5.3 \u4ebf\u53c2\u6570）= 129,280 \u4e2a logits\u2014\u2014\u5168\u7cfb\u5217\u6700\u540e\u4e00\u6b21\u300c\u4e00\u884c\u4e58\u4e00\u5217\u300d，\u6f14\u793a 8 \u4e2a\u5019\u9009\u4e2d\u300c\u6674\u300d2.20 \u6700\u9ad8<\/li>\n
- Softmax：\u624b\u7b97\u4e09\u6b65（\u9010\u683c e^z \u2192 \u6c42\u548c 20.91 \u2192 \u76f8\u9664）\u2014\u2014\u6307\u6570\u653e\u5927\u5dee\u8ddd，\u300c\u6674\u300d9.03\/20.91 = 43.2% \u4e00\u5bb6\u72ec\u5927；\u7bc79 \u7684 6 \u6761\u7ea6\u675f\u94fe\u5728\u6982\u7387\u4e0a\u5168\u90e8\u5151\u73b0（"\u4eca"\u4ec5 5.7% \u56e0\u4e3a\u65f6\u95f4\u5df2\u4ea4\u4ee3\u3001"\u5f88"\u57ab\u5e95\u56e0\u4e3a\u8bcd\u6027\u4e0d\u7b26）<\/li>\n
- \u9009 token \u4e09\u7b56\u7565：\u8d2a\u5fc3（\u786e\u5b9a）\u3001top-k（\u786c\u95e8\u69db）\u3001top-p（\u81ea\u9002\u5e94\u622a\u65ad）；\u6e29\u5ea6\u662f\u5206\u5e03\u96c6\u4e2d\u5ea6\u65cb\u94ae（T=0.5\u219278.6%，T=2.0\u219225.0%）<\/li>\n
- \u9996 token =\u300c\u6674\u300d：\u4ece\u5206\u8bcd\u5230\u8f93\u51fa，10 \u7bc7\u5168\u94fe\u8def\u7b2c\u4e00\u6b21\u5b8c\u6574\u8dd1\u901a<\/li>\n
- \u81ea\u56de\u5f52：\u8f93\u51fa\u62fc\u56de\u8f93\u5165，\u9010 token \u751f\u6210\u5b8c\u6574\u56de\u7b54；\u6bcf\u6b65\u90fd\u662f\u5168\u91cf\u524d\u5411<\/li>\n
- KV-Cache：\u56e0\u679c\u63a9\u7801\u4fdd\u8bc1\u65e7 K\/V \u6c38\u4e0d\u53d8 \u2192 \u7f13\u5b58\u590d\u7528\u3001\u53ea\u7b97\u65b0 token，\u957f\u56de\u7b54\u4ece\u5e73\u65b9\u7ea7\u964d\u56de\u8fd1\u7ebf\u6027；\u4ee3\u4ef7\u662f\u663e\u5b58（2\u00d7\u5c42\u6570\u00d7\u957f\u5ea6\u00d7\u5bbd\u5ea6）<\/li>\n
- GQA \u4e0e MLA（\u7626\u8eab\u4e24\u628a\u5200）：GQA \u4ece\u5934\u6570\u4e0b\u624b（32 Q \u5934\u5171\u4eab 8 \u7ec4 K\/V，\u7f13\u5b58\u7701 4 \u500d，LLaMA-3\/Mistral\/Qwen \u5168\u7cfb）；MLA \u4ece\u7ef4\u5ea6\u4e0b\u624b（\u7f13\u5b58 576 \u7ef4\u6f5c\u5728\u5411\u91cf\u800c\u975e K\/V \u672c\u8eab，\u51cf\u7ea6 93%，DeepSeek \u62db\u724c）；MHA\/MQA\/GQA\/MLA \u56db\u65b9\u6848\u5bf9\u8d26<\/li>\n
- 230B \u95ed\u73af：Seed-1.6 \u4e0e LLaMA-3-8B \u5171\u4eab\u5168\u90e8\u539f\u7406，\u533a\u522b\u53ea\u662f\u5bbd\u5ea6\/\u6df1\u5ea6\/MoE \u4e09\u65cb\u94ae\u2014\u2014\u201c\u77e5\u8bc6\u591a\u5b58\u3001\u8ba1\u7b97\u5c11\u505a\u201d<\/li>\n
- \u4e0b\u4e00\u7cfb\u5217\u9884\u544a：17 \u4e2a\u95ee\u9898\u5206\u56db\u7ec4\u2014\u2014\u61c2\u4f60（Prompt\/\u89d2\u8272\u626e\u6f14\/Skill\/Memory）\u3001\u6709\u636e（MCP\/RAG\/SubAgent）\u3001\u770b\u89c1\u66f4\u591a（\u591a\u6a21\u6001\/Agent\/GUI\/AI \u5c0f\u9547）\u3001\u53d8\u5f3a（\u9884\u8bad\u7ec3\/SFT\/RL\/\u81ea\u8fdb\u5316\/\u5feb\u6162\u601d\u8003）<\/li>\n<\/ul>\n
\u7cfb\u5217\u5b8c\u7ed3<\/h3>\n
\u539f\u7406\u7bc7\u5230\u6b64\u5b8c\u7ed3，\u5e94\u7528\u4e0e\u8bad\u7ec3\u7bc7（17 \u4e2a\u5019\u9009\u4e3b\u9898，\u89c1\u7b2c 9 \u8282）\u62e9\u65e5\u5f00\u5199\u3002<\/p>\n
\u7cfb\u5217\u76ee\u5f55（\u5168 10 \u7bc7）：<\/p>\n
- \u4ece\u4e00\u6761\u76f4\u7ebf\u5230\u9ad8\u7ef4\u7a7a\u95f4<\/li>\n
- \u5206\u8bcd\u4e0e token \u8868<\/li>\n
- \u9690\u85cf\u5c42\u4e0e\u5d4c\u5165<\/li>\n
- \u4f4d\u7f6e\u7f16\u7801 RoPE<\/li>\n
- Transformer Block \u5168\u666f\u4e0e\u5c42\u5f52\u4e00\u5316<\/li>\n
- QKV \u4e09\u5251\u5ba2<\/li>\n
- \u6ce8\u610f\u529b\u6743\u91cd\u4e0e\u591a\u5934\u673a\u5236<\/li>\n
- \u6b8b\u5dee\u8fde\u63a5\u4e0e\u524d\u9988\u7f51\u7edc<\/li>\n
- \u8f93\u51fa\u77e9\u9635\u4e0e\u591a\u5c42\u5806\u53e0<\/li>\n
- \u9996 token \u8bde\u751f\u4e0e KV-Cache（\u5f53\u524d\u7bc7，\u5b8c\u7ed3）<\/li>\n
\n\u7248\u6743\u58f0\u660e：\u672c\u6587\u4e3a\u535a\u4e3b\u539f\u521b\u6587\u7ae0，\u9075\u5faa CC 4.0 BY-SA \u7248\u6743\u534f\u8bae，\u8f6c\u8f7d\u8bf7\u9644\u4e0a\u539f\u6587\u51fa\u5904\u94fe\u63a5\u548c\u672c\u58f0\u660e\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"
\u4ece\u4e00\u6761\u76f4\u7ebf\u5230\u5927\u6a21\u578b\u8f93\u51fa\u4e00\u4e2atoken（\u5341）：\u9996 token \u8bde\u751f\u4e0e KV-Cache \u5efa\u8bae\u5148\u770b：\u4ece\u4e00\u6761\u76f4\u7ebf\u5230\u5927\u6a21\u578b\u8f93\u51fa\u4e00\u4e2atoken（\u4e5d）：\u8f93\u51fa\u77e9\u9635\u4e0e\u591a\u5c42\u5806\u53e0 \u4e0a\u4e00\u7bc7\u62ff\u5230\u4e86\u8f93\u51fa\u77e9\u9635，6 \u6761\u7ea6\u675f\u94fe\u5168\u90e8\u6c89\u6dc0\u5728\u6700\u540e\u4e00\u884c\u3002\u8fd9\u4e00\u7bc7\u8ba9\u9996 token \u771f\u6b63\u8bde\u751f\u2014\u2014\u7136\u540e\u662f\u81ea\u56de\u5f52\u3001KV-Cache，\u4ee5\u53ca\u6574\u4e2a\u7cfb\u5217\u7684\u6536\u5b98\u3002 \u4ece\u4e00\u6761\u76f4\u7ebf\u5230\u5927\u6a21\u578b\u8f93\u51fa\u4e00\u4e2atoken（<\/p>\n","protected":false},"author":2,"featured_media":97219,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[10187,10782,11153,11154,66,841],"topic":[],"class_list":["post-97230","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-server","tag-codingplan","tag-kv-cache","tag-softmax","tag-11154","tag-ai","tag-transformer"],"yoast_head":"\n
\u4ece\u4e00\u6761\u76f4\u7ebf\u5230\u5927\u6a21\u578b\u8f93\u51fa\u4e00\u4e2atoken\uff08\u5341\uff09\uff1a\u9996 token \u8bde\u751f\u4e0e KV-Cache - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.wsisp.com\/helps\/97230.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"\u4ece\u4e00\u6761\u76f4\u7ebf\u5230\u5927\u6a21\u578b\u8f93\u51fa\u4e00\u4e2atoken\uff08\u5341\uff09\uff1a\u9996 token \u8bde\u751f\u4e0e KV-Cache - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"\u4ece\u4e00\u6761\u76f4\u7ebf\u5230\u5927\u6a21\u578b\u8f93\u51fa\u4e00\u4e2atoken（\u5341）：\u9996 token \u8bde\u751f\u4e0e KV-Cache \u5efa\u8bae\u5148\u770b：\u4ece\u4e00\u6761\u76f4\u7ebf\u5230\u5927\u6a21\u578b\u8f93\u51fa\u4e00\u4e2atoken（\u4e5d）：\u8f93\u51fa\u77e9\u9635\u4e0e\u591a\u5c42\u5806\u53e0 \u4e0a\u4e00\u7bc7\u62ff\u5230\u4e86\u8f93\u51fa\u77e9\u9635，6 \u6761\u7ea6\u675f\u94fe\u5168\u90e8\u6c89\u6dc0\u5728\u6700\u540e\u4e00\u884c\u3002\u8fd9\u4e00\u7bc7\u8ba9\u9996 token \u771f\u6b63\u8bde\u751f\u2014\u2014\u7136\u540e\u662f\u81ea\u56de\u5f52\u3001KV-Cache，\u4ee5\u53ca\u6574\u4e2a\u7cfb\u5217\u7684\u6536\u5b98\u3002 \u4ece\u4e00\u6761\u76f4\u7ebf\u5230\u5927\u6a21\u578b\u8f93\u51fa\u4e00\u4e2atoken（\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/97230.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-08-28T23:34:06+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260828233341-6a921ad54728a.png\" \/>\n<meta name=\"author\" content=\"admin\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"\u4f5c\u8005\" \/>\n\t<meta name=\"twitter:data1\" content=\"admin\" \/>\n\t<meta name=\"twitter:label2\" content=\"\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4\" \/>\n\t<meta name=\"twitter:data2\" content=\"7 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/97230.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/97230.html\",\"name\":\"\u4ece\u4e00\u6761\u76f4\u7ebf\u5230\u5927\u6a21\u578b\u8f93\u51fa\u4e00\u4e2atoken\uff08\u5341\uff09\uff1a\u9996 token \u8bde\u751f\u4e0e KV-Cache - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-08-28T23:34:06+00:00\",\"dateModified\":\"2026-08-28T23:34:06+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/97230.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/97230.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/97230.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u4ece\u4e00\u6761\u76f4\u7ebf\u5230\u5927\u6a21\u578b\u8f93\u51fa\u4e00\u4e2atoken\uff08\u5341\uff09\uff1a\u9996 token \u8bde\u751f\u4e0e KV-Cache\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\",\"url\":\"https:\/\/www.wsisp.com\/helps\/\",\"name\":\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"description\":\"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}\"},\"query-input\":\"required name=search_term_string\"}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\",\"name\":\"admin\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"contentUrl\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"caption\":\"admin\"},\"sameAs\":[\"http:\/\/wp.wsisp.com\"],\"url\":\"https:\/\/www.wsisp.com\/helps\/author\/admin\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\u4ece\u4e00\u6761\u76f4\u7ebf\u5230\u5927\u6a21\u578b\u8f93\u51fa\u4e00\u4e2atoken\uff08\u5341\uff09\uff1a\u9996 token \u8bde\u751f\u4e0e KV-Cache - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.wsisp.com\/helps\/97230.html","og_locale":"zh_CN","og_type":"article","og_title":"\u4ece\u4e00\u6761\u76f4\u7ebf\u5230\u5927\u6a21\u578b\u8f93\u51fa\u4e00\u4e2atoken\uff08\u5341\uff09\uff1a\u9996 token \u8bde\u751f\u4e0e KV-Cache - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"\u4ece\u4e00\u6761\u76f4\u7ebf\u5230\u5927\u6a21\u578b\u8f93\u51fa\u4e00\u4e2atoken（\u5341）：\u9996 token \u8bde\u751f\u4e0e KV-Cache \u5efa\u8bae\u5148\u770b：\u4ece\u4e00\u6761\u76f4\u7ebf\u5230\u5927\u6a21\u578b\u8f93\u51fa\u4e00\u4e2atoken（\u4e5d）：\u8f93\u51fa\u77e9\u9635\u4e0e\u591a\u5c42\u5806\u53e0 \u4e0a\u4e00\u7bc7\u62ff\u5230\u4e86\u8f93\u51fa\u77e9\u9635，6 \u6761\u7ea6\u675f\u94fe\u5168\u90e8\u6c89\u6dc0\u5728\u6700\u540e\u4e00\u884c\u3002\u8fd9\u4e00\u7bc7\u8ba9\u9996 token \u771f\u6b63\u8bde\u751f\u2014\u2014\u7136\u540e\u662f\u81ea\u56de\u5f52\u3001KV-Cache，\u4ee5\u53ca\u6574\u4e2a\u7cfb\u5217\u7684\u6536\u5b98\u3002 \u4ece\u4e00\u6761\u76f4\u7ebf\u5230\u5927\u6a21\u578b\u8f93\u51fa\u4e00\u4e2atoken（","og_url":"https:\/\/www.wsisp.com\/helps\/97230.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-08-28T23:34:06+00:00","og_image":[{"url":"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/08\/20260828233341-6a921ad54728a.png"}],"author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"7 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/97230.html","url":"https:\/\/www.wsisp.com\/helps\/97230.html","name":"\u4ece\u4e00\u6761\u76f4\u7ebf\u5230\u5927\u6a21\u578b\u8f93\u51fa\u4e00\u4e2atoken\uff08\u5341\uff09\uff1a\u9996 token \u8bde\u751f\u4e0e KV-Cache - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-08-28T23:34:06+00:00","dateModified":"2026-08-28T23:34:06+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/97230.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/97230.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/97230.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"\u4ece\u4e00\u6761\u76f4\u7ebf\u5230\u5927\u6a21\u578b\u8f93\u51fa\u4e00\u4e2atoken\uff08\u5341\uff09\uff1a\u9996 token \u8bde\u751f\u4e0e KV-Cache"}]},{"@type":"WebSite","@id":"https:\/\/www.wsisp.com\/helps\/#website","url":"https:\/\/www.wsisp.com\/helps\/","name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","description":"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}"},"query-input":"required name=search_term_string"}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41","name":"admin","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/","url":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","contentUrl":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","caption":"admin"},"sameAs":["http:\/\/wp.wsisp.com"],"url":"https:\/\/www.wsisp.com\/helps\/author\/admin"}]}},"_links":{"self":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/97230","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/comments?post=97230"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/97230\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media\/97219"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=97230"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=97230"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=97230"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=97230"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}