{"id":113,"date":"2026-06-07T13:27:13","date_gmt":"2026-06-07T13:27:13","guid":{"rendered":"https:\/\/blog.joezhou.top\/113-2\/"},"modified":"2026-06-07T13:27:13","modified_gmt":"2026-06-07T13:27:13","slug":"113-2","status":"publish","type":"post","link":"https:\/\/www.joezhou.top\/?p=113","title":{"rendered":"\u5927\u6a21\u578b\u63a8\u7406\u4f18\u5316\u6280\u672f\uff1a\u4ece\u91cf\u5316\u5230\u670d\u52a1\u90e8\u7f72\u7684\u5de5\u7a0b\u5b9e\u8df5"},"content":{"rendered":"<h2>\u4e00\u3001\u5f15\u8a00<\/h2>\n<p>\u968f\u7740\u5927\u8bed\u8a00\u6a21\u578b\u53c2\u6570\u89c4\u6a21\u4ece\u6570\u5341\u4ebf\u8dc3\u5347\u81f3\u4e07\u4ebf\u7ea7\u522b\uff0c\u6a21\u578b&quot;\u80fd\u505a\u4ec0\u4e48&quot;\u5df2\u7ecf\u4e0d\u662f\u74f6\u9888\uff0c&quot;\u600e\u4e48\u7528\u5f97\u8d77&quot;\u6210\u4e86\u771f\u6b63\u7684\u6311\u6218\u3002\u90e8\u7f72\u4e00\u4e2a70B\u53c2\u6570\u7684Llama 2\u6a21\u578b\uff0c\u5982\u679c\u7528FP16\u7cbe\u5ea6\u63a8\u7406\uff0c\u4ec5\u52a0\u8f7d\u6a21\u578b\u6743\u91cd\u5c31\u9700\u8981\u7ea6140GB\u663e\u5b58\u2014\u2014\u8fd9\u610f\u5473\u7740\u5373\u4f7f\u5355\u5f20H100\uff0880GB\uff09\u4e5f\u65e0\u6cd5\u8fd0\u884c\uff0c\u81f3\u5c11\u9700\u8981\u4e24\u5f20\u9ad8\u7aefGPU\u3002\u5bf9\u4e8e\u5927\u591a\u6570\u4f01\u4e1a\u548c\u5f00\u53d1\u8005\u6765\u8bf4\uff0c\u8fd9\u6837\u7684\u786c\u4ef6\u95e8\u69db\u51e0\u4e4e\u662f\u4e0d\u53ef\u627f\u53d7\u7684\u3002<\/p>\n<p>\u6b63\u56e0\u5982\u6b64\uff0c\u5927\u6a21\u578b\u63a8\u7406\u4f18\u5316\u5728\u8fc7\u53bb\u4e24\u5e74\u6210\u4e3aAI\u5de5\u7a0b\u5316\u9886\u57df\u6700\u706b\u70ed\u7684\u65b9\u5411\u4e4b\u4e00\u3002\u4ece\u91cf\u5316\u6280\u672f\u5230KV Cache\u4f18\u5316\uff0c\u4ece\u8ba1\u7b97\u56fe\u526a\u88c1\u5230\u670d\u52a1\u6846\u67b6\u8c03\u5ea6\uff0c\u6574\u4e2a\u884c\u4e1a\u6b63\u5728\u7ecf\u5386\u4e00\u573a&quot;\u8ba9\u5927\u6a21\u578b\u98de\u5165\u5bfb\u5e38\u767e\u59d3\u5bb6&quot;\u7684\u8fdb\u7a0b\u3002\u672c\u6587\u5c06\u4ece\u5de5\u7a0b\u5b9e\u8df5\u7684\u89d2\u5ea6\uff0c\u7cfb\u7edf\u68b3\u7406\u5927\u6a21\u578b\u63a8\u7406\u4f18\u5316\u7684\u6838\u5fc3\u6280\u672f\u8def\u7ebf\u3001\u5404\u5bb6\u65b9\u6848\u7684\u9009\u578b\u5bf9\u6bd4\uff0c\u4ee5\u53ca\u5b9e\u9645\u90e8\u7f72\u4e2d\u7684\u7ecf\u9a8c\u6559\u8bad\u3002<\/p>\n<h2>\u4e8c\u3001\u4e3a\u4ec0\u4e48\u5927\u6a21\u578b\u63a8\u7406\u8fd9\u4e48&quot;\u8d35&quot;<\/h2>\n<p>\u7406\u89e3\u4f18\u5316\u65b9\u6848\u4e4b\u524d\uff0c\u5148\u8981\u7406\u89e3\u7b97\u529b\u90fd\u6d88\u8017\u5728\u54ea\u91cc\u3002\u5927\u6a21\u578b\u63a8\u7406\u53ef\u4ee5\u5206\u89e3\u4e3a\u4e24\u4e2a\u6838\u5fc3\u9636\u6bb5\uff1a<\/p>\n<h3>2.1 Prefill\uff08\u9884\u586b\u5145\uff09\u9636\u6bb5<\/h3>\n<p>\u5f53\u7528\u6237\u8f93\u5165\u63d0\u793a\u8bcd\uff08prompt\uff09\u65f6\uff0c\u6a21\u578b\u9700\u8981\u5e76\u884c\u8ba1\u7b97\u6240\u6709\u8f93\u5165token\u7684\u6ce8\u610f\u529b\uff0c\u8ba1\u7b97\u51fa\u7b2c\u4e00\u4e2a\u8f93\u51fatoken\u3002\u8fd9\u4e2a\u9636\u6bb5\u5c5e\u4e8e\u8ba1\u7b97\u5bc6\u96c6\u578b\u2014\u2014\u74f6\u9888\u5728\u77e9\u9635\u4e58\u6cd5\u7b97\u529b\uff08FLOPs\uff09\u3002\u6279\u91cf\u8d8a\u5927\u7684\u8f93\u5165\uff0c\u8ba1\u7b97\u6548\u7387\u8d8a\u9ad8\u3002<\/p>\n<h3>2.2 Decode\uff08\u89e3\u7801\uff09\u9636\u6bb5<\/h3>\n<p>\u751f\u6210\u7b2c\u4e00\u4e2atoken\u540e\uff0c\u6a21\u578b\u9010token\u751f\u6210\u540e\u7eed\u5185\u5bb9\u3002\u6bcf\u4e2atoken\u751f\u6210\u9700\u8981\uff1a\u5c06\u5f53\u524dtoken\u7684\u9690\u85cf\u72b6\u6001\u4e0e\u6240\u6709\u5386\u53f2token\u7684Key\/Value\u505a\u6ce8\u610f\u529b\u8ba1\u7b97\u3002\u8fd9\u4e2a\u9636\u6bb5\u5c5e\u4e8e\u8bbf\u5b58\u5bc6\u96c6\u578b\u2014\u2014\u74f6\u9888\u5728\u4e8e\u4ece\u663e\u5b58\u4e2d\u8bfb\u53d6KV Cache\u7684\u901f\u5ea6\u3002<\/p>\n<p>KV Cache\u5c31\u662f\u8fd9\u4e2a&quot;\u5386\u53f2\u8bb0\u5f55&quot;\u3002\u5047\u8bbe\u5e8f\u5217\u957f\u5ea6\u4e3a4096\u3001\u9690\u5c42\u7ef4\u5ea6\u4e3a4096\uff0c\u5bf9\u4e8e70B\u6a21\u578b\uff08\u6570\u5341\u5c42\uff09\uff0cKV Cache\u5360\u7528\u7ea6\u4e3a\uff1a<\/p>\n<ul>\n<li>\u6bcf\u5c42\uff1a2 \u00d7 4096 \u00d7 4096 \u00d7 2\u5b57\u8282\uff08FP16\uff09= 64MB<\/li>\n<li>\u517180\u5c42\uff1a\u7ea65GB<\/li>\n<\/ul>\n<p>\u8fd9\u8fd8\u53ea\u662f<strong>\u4e00\u4e2a\u8bf7\u6c42<\/strong>\u7684KV Cache\u3002\u5982\u679c\u670d\u52a1\u8981\u540c\u65f6\u5904\u7406\u6570\u5341\u4e2a\u5e76\u53d1\u8bf7\u6c42\uff0cKV Cache\u7684\u663e\u5b58\u5360\u7528\u4f1a\u8fc5\u901f\u6210\u4e3a\u74f6\u9888\u3002\u8fd9\u5c31\u662f\u4e3a\u4ec0\u4e48\u5927\u6a21\u578b\u63a8\u7406\u65e2&quot;\u5403&quot;\u7b97\u529b\u53c8&quot;\u5403&quot;\u663e\u5b58\u3002<\/p>\n<h2>\u4e09\u3001\u91cf\u5316\u6280\u672f\uff1a\u7528\u7cbe\u5ea6\u6362\u6548\u7387<\/h2>\n<p>\u91cf\u5316\u662f\u76ee\u524d\u6700\u6210\u719f\u3001\u5e94\u7528\u6700\u5e7f\u6cdb\u7684\u63a8\u7406\u4f18\u5316\u624b\u6bb5\u3002\u5b83\u7684\u6838\u5fc3\u601d\u60f3\u5f88\u7b80\u5355\uff1a\u7528\u66f4\u5c11\u7684\u6bd4\u7279\u4f4d\u6765\u8868\u793a\u6a21\u578b\u6743\u91cd\u548c\u6fc0\u6d3b\u503c\u3002<\/p>\n<h3>3.1 \u6743\u91cd\u91cf\u5316\uff08Weight-Only Quantization\uff09<\/h3>\n<p>\u53ea\u5bf9\u6a21\u578b\u6743\u91cd\u8fdb\u884c\u91cf\u5316\uff0c\u63a8\u7406\u65f6\u5148\u5c06\u91cf\u5316\u6743\u91cd\u53cd\u91cf\u5316\u56deFP16\u518d\u8fdb\u884c\u8ba1\u7b97\u3002<\/p>\n<p><strong>INT8\u6743\u91cd\u91cf\u5316\uff1a<\/strong><\/p>\n<ul>\n<li>\u6a21\u578b\u5927\u5c0f\u51cf\u534a\uff08\u4ee570B\u4e3a\u4f8b\uff1a140GB \u2192 70GB\uff09<\/li>\n<li>\u63a8\u7406\u901f\u5ea6\u63d0\u5347\u4e0d\u660e\u663e\uff08\u8ba1\u7b97\u4ecd\u662fFP16\uff09\uff0c\u4f46\u53ef\u4ee5\u5bb9\u7eb3\u66f4\u5927\u7684\u6a21\u578b<\/li>\n<li>\u7cbe\u5ea6\u635f\u5931\u6781\u5c0f\uff0c\u5927\u90e8\u5206\u6a21\u578b\u51e0\u4e4e\u65e0\u611f<\/li>\n<\/ul>\n<p><strong>INT4\u6743\u91cd\u91cf\u5316\uff1a<\/strong><\/p>\n<ul>\n<li>\u6a21\u578b\u5927\u5c0f\u964d\u81f3\u539f\u6765\u76841\/4\uff0870B\uff1a140GB \u2192 35GB\uff09<\/li>\n<li>\u4e00\u5f20A100 80GB\u53ef\u4ee5\u8dd170B\u6a21\u578b\uff0c\u751a\u81f3\u5355\u5f204090 24GB\u53ef\u8dd1Qwen 32B<\/li>\n<li>\u7cbe\u5ea6\u635f\u5931\u5f00\u59cb\u53d8\u5f97\u663e\u8457\uff0c\u9700\u8981\u6ce8\u610f\u6821\u51c6\u6570\u636e\u96c6\u7684\u9009\u62e9<\/li>\n<\/ul>\n<p><strong>\u4e3b\u6d41\u65b9\u6848\u5bf9\u6bd4\uff1a<\/strong><\/p>\n<table>\n<thead>\n<tr>\n<th>\u65b9\u6848<\/th>\n<th>\u7cbe\u5ea6\u7ea7\u522b<\/th>\n<th>\u786c\u4ef6\u8981\u6c42<\/th>\n<th>\u7cbe\u5ea6\u4fdd\u7559<\/th>\n<th>\u9002\u7528\u573a\u666f<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>GPTQ<\/td>\n<td>4-bit<\/td>\n<td>\u4efb\u4f55GPU<\/td>\n<td>\u4f18\u79c0<\/td>\n<td>\u79bb\u7ebf\u91cf\u5316\u7684\u90e8\u7f72\u573a\u666f<\/td>\n<\/tr>\n<tr>\n<td>AWQ<\/td>\n<td>4-bit<\/td>\n<td>\u4efb\u4f55GPU<\/td>\n<td>\u66f4\u4f18<\/td>\n<td>\u5bf9\u7cbe\u5ea6\u654f\u611f\u7684\u90e8\u7f72\u573a\u666f<\/td>\n<\/tr>\n<tr>\n<td>GGML\/GGUF<\/td>\n<td>2~8-bit<\/td>\n<td>CPU\u4f18\u5148<\/td>\n<td>\u826f\u597d<\/td>\n<td>\u672c\u5730\/\u8fb9\u7f18\u8bbe\u5907\u90e8\u7f72<\/td>\n<\/tr>\n<tr>\n<td>bitsandbytes<\/td>\n<td>4\/8-bit<\/td>\n<td>NVIDIA GPU<\/td>\n<td>\u826f\u597d<\/td>\n<td>\u5feb\u901f\u539f\u578b\u9a8c\u8bc1<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p><strong>\u5b9e\u6218\u7ecf\u9a8c\uff1a<\/strong> \u5982\u679c\u8ffd\u6c42\u6781\u81f4\u7cbe\u5ea6\u4fdd\u7559\uff0c\u63a8\u8350AWQ + \u6821\u51c6\u6570\u636e\u96c6\u5bf9\u9f50\u4e0b\u6e38\u4efb\u52a1\u3002\u5982\u679c\u76ee\u6807\u662f\u5c3d\u53ef\u80fd\u6269\u5927\u6a21\u578b\u89c4\u6a21\uff0cGPTQ 4-bit\u914d\u5408\u5206\u7ec4\u5927\u5c0f(group size) 128\u662f\u6700\u7a33\u59a5\u7684\u9009\u62e9\u3002<\/p>\n<h3>3.2 \u91cf\u5316\u611f\u77e5\u8bad\u7ec3\uff08QAT\uff09<\/h3>\n<p>\u76f8\u6bd4\u8bad\u7ec3\u540e\u91cf\u5316(PTQ)\uff0cQAT\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u6a21\u62df\u91cf\u5316\u8bef\u5dee\uff0c\u8ba9\u6a21\u578b\u4e3b\u52a8\u9002\u5e94\u4f4e\u7cbe\u5ea6\u8868\u793a\u3002\u901a\u5e38\u80fd\u83b7\u5f97\u6bd4PTQ\u9ad81~2\u4e2a\u767e\u5206\u70b9\u7684\u57fa\u51c6\u5206\u6570\uff0c\u7f3a\u70b9\u662f\u8bad\u7ec3\u6210\u672c\u8f83\u9ad8\u3002<\/p>\n<p>\u4e00\u822c\u63a8\u8350\u6d41\u7a0b\uff1a\u5148\u7528FP16\u8dd1\u5b8c\u6a21\u578b\u8bad\u7ec3\uff0c\u518d\u505aPTQ\u5feb\u901f\u9a8c\u8bc1\u6548\u679c\uff0c\u5982\u679c\u7cbe\u5ea6\u635f\u5931\u4e0d\u53ef\u63a5\u53d7\uff0c\u518d\u542f\u7528QAT\u5fae\u8c03\u3002<\/p>\n<h3>3.3 \u6fc0\u6d3b\u91cf\u5316<\/h3>\n<p>\u9664\u4e86\u6743\u91cd\u4e4b\u5916\uff0c\u5bf9\u6fc0\u6d3b\u503c(activation)\u505a\u91cf\u5316\u53ef\u4ee5\u8fdb\u4e00\u6b65\u63d0\u5347\u8ba1\u7b97\u6548\u7387\u3002FP8\uff088\u4f4d\u6d6e\u70b9\u6570\uff09\u662fNVIDIA Hopper\u67b6\u6784\u7684\u539f\u751f\u652f\u6301\u683c\u5f0f\u3002\u5728H100\u4e0a\u5229\u7528FP8\u505a\u63a8\u7406\uff0c\u76f8\u6bd4FP16\u5728\u77e9\u9635\u8ba1\u7b97\u4e0a\u53ef\u83b7\u5f97\u7ea62\u500d\u7684\u541e\u5410\u63d0\u5347\u3002<\/p>\n<p>\u4e0d\u8fc7\u6fc0\u6d3b\u91cf\u5316\u6bd4\u6743\u91cd\u91cf\u5316\u66f4&quot;\u654f\u611f&quot;\uff0c\u56e0\u4e3a\u6fc0\u6d3b\u503c\u4f1a\u56e0\u4e3a\u8f93\u5165\u5206\u5e03\u4e0d\u540c\u800c\u5927\u5e45\u6ce2\u52a8\u3002\u5b9e\u8df5\u4e2d\u9700\u8981\u505a\u52a8\u6001\u91cf\u5316\uff08\u6bcftoken\/\u6bcf\u5c42\u8ba1\u7b97scale\uff09\u6216\u9759\u6001\u91cf\u5316\uff08\u7528\u6821\u51c6\u6570\u636e\u9884\u8ba1\u7b97scale\uff09\u3002<\/p>\n<h2>\u56db\u3001KV Cache \u4f18\u5316\uff1a\u8ba9\u4e0a\u4e0b\u6587\u7a97\u53e3&quot;\u88c5\u5f97\u4e0b&quot;<\/h2>\n<p>KV Cache\u4f18\u5316\u662f\u63d0\u5347\u957f\u5e8f\u5217\u63a8\u7406\u6548\u7387\u7684\u5173\u952e\u3002\u968f\u7740\u4e0a\u4e0b\u6587\u7a97\u53e3\u4ece4K\u6269\u5c55\u5230128K\u751a\u81f31M\uff0cKV Cache\u7684\u663e\u5b58\u5360\u7528\u5448\u7ebf\u6027\u589e\u957f\u3002<\/p>\n<h3>4.1 KV Cache\u91cf\u5316<\/h3>\n<p>\u5bf9KV Cache\u505aINT8\u91cf\u5316\uff0c\u53ef\u4ee5\u5c06\u5176\u5360\u7528\u964d\u4f4e\u4e00\u534a\u3002KV Cache\u91cf\u5316\u6bd4\u6743\u91cd\u91cf\u5316\u66f4\u96be\u505a\uff0c\u56e0\u4e3a\u7cbe\u5ea6\u635f\u5931\u4f1a\u9010\u5c42\u7d2f\u79ef\u3002\u5f53\u524d\u8f83\u6210\u719f\u7684\u65b9\u6848\u6709\uff1a<\/p>\n<ul>\n<li><strong>KIVI\uff1a<\/strong> \u5bf9K\u548cV\u5206\u522b\u91cf\u5316\uff0cK\u7528\u91cf\u5316+\u5206\u7ec4\u65b9\u5f0f\u538b\u7f29\uff0cV\u4fdd\u6301\u8f83\u9ad8\u7cbe\u5ea6<\/li>\n<li><strong>KVQuant\uff1a<\/strong> \u652f\u6301INT2~INT8\u6df7\u5408\u7cbe\u5ea6\u91cf\u5316\uff0c\u6839\u636e\u6ce8\u610f\u529b\u5934\u7684\u654f\u611f\u5ea6\u81ea\u9002\u5e94\u5206\u914dbit\u6570<\/li>\n<\/ul>\n<h3>4.2 Multi-Query Attention (MQA) \u4e0e Grouped-Query Attention (GQA)<\/h3>\n<p>\u539f\u59cbTransformer\u8bba\u6587\u4f7f\u7528\u591a\u5934\u6ce8\u610f\u529b(MHA)\uff0cKV\u5934\u6570\u7b49\u4e8eQuery\u5934\u6570\u3002MQA\u548cGQA\u5927\u5e45\u51cf\u5c11\u4e86KV\u5934\u7684\u6570\u91cf\uff1a<\/p>\n<ul>\n<li>MQA\uff1a\u6240\u6709Query\u5934\u5171\u4eab1\u4e2aKV\u5934\uff0cKV Cache\u964d\u81f31\/n\uff08n=\u5934\u6570\uff09<\/li>\n<li>GQA\uff1a\u4e2d\u95f4\u65b9\u6848\u2014\u2014n\u4e2aQuery\u5934\u5171\u4eab1\u4e2aKV\u5934\uff0c\u5e73\u8861\u4e86\u6548\u7387\u4e0e\u8d28\u91cf<\/li>\n<\/ul>\n<p>\u5927\u591a\u6570\u73b0\u4ee3\u6a21\u578b\uff08Llama 2\/3\u3001Mistral\u3001Qwen 2\uff09\u5747\u5df2\u91c7\u7528GQA\u3002\u63a8\u7406\u6846\u67b6\u53ef\u4ee5\u901a\u8fc7\u8fd9\u4e00\u7279\u6027\u8fdb\u4e00\u6b65\u505a\u5e76\u884c\u4f18\u5316\u3002<\/p>\n<h3>4.3 PagedAttention \u4e0e vLLM<\/h3>\n<p>vLLM\u662f\u5f53\u524d\u6700\u6d41\u884c\u7684\u63a8\u7406\u670d\u52a1\u6846\u67b6\u4e4b\u4e00\uff0c\u5176\u6838\u5fc3\u521b\u65b0PagedAttention\u501f\u9274\u4e86\u64cd\u4f5c\u7cfb\u7edf\u865a\u62df\u5185\u5b58\u7684\u601d\u8def\uff1a\u4e0d\u518d\u4e3a\u6bcf\u4e2a\u8bf7\u6c42\u9884\u5206\u914d\u5b8c\u6574\u7684KV Cache\u8fde\u7eed\u7a7a\u95f4\uff0c\u800c\u662f\u4ee5&quot;\u9875&quot;(page)\u4e3a\u5355\u4f4d\u6309\u9700\u5206\u914d\u3002<\/p>\n<p><strong>\u6548\u679c\uff1a<\/strong> <\/p>\n<ul>\n<li>\u663e\u5b58\u5229\u7528\u7387\u4ece60%\u63d0\u5347\u523095%+<\/li>\n<li>\u652f\u6301\u66f4\u5927\u7684\u5e76\u53d1\u6570\u548c\u66f4\u957f\u7684\u5e8f\u5217<\/li>\n<li>\u914d\u5408\u8fde\u7eed\u6279\u5904\u7406(Continuous Batching)\uff0c\u541e\u5410\u91cf\u53ef\u8fbe\u4f20\u7edf\u65b9\u6848\u7684\u6570\u500d<\/li>\n<\/ul>\n<h2>\u4e94\u3001\u8ba1\u7b97\u4f18\u5316\uff1a\u4ece\u7b97\u6cd5\u5230\u7b97\u5b50\u5c42\u9762<\/h2>\n<h3>5.1 Flash Attention<\/h3>\n<p>Flash Attention\u662f\u8fd1\u5e74\u6765Transformer\u63a8\u7406\u52a0\u901f\u6700\u91cd\u8981\u7684\u7b97\u6cd5\u521b\u65b0\u3002\u5b83\u901a\u8fc7\u5206\u5757\u8ba1\u7b97(tiling)\u548c\u878d\u5408\u8f6f\u6700\u5927\u5316(fused softmax)\u6280\u672f\uff0c\u5c06\u6ce8\u610f\u529b\u8ba1\u7b97\u7684\u663e\u5b58\u5360\u7528\u4eceO(n\u00b2)\u964d\u81f3O(n)\u3002\u5728H100\u4e0a\u914d\u5408FP8\u4f7f\u7528\uff0cFlash Attention-3\u53ef\u4ee5\u5b9e\u73b0\u7406\u8bba\u5cf0\u503c\u768460%+\u5229\u7528\u7387\u3002<\/p>\n<h3>5.2 \u7b97\u5b50\u878d\u5408\u4e0e\u7f16\u8bd1<\/h3>\n<ul>\n<li><strong>\u7b97\u5b50\u878d\u5408(Fused Ops)\uff1a<\/strong> \u5c06\u591a\u4e2a\u8fde\u7eed\u5c0f\u7b97\u5b50\u5408\u5e76\u4e3a\u4e00\u4e2a\u5927\u7b97\u5b50\uff0c\u51cf\u5c11Kernel Launch\u5f00\u9500\u548c\u663e\u5b58\u5e26\u5bbd\u5f80\u8fd4\u3002\u4f8b\u5982LayerNorm + Residual Add\u4e00\u6b21\u6027\u5b8c\u6210<\/li>\n<li><strong>\u56fe\u4f18\u5316\uff1a<\/strong> \u901a\u8fc7TorchScript\u6216\u4e13\u7528\u7f16\u8bd1\u5668\u4f18\u5316\u8ba1\u7b97\u56fe\u2014\u2014\u53bb\u9664\u5197\u4f59\u8fd0\u7b97\u3001\u6298\u53e0\u5e38\u91cf\u8868\u8fbe\u5f0f\u3001\u91cd\u6392\u6267\u884c\u987a\u5e8f<\/li>\n<li><strong>CUDA Graph\uff1a<\/strong> \u4e00\u6b21\u7f16\u8bd1\u3001\u591a\u6b21\u6267\u884c\uff0c\u6d88\u9664\u6bcf\u6b21\u63a8\u7406\u542f\u52a8\u65f6\u7684GPU Kernel\u8c03\u5ea6\u5f00\u9500<\/li>\n<\/ul>\n<h3>5.3 \u6a21\u578b\u526a\u88c1\u4e0e\u84b8\u998f<\/h3>\n<ul>\n<li><strong>\u526a\u88c1(Pruning)\uff1a<\/strong> \u79fb\u9664\u4e0d\u91cd\u8981\u7684\u6743\u91cd\u8fde\u63a5\u6216\u6ce8\u610f\u529b\u5934\uff0c\u5927\u5e45\u51cf\u5c11\u53c2\u6570\u6570\u91cf\u3002\u7ed3\u6784\u5316\u526a\u88c1\u53ef\u76f4\u63a5\u51cf\u5c0f\u6a21\u578b\u5c3a\u5bf8\uff0c\u4f46\u8bad\u7ec3\u6210\u672c\u9ad8<\/li>\n<li><strong>\u84b8\u998f(Distillation)\uff1a<\/strong> \u8bad\u7ec3\u4e00\u4e2a\u5c0f\u6a21\u578b($\u5c0f)\u53bb\u6a21\u4eff\u5927\u6a21\u578b($\u5927)\u7684\u8f93\u51fa\u5206\u5e03\u3002\u5178\u578b\u5982Phi-3\u7cfb\u5217\uff0c\u75283.8B\u53c2\u6570\u8fbe\u52307B\u7ea7\u522b\u6027\u80fd<\/li>\n<\/ul>\n<h2>\u516d\u3001\u670d\u52a1\u67b6\u6784\u4f18\u5316\uff1a\u4ece\u5355\u673a\u5230\u96c6\u7fa4<\/h2>\n<p>\u5355\u4e2a\u6a21\u578b\u7684\u63a8\u7406\u4f18\u5316\u8fdc\u8fdc\u4e0d\u591f\uff0c\u7cfb\u7edf\u7ea7\u7684\u67b6\u6784\u8bbe\u8ba1\u51b3\u5b9a\u4e86\u5b9e\u9645\u670d\u52a1\u7684\u541e\u5410\u548c\u5ef6\u8fdf\u3002<\/p>\n<h3>6.1 Continuous Batching\uff08\u8fde\u7eed\u6279\u5904\u7406\uff09<\/h3>\n<p>\u4f20\u7edf\u6279\u5904\u7406\u5728\u8bf7\u6c42\u4e0d\u9971\u548c\u65f6\u6d6a\u8d39\u7b97\u529b\uff0c\u8fde\u7eed\u6279\u5904\u7406\u5219\u4f1a\u5728\u6bcf\u4e2a\u63a8\u7406\u6b65\u9aa4\u540e\u68c0\u67e5\u662f\u5426\u6709\u65b0\u8bf7\u6c42\u5230\u8fbe\uff0c\u5141\u8bb8&quot;\u63d2\u961f&quot;\u5904\u7406\u3002vLLM\u548cTensorRT-LLM\u90fd\u652f\u6301\u8fd9\u4e00\u673a\u5236\uff0c\u5728\u5927\u5e76\u53d1\u573a\u666f\u4e0b\u541e\u5410\u63d0\u5347\u53ef\u8fbe10\u500d\u3002<\/p>\n<h3>6.2 \u524d\u7f00\u7f13\u5b58\uff08Prefix Caching\uff09<\/h3>\n<p>\u5f53\u5927\u91cf\u8bf7\u6c42\u5171\u4eab\u76f8\u540c\u7684\u524d\u7f00\uff08\u5982AI Agent\u7684\u7cfb\u7edf\u63d0\u793a\u8bcd\u76f8\u540c\u7684\u5f88\u957f\uff09\u65f6\uff0c\u53ef\u4ee5\u7f13\u5b58\u8fd9\u4e00\u90e8\u5206\u7684KV Cache\uff0c\u540e\u7eed\u8bf7\u6c42\u76f4\u63a5\u590d\u7528\u3002\u5728Chatbot\u3001Agent\u5e94\u7528\u4e2d\u6548\u679c\u663e\u8457\u3002<\/p>\n<h3>6.3 \u63a8\u6d4b\u89e3\u7801\uff08Speculative Decoding\uff09<\/h3>\n<p>\u6838\u5fc3\u601d\u8def\uff1a\u7528\u4e00\u4e2a\u8f7b\u91cf\u7ea7\u8349\u7a3f\u6a21\u578b\uff08\u901a\u5e38\u5c0f\u51e0\u5341\u500d\uff09\u5feb\u901f\u751f\u6210\u5019\u9009\u5e8f\u5217\uff0c\u518d\u7528\u5927\u6a21\u578b\u4e00\u6b21\u6027\u9a8c\u8bc1\u3002\u5982\u679c\u8349\u7a3f\u6a21\u578b\u51c6\u786e\u7387\u9ad8\uff0c\u63a8\u7406\u901f\u5ea6\u53ef\u63d0\u53472~3\u500d\u800c\u4e0d\u635f\u5931\u8d28\u91cf\u3002<\/p>\n<h3>6.4 \u90e8\u7f72\u6846\u67b6\u9009\u578b\u5bf9\u6bd4<\/h3>\n<table>\n<thead>\n<tr>\n<th>\u6846\u67b6<\/th>\n<th>\u4eae\u70b9<\/th>\n<th>\u9002\u5408\u573a\u666f<\/th>\n<th>\u5c40\u9650\u6027<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>vLLM<\/td>\n<td>PagedAttention\uff0c\u4e0a\u624b\u7b80\u5355<\/td>\n<td>\u901a\u7528\u63a8\u7406\u670d\u52a1<\/td>\n<td>\u591a\u673a\u6269\u5c55\u4e0d\u5982TGI<\/td>\n<\/tr>\n<tr>\n<td>TensorRT-LLM<\/td>\n<td>NVIDIA\u751f\u6001\u6df1\u5ea6\u4f18\u5316<\/td>\n<td>\u751f\u4ea7\u7ea7\u9ad8\u6027\u80fd\u670d\u52a1<\/td>\n<td>\u5b66\u4e60\u6210\u672c\u9ad8\uff0c\u6a21\u578b\u652f\u6301\u6709\u9650<\/td>\n<\/tr>\n<tr>\n<td>llama.cpp<\/td>\n<td>CPU\u63a8\u7406\/\u8fb9\u7f18\u8bbe\u5907<\/td>\n<td>\u672c\u5730\/\u8d44\u6e90\u53d7\u9650\u573a\u666f<\/td>\n<td>GPU\u5229\u7528\u7387\u8f83\u4f4e<\/td>\n<\/tr>\n<tr>\n<td>SGLang<\/td>\n<td>\u524d\u7f00\u7f13\u5b58+\u7ed3\u6784\u5316LLM\u7f16\u7a0b<\/td>\n<td>\u590d\u6742\u63a8\u7406\u4e0eAgent\u573a\u666f<\/td>\n<td>\u751f\u6001\u5c1a\u5728\u5feb\u901f\u53d1\u5c55<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h2>\u4e03\u3001\u5b9e\u9645\u90e8\u7f72\u7684\u907f\u5751\u6307\u5357<\/h2>\n<h3>7.1 \u663e\u5b58\u4f30\u7b97\u516c\u5f0f<\/h3>\n<p>\u4ee5INT4\u91cf\u5316\u540e\u768470B\u6a21\u578b\u4e3a\u4f8b\uff0c\u9700\u8981\u4f30\u7b97\u4e09\u90e8\u5206\u663e\u5b58\uff1a<\/p>\n<ul>\n<li>\u6a21\u578b\u6743\u91cd\uff1a70B \u00d7 0.5\u5b57\u8282(4-bit) \u2248 35GB<\/li>\n<li>KV Cache\uff1abatch_size \u00d7 max_seq_len \u00d7 layers \u00d7 hidden_dim \u00d7 KV_heads_ratio \u00d7 2\u5b57\u8282<\/li>\n<li>\u4e2d\u95f4\u6fc0\u6d3b\uff1a\u7ea6\u7b49\u4e8e\u6a21\u578b\u6743\u91cd\u4f7f\u7528FP16\u65f6\u76843~8\u500d\uff08\u53d6\u51b3\u4e8ebatch size\u548c\u5e8f\u5217\u957f\u5ea6\uff09<\/li>\n<\/ul>\n<p><strong>\u4e00\u4e2a\u5e38\u89c1\u9677\u9631\uff1a<\/strong> \u5f88\u591a\u4eba\u5728\u670d\u52a1\u7aef\u53ea\u9884\u7559\u4e86\u6a21\u578b\u6743\u91cd\u7684\u663e\u5b58\uff0c\u5ffd\u7565KV Cache\u548c\u4e2d\u95f4\u6fc0\u6d3b\uff0c\u5bfc\u81f4OOM\u3002<\/p>\n<h3>7.2 \u91cf\u5316\u7cbe\u5ea6\u9a8c\u8bc1\u6d41\u7a0b<\/h3>\n<p>\u91cf\u5316\u540e\u4e0d\u662f&quot;\u8dd1\u901a\u4e86\u5c31\u884c&quot;\uff0c\u4e00\u5b9a\u8981\u505a\u7cfb\u7edf\u6027\u7cbe\u5ea6\u9a8c\u8bc1\uff1a<\/p>\n<ol>\n<li>\u5148\u5728\u6807\u51c6\u57fa\u51c6\u96c6\uff08MMLU\u3001HumanEval\u7b49\uff09\u4e0a\u6d4b\u8bd5\u91cf\u5316\u524d\u540e\u7684\u5206\u6570\u53d8\u5316<\/li>\n<li>\u518d\u7528\u5b9e\u9645\u4e1a\u52a1\u6570\u636e\u9a8c\u8bc1\u2014\u2014\u56e0\u4e3a\u91cf\u5316\u5bf9\u957f\u5c3e\u5206\u5e03\u7684\u9886\u57df\u77e5\u8bc6\u5f71\u54cd\u53ef\u80fd\u88ab\u57fa\u51c6\u96c6\u63a9\u76d6<\/li>\n<li>\u505a\u7aef\u5230\u7aef\u7684A\/B\u6d4b\u8bd5\u2014\u2014\u5c06\u91cf\u5316\u524d\u540e\u7684\u6a21\u578b\u653e\u5728\u540c\u4e00\u4e1a\u52a1\u7ba1\u9053\u4e2d\u76f2\u6d4b<\/li>\n<\/ol>\n<h3>7.3 \u670d\u52a1\u8d28\u91cf(QoS)\u4fdd\u969c<\/h3>\n<p>\u5728\u5b9e\u9645\u8fd0\u8425\u4e2d\uff0c\u6700\u5927\u5ba2\u6237\u7aef\u541e\u5410\u91cf(TPS)\u548c\u5ef6\u8fdf(P99)\u5f80\u5f80\u662f\u77db\u76fe\u7684\u3002\u5efa\u8bae\uff1a<\/p>\n<ul>\n<li>\u7528\u8bf7\u6c42\u6392\u961f+\u52a8\u6001\u6279\u5904\u7406\u5e73\u8861\u541e\u5410\u548c\u5ef6\u8fdf<\/li>\n<li>\u8bbe\u7f6e\u6700\u5927\u5e76\u53d1\u6570\u9650\u5236\uff0c\u907f\u514d\u6240\u6709GPU\u7ebf\u7a0b\u88ab\u8bf7\u6c42\u8017\u5c3d\u5bfc\u81f4\u5ef6\u8fdf\u98d9\u5347<\/li>\n<li>\u5b9e\u65bd\u964d\u7ea7\u7b56\u7565\u2014\u2014\u9ad8\u8d1f\u8f7d\u65f6\u5207\u6362\u4e3a\u5c0f\u6a21\u578b\u6216\u964d\u6863\u91cf\u5316<\/li>\n<\/ul>\n<h2>\u516b\u3001\u672a\u6765\u8d8b\u52bf\u4e0e\u603b\u7ed3<\/h2>\n<p>\u5927\u6a21\u578b\u63a8\u7406\u4f18\u5316\u6b63\u671d\u7740\u4e09\u4e2a\u65b9\u5411\u52a0\u901f\u6f14\u8fdb\uff1a<\/p>\n<ol>\n<li><strong>\u786c\u4ef6\u4e13\u7528\u5316\uff1a<\/strong> NVIDIA B200\u5f15\u5165FP4\u7cbe\u5ea6\u652f\u6301\uff0cGroq\u7684LPU\u4e3a\u63a8\u7406\u573a\u666f\u505a\u6781\u81f4\u4f18\u5316<\/li>\n<li><strong>\u7b97\u6cd5\u4e0e\u7cfb\u7edf\u534f\u540c\uff1a<\/strong> \u7a00\u758f\u6ce8\u610f\u529b\u3001\u72b6\u6001\u7a7a\u95f4\u6a21\u578b(\u5982Mamba)\u6b63\u4ece\u6839\u6e90\u6539\u53d8Transformer\u7684\u63a8\u7406\u8ba1\u7b97\u6a21\u5f0f<\/li>\n<li><strong>\u81ea\u52a8\u5316\u8c03\u4f18\uff1a<\/strong> AutoQuant\u3001\u81ea\u52a8\u641c\u7d22\u6700\u4f18\u91cf\u5316\u7b56\u7565\u548c\u8c03\u5ea6\u53c2\u6570\u6b63\u5728\u6210\u4e3a\u6807\u914d\u5de5\u5177<\/li>\n<\/ol>\n<p>\u603b\u7ed3\u6765\u8bf4\uff0c\u5927\u6a21\u578b\u63a8\u7406\u4f18\u5316\u5df2\u7ecf\u4ece&quot;\u80fd\u4e0d\u80fd\u8dd1&quot;\u7684\u9636\u6bb5\u8fdb\u5165&quot;\u600e\u4e48\u8dd1\u5f97\u66f4\u5feb\u66f4\u7701&quot;\u7684\u65b0\u9636\u6bb5\u3002\u5bf9\u4e8e\u5de5\u7a0b\u5e08\u800c\u8a00\uff0c\u7406\u89e3\u91cf\u5316\u3001KV Cache\u4f18\u5316\u3001\u8ba1\u7b97\u56fe\u4f18\u5316\u548c\u670d\u52a1\u8c03\u5ea6\u8fd9\u56db\u4e2a\u7ef4\u5ea6\uff0c\u7ed3\u5408\u5b9e\u9645\u4e1a\u52a1\u573a\u666f\u505a\u7cfb\u7edf\u7ea7\u7684\u6743\u8861\u51b3\u7b56\uff0c\u624d\u662f\u505a\u597d\u5927\u6a21\u578b\u90e8\u7f72\u7684\u5173\u952e\u3002\u6ca1\u6709\u94f6\u5f39\uff0c\u4f46\u6709\u4e86\u5de5\u5177\u7bb1\u2014\u2014\u9009\u5bf9\u5de5\u5177\u3001\u7528\u597d\u5de5\u5177\uff0c\u5c31\u662f\u6700\u4f18\u89e3\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u4e00\u3001\u5f15\u8a00 \u968f\u7740\u5927\u8bed\u8a00\u6a21\u578b\u53c2\u6570\u89c4\u6a21\u4ece\u6570\u5341\u4ebf\u8dc3\u5347\u81f3\u4e07\u4ebf\u7ea7\u522b\uff0c\u6a21\u578b&quot;\u80fd\u505a\u4ec0\u4e48&quot;\u5df2\u7ecf\u4e0d\u662f\u74f6\u9888\uff0c&#038;q [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[5],"tags":[18,26,25,27],"class_list":["post-113","post","type-post","status-publish","format-standard","hentry","category-ai","tag-ai","tag-26","tag-25","tag-27"],"_links":{"self":[{"href":"https:\/\/www.joezhou.top\/index.php?rest_route=\/wp\/v2\/posts\/113","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.joezhou.top\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.joezhou.top\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.joezhou.top\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.joezhou.top\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=113"}],"version-history":[{"count":0,"href":"https:\/\/www.joezhou.top\/index.php?rest_route=\/wp\/v2\/posts\/113\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.joezhou.top\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=113"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.joezhou.top\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=113"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.joezhou.top\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=113"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}