{"id":1698,"date":"2026-06-09T11:37:54","date_gmt":"2026-06-09T03:37:54","guid":{"rendered":"https:\/\/www.izhuhn.cn\/?p=1698"},"modified":"2026-06-10T00:50:17","modified_gmt":"2026-06-09T16:50:17","slug":"atlas-npu%e5%a4%9a%e5%ae%9e%e4%be%8bpaddleocr%e9%83%a8%e7%bd%b2%e4%b8%8e%e8%b4%9f%e8%bd%bd%e5%9d%87%e8%a1%a1%e5%ae%9e%e6%88%98%e7%ae%80%e8%ae%b0","status":"publish","type":"post","link":"https:\/\/www.izhuhn.cn\/index.php\/2026\/06\/09\/atlas-npu%e5%a4%9a%e5%ae%9e%e4%be%8bpaddleocr%e9%83%a8%e7%bd%b2%e4%b8%8e%e8%b4%9f%e8%bd%bd%e5%9d%87%e8%a1%a1%e5%ae%9e%e6%88%98%e7%ae%80%e8%ae%b0\/","title":{"rendered":"NPU\u591a\u5b9e\u4f8bPaddleOCR\u90e8\u7f72\u4e0e\u8d1f\u8f7d\u5747\u8861\u5b9e\u6218\u7b80\u8bb0"},"content":{"rendered":"\n<h2 class=\"wp-block-heading\">\u4e00\u3001\u80cc\u666f\u4e0e\u52a8\u673a<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u5927\u89c4\u6a21\u4f01\u4e1a\u6587\u6863\u5904\u7406\u7684OCR\u573a\u666f\u4e2d\uff0c\u9700\u8981\u5bf9\u6d77\u91cf\u626b\u63cf\u4ef6\/\u6587\u6863\u56fe\u7247\u8fdb\u884c\u9ad8\u5e76\u53d1\u8bc6\u522b\u3002\u5355\u5b9e\u4f8bPaddleOCR\u9762\u5bf9\u5927\u91cf\u8bf7\u6c42\u65f6\u541e\u5410\u91cf\u6709\u9650\uff0c\u56e0\u6b64\u9700\u8981\u5728 **\u534e\u4e3aAtlas 800 TA2**\uff088\u5361Ascend NPU\uff09\u670d\u52a1\u5668\u4e0a\u90e8\u7f72\u591a\u5b9e\u4f8bPaddleOCR\uff0c\u5e76\u901a\u8fc7\u81ea\u7814\u7684 **uvicorn\u8d1f\u8f7d\u5747\u8861\u7f51\u5173** \u7edf\u4e00\u5bf9\u5916\u66b4\u9732API\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6574\u4e2a\u65b9\u6848\u7528\u5230\u4e86\u4e09\u5c42\u67b6\u6784\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>**vllm-ascend**\uff08\u63a8\u7406\u5f15\u64ce\uff09\uff1a\u6bcf\u5f20NPU\u5361\u8fd0\u884c2\u4e2avllm\u5b9e\u4f8b\uff0c\u8d1f\u8d23VL\u6a21\u578b\u7684\u63a8\u7406\u52a0\u901f<\/li>\n\n\n\n<li>**paddlex**\uff08OCR pipeline\uff09\uff1a\u6bcf\u5361\u8fd0\u884c4-6\u4e2apaddlex\u5b9e\u4f8b\uff0c\u8c03\u7528vllm\u5b8c\u6210OCR\u5168\u6d41\u7a0b<\/li>\n\n\n\n<li>**uvicorn\u7f51\u5173**\uff08\u8d1f\u8f7d\u5747\u8861\uff09\uff1aFastAPI\u5b9e\u73b0\u7684\u667a\u80fd\u8d1f\u8f7d\u5747\u8861\u5c42\uff0c\u63d0\u4f9b\u7edf\u4e00\u7684HTTP API<\/li>\n<\/ul>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e8c\u3001\u786c\u4ef6\u73af\u5883<\/h2>\n\n\n\n<figure class=\"wp-block-table\"><table>\n<thead><tr><th>\u9879\u76ee<\/th><th>\u914d\u7f6e<\/th><\/tr><\/thead>\n<tbody>\n<tr><td>\u670d\u52a1\u5668<\/td><td>Atlas 800 TA2<\/td><\/tr>\n<tr><td>NPU<\/td><td>\u534e\u4e3aAscend 910B \u00d7 8\uff08\u53610~\u53617\uff09<\/td><\/tr>\n<tr><td>\u64cd\u4f5c\u7cfb\u7edf<\/td><td>EulerOS \/ openEuler<\/td><\/tr>\n<tr><td>\u5bb9\u5668\u8fd0\u884c\u65f6<\/td><td>Docker\uff08\u975eK8s\uff09<\/td><\/tr>\n<tr><td>\u955c\u50cf<\/td><td>`vllm-ascend:v0.19.1rc1` \/ `paddleocrvl:20260529`<\/td><\/tr>\n<tr><td>\u6a21\u578b<\/td><td>PaddleOCR-VL-1.5-0.9B<\/td><\/tr>\n<\/tbody>\n<\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">&gt; \u6ce8\u610f\uff1a\u670d\u52a1\u5668\u5df2\u88ab\u5176\u4ed6\u5e73\u53f0\u7eb3\u7ba1\uff0c\u4e0d\u80fd\u4f7f\u7528Kubernetes\uff0c\u56e0\u6b64\u91c7\u7528\u7eafDocker + Compose\u7f16\u6392\u65b9\u6848\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e09\u3001\u6574\u4f53\u67b6\u6784<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u6574\u4e2a\u7cfb\u7edf\u5206\u4e3a\u56db\u5c42\uff1aUvicorn \u7f51\u5173 \u2192 paddlex OCR Pipeline \u2192 vllm \u63a8\u7406\u5f15\u64ce \u2192 NPU \u786c\u4ef6\u3002\u4e0b\u9762\u5206\u522b\u7528\u9ad8\u5c42\u603b\u89c8\u548c\u5355\u5361\u8be6\u7ec6\u62d3\u6251\u6765\u5c55\u793a\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">8\u5361\u5b8c\u6574\u62d3\u6251\u603b\u89c8<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e0b\u56fe\u5c55\u793a\u5168\u90e88\u5f20NPU\u5361\u300116\u4e2avllm\u5b9e\u4f8b\u300148\u4e2apaddlex\u5b9e\u4f8b\u7684\u5b8c\u6574\u62d3\u6251\u5173\u7cfb\u3002\u6bcf\u5217\u7684\"\u2713\"\u8868\u793a\u5df2\u6709\u5b9e\u4f8b\uff0c\"\u65b0\u589e\"\u8868\u793a\u672c\u6b21\u6269\u5bb9\u589e\u52a0\u7684\u5b9e\u4f8b\u3002<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>\n                         \u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510\n                         \u2502                     Uvicorn \u8d1f\u8f7d\u5747\u8861\u7f51\u5173 :9101                           \u2502\n                         \u2502               round_robin \u00d7 48 upstreams + failover                     \u2502\n                         \u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518\n                                                             \u2502\n    \u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510\n    \u2502   \u53610      \u2502   \u53611      \u2502   \u53612      \u2502   \u53613      \u2502   \u53614      \u2502   \u53615      \u2502   \u53616      \u2502   \u53617      \u2502\n    \u2502 (\u5168\u90e8\u65b0\u589e) \u2502 (\u5168\u90e8\u65b0\u589e) \u2502 (\u5df2\u6709+\u65b0\u589e)\u2502 (\u5df2\u6709+\u65b0\u589e)\u2502 (\u5df2\u6709+\u65b0\u589e)\u2502 (\u5df2\u6709+\u65b0\u589e)\u2502 (\u5df2\u6709+\u65b0\u589e)\u2502 (\u5df2\u6709+\u65b0\u589e)\u2502\n    \u2502 davinci0   \u2502 davinci1   \u2502 davinci2   \u2502 davinci3   \u2502 davinci4   \u2502 davinci5   \u2502 davinci6   \u2502 davinci7   \u2502\n    \u251c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2524\n    \u2502            \u2502            \u2502            \u2502            \u2502            \u2502            \u2502            \u2502            \u2502\n    \u2502  vllm:9031 \u2502  vllm:9032 \u2502  vllm:9025\u2713\u2502  vllm:9026\u2713\u2502  vllm:9027\u2713\u2502  vllm:9028\u2713\u2502  vllm:9029\u2713\u2502  vllm:9030\u2713\u2502\n    \u2502   \u251c\u25009126   \u2502   \u251c\u25009127   \u2502   \u251c\u25009120   \u2502   \u251c\u25009121   \u2502   \u251c\u25009122   \u2502   \u251c\u25009123   \u2502   \u251c\u25009124   \u2502   \u251c\u25009125   \u2502\n    \u2502   \u251c\u25009136   \u2502   \u251c\u25009137   \u2502   \u251c\u25009130   \u2502   \u251c\u25009131   \u2502   \u251c\u25009132   \u2502   \u251c\u25009133   \u2502   \u251c\u25009134   \u2502   \u251c\u25009135   \u2502\n    \u2502   \u2514\u25009146   \u2502   \u2514\u25009147   \u2502   \u2514\u25009140   \u2502   \u2514\u25009141   \u2502   \u2514\u25009142   \u2502   \u2514\u25009143   \u2502   \u2514\u25009144   \u2502   \u2514\u25009145   \u2502\n    \u2502            \u2502            \u2502            \u2502            \u2502            \u2502            \u2502            \u2502            \u2502\n    \u2502  vllm:9041 \u2502  vllm:9042 \u2502  vllm:9035 \u2502  vllm:9036 \u2502  vllm:9037 \u2502  vllm:9038 \u2502  vllm:9039 \u2502  vllm:9040 \u2502\n    \u2502   \u251c\u25009156   \u2502   \u251c\u25009157   \u2502   \u251c\u25009150   \u2502   \u251c\u25009151   \u2502   \u251c\u25009152   \u2502   \u251c\u25009153   \u2502   \u251c\u25009154   \u2502   \u251c\u25009155   \u2502\n    \u2502   \u251c\u25009166   \u2502   \u251c\u25009167   \u2502   \u251c\u25009160   \u2502   \u251c\u25009161   \u2502   \u251c\u25009162   \u2502   \u251c\u25009163   \u2502   \u251c\u25009164   \u2502   \u251c\u25009165   \u2502\n    \u2502   \u2514\u25009176   \u2502   \u2514\u25009177   \u2502   \u2514\u25009170   \u2502   \u2514\u25009171   \u2502   \u2514\u25009172   \u2502   \u2514\u25009173   \u2502   \u2514\u25009174   \u2502   \u2514\u25009175   \u2502\n    \u251c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2524\n    \u2502  paddle\u00d76  \u2502  paddle\u00d76  \u2502  paddle\u00d76  \u2502  paddle\u00d76  \u2502  paddle\u00d76  \u2502  paddle\u00d76  \u2502  paddle\u00d76  \u2502  paddle\u00d76  \u2502\n    \u2502  vllm\u00d72    \u2502  vllm\u00d72    \u2502  vllm\u00d72    \u2502  vllm\u00d72    \u2502  vllm\u00d72    \u2502  vllm\u00d72    \u2502  vllm\u00d72    \u2502  vllm\u00d72    \u2502\n    \u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2534\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2534\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2534\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2534\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2534\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2534\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2534\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518\n                                                             \u2502\n                                                    \u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2534\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510\n                                                    \u2502  Ascend 910B\u00d78  \u2502\n                                                    \u2502  Atlas 800 TA2  \u2502\n                                                    \u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518\n  \u2713 = \u5df2\u6709\u5b9e\u4f8b  \u2502  \u65e0\u6807\u8bb0 = \u65b0\u589e\u5b9e\u4f8b  \u2502  \u251c\u2500 = paddlex \u2192 vllm \u8def\u7531\u7ed1\u5b9a\n<\/code><\/pre>\n\n\n\n<h3 class=\"wp-block-heading\">\u5355\u5361\u5185\u90e8\u8be6\u7ec6\u62d3\u6251\uff08\u4ee5\u53612\u4e3a\u4f8b\uff0c8\u5361\u683c\u5c40\u5b8c\u5168\u76f8\u540c\uff09<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u6bcf\u5f20NPU\u5361\u4e0a\u7684\u62d3\u6251<strong>\u5b8c\u5168\u4e00\u81f4<\/strong>\uff1a2\u4e2avllm\u5b9e\u4f8b\u5e73\u5206\u663e\u5b58\uff0c6\u4e2apaddlex\u5b9e\u4f8b\u5206\u4e24\u7ec4\u5206\u522b\u6307\u5411\u4e24\u4e2avllm\u5b9e\u4f8b\u3002\u4e0b\u9762\u662f\u5355\u5361\u5185\u90e8\u5b8c\u6574\u94fe\u8def\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>\n                      \u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510\n                      \u2502          Uvicorn \u7f51\u5173 :9101                  \u2502\n                      \u2502   48\u4e2a upstreams \u4e2d\u5305\u62ec\u672c\u5361\u76846\u4e2apaddlex      \u2502\n                      \u2514\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2518\n                            \u2502      \u2502      \u2502      \u2502      \u2502      \u2502\n                   \u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u25bc\u2500\u2500\u2510 \u250c\u2500\u25bc\u2500\u2500\u2500\u2500\u2500\u2500\u2510\u250c\u2500\u25bc\u2500\u2500\u2500\u2500\u2500\u2500\u2510\u250c\u2500\u25bc\u2500\u2500\u2500\u2500\u2500\u2500\u2510\u250c\u2500\u25bc\u2500\u2500\u2510  \u250c\u2500\u25bc\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510\n                   \u2502paddlex    \u2502 \u2502paddlex \u2502\u2502paddlex \u2502\u2502paddlex \u2502\u2502pad \u2502  \u2502paddlex     \u2502\n                   \u2502:9120      \u2502 \u2502:9130   \u2502\u2502:9140   \u2502\u2502:9150   \u2502\u2502:.. \u2502  \u2502:9170       \u2502\n                   \u2502paddle-    \u2502 \u2502paddle- \u2502\u2502paddle- \u2502\u2502paddle- \u2502\u2502    \u2502  \u2502paddle-     \u2502\n                   \u25029120-npu-2 \u2502 \u25029130-.. \u2502\u25029140-.. \u2502\u25029150-.. \u2502\u2502    \u2502  \u25029170-npu-2  \u2502\n                   \u2514\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2518 \u2514\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2518\u2514\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2518\u2514\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2518\u2514\u2500\u2500\u252c\u2500\u2518  \u2514\u2500\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2518\n                         \u2502           \u2502         \u2502         \u2502       \u2502            \u2502\n                         \u2502  \u524d3\u4e2a\u5b9e\u4f8b \u2502         \u2502  \u540e3\u4e2a\u5b9e\u4f8b\u2502       \u2502            \u2502\n                         \u2502  \u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2534\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2534\u2500\u2500\u2510  \u250c\u2500\u2500\u2500\u2534\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2534\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2534\u2500\u2500\u2510\n                         \u2502  \u2502                     \u2502  \u2502                            \u2502\n                         \u2514\u2500\u2500\u25ba  pipeline_config    \u2502  \u2502   pipeline_config          \u2502\n                            \u2502  server_url \u2192 :9025 \u2502  \u2502   server_url \u2192 :9035       \u2502\n                            \u2502                     \u2502  \u2502                            \u2502\n                            \u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518  \u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518\n                                       \u2502                           \u2502\n                                       \u25bc                           \u25bc\n                            \u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510     \u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510\n                            \u2502 vllm-ascend      \u2502     \u2502 vllm-ascend      \u2502\n                            \u2502 :9025 (\u5b9e\u4f8b1)     \u2502     \u2502 :9035 (\u5b9e\u4f8b2)     \u2502\n                            \u2502 gpu-mem-util 0.45 \u2502     \u2502 gpu-mem-util 0.45 \u2502\n                            \u2502 \u5bb9\u5668: vllm-ascend-\u2502     \u2502 \u5bb9\u5668: vllm-ascend-\u2502\n                            \u2502 paddleocr-9025-   \u2502     \u2502 paddleocr-9035-   \u2502\n                            \u2502       npu-2       \u2502     \u2502       npu-2       \u2502\n                            \u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518     \u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518\n                                     \u2502                        \u2502\n                                     \u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u252c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518\n                                              \u2502\n                                    \u250c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u25bc\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2510\n                                    \u2502   NPU \u53612          \u2502\n                                    \u2502   \/dev\/davinci2    \u2502\n                                    \u2502   Ascend 910B      \u2502\n                                    \u2514\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2518\n<\/code><\/pre>\n\n\n\n<h3 class=\"wp-block-heading\">8\u5361\u5b8c\u6574\u7aef\u53e3\u77e9\u9635\u4e0e\u8def\u7531\u5bf9\u7167\u8868<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u6bcf\u4e00\u884c\u662f\u4e00\u6761paddlex\u2192vllm\u8def\u7531\u7ed1\u5b9a\u5173\u7cfb\uff0c\u7aef\u53e3\u53f7\u672c\u8eab\u5c31\u7f16\u7801\u4e86\u5f52\u5c5e\u4fe1\u606f\uff0881xx=vllm, 82~87xx=paddlex\uff09\u3002<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>\n  \u5361\u53f7 \u2502 paddlex 6\u5b9e\u4f8b\u7aef\u53e3       \u2502 vllm 2\u5b9e\u4f8b\u7aef\u53e3          \u2502 vllm\u5b9e\u4f8b1 \u5bb9\u5668             \u2502 vllm\u5b9e\u4f8b2 \u5bb9\u5668\n  \u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n   \u53610 \u2502 9126  9136  9146       \u2502 9031 (\u5b9e\u4f8b1)           \u2502 ...-9031-npu-0            \u2502 ...-9041-npu-0\n       \u2502 9156  9166  9176       \u2502 9041 (\u5b9e\u4f8b2)           \u2502                           \u2502\n  \u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n   \u53611 \u2502 9127  9137  9147       \u2502 9032 (\u5b9e\u4f8b1)           \u2502 ...-9032-npu-1            \u2502 ...-9042-npu-1\n       \u2502 9157  9167  9177       \u2502 9042 (\u5b9e\u4f8b2)           \u2502                           \u2502\n  \u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n   \u53612 \u2502 9120  9130  9140       \u2502 9025 (\u5b9e\u4f8b1) \u2705\u5df2\u6709     \u2502 ...-9025-npu-2 \u2705\u5df2\u6709      \u2502 ...-9035-npu-2 \u65b0\u589e\n       \u2502 9150  9160  9170       \u2502 9035 (\u5b9e\u4f8b2) \u65b0\u589e       \u2502                           \u2502\n  \u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n   \u53613 \u2502 9121  9131  9141       \u2502 9026 (\u5b9e\u4f8b1) \u2705\u5df2\u6709     \u2502 ...-9026-npu-3 \u2705\u5df2\u6709      \u2502 ...-9036-npu-3 \u65b0\u589e\n       \u2502 9151  9161  9171       \u2502 9036 (\u5b9e\u4f8b2) \u65b0\u589e       \u2502                           \u2502\n  \u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n   \u53614 \u2502 9122  9132  9142       \u2502 9027 (\u5b9e\u4f8b1) \u2705\u5df2\u6709     \u2502 ...-9027-npu-4 \u2705\u5df2\u6709      \u2502 ...-9037-npu-4 \u65b0\u589e\n       \u2502 9152  9162  9172       \u2502 9037 (\u5b9e\u4f8b2) \u65b0\u589e       \u2502                           \u2502\n  \u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n   \u53615 \u2502 9123  9133  9143       \u2502 9028 (\u5b9e\u4f8b1) \u2705\u5df2\u6709     \u2502 ...-9028-npu-5 \u2705\u5df2\u6709      \u2502 ...-9038-npu-5 \u65b0\u589e\n       \u2502 9153  9163  9173       \u2502 9038 (\u5b9e\u4f8b2) \u65b0\u589e       \u2502                           \u2502\n  \u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n   \u53616 \u2502 9124  9134  9144       \u2502 9029 (\u5b9e\u4f8b1) \u2705\u5df2\u6709     \u2502 ...-9029-npu-6 \u2705\u5df2\u6709      \u2502 ...-9039-npu-6 \u65b0\u589e\n       \u2502 9154  9164  9174       \u2502 9039 (\u5b9e\u4f8b2) \u65b0\u589e       \u2502                           \u2502\n  \u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n   \u53617 \u2502 9125  9135  9145       \u2502 9030 (\u5b9e\u4f8b1) \u2705\u5df2\u6709     \u2502 ...-9030-npu-7 \u2705\u5df2\u6709      \u2502 ...-9040-npu-7 \u65b0\u589e\n       \u2502 9155  9165  9175       \u2502 9040 (\u5b9e\u4f8b2) \u65b0\u589e       \u2502                           \u2502\n  \u2500\u2500\u2500\u2500\u2500\u2534\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2534\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2534\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2534\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n  \u603b\u8ba1: 16 vllm-ascend \u5b9e\u4f8b + 48 paddlex \u5b9e\u4f8b\uff0c\u5206\u5e03\u5728 8 \u5f20 Ascend 910B NPU \u5361\u4e0a\n  \u547d\u540d\u89c4\u5219: vllm-ascend-paddleocr-{\u7aef\u53e3}-npu-{\u5361\u53f7}  \/  paddle-{\u7aef\u53e3}-npu-{\u5361\u53f7}\n<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b8c\u6574\u62d3\u6251\u53ef\u4ee5\u6982\u62ec\u4e3a\uff1a<strong>Uvicorn\u7f51\u5173\u8f6e\u8be248\u4e2apaddlex \u2192 \u6bcf6\u4e2apaddlex\u5171\u4eab\u4e00\u5f20NPU\u5361 \u2192 \u5361\u4e0a2\u4e2avllm\u5404\u81ea\u670d\u52a13\u4e2apaddlex \u2192 vllm\u901a\u8fc7Ascend 910B NPU\u5b8c\u6210\u6a21\u578b\u63a8\u7406\u3002<\/strong>\u6269\u5bb9\u65f6\u53ea\u9700\u5c06\u65b0\u5b9e\u4f8b\u5982\u7aef\u53e3\u53f7\u52a0\u5165 config.json \u7684 upstreams \u5217\u8868\uff0c\u7f51\u5173\u5373\u53ef\u81ea\u52a8\u7eb3\u5165\u8f6e\u8be2\uff0c\u65e0\u9700\u91cd\u542f\u5176\u4ed6\u670d\u52a1\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u63a8\u7406\u5f15\u64ce\u5c42\uff1avllm-ascend<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">vllm-ascend \u662f\u57fa\u4e8e\u534e\u4e3a\u6607\u817eNPU\u5b9e\u73b0\u7684\u9ad8\u6027\u80fd\u63a8\u7406\u5f15\u64ce\uff0c\u627f\u8f7d PaddleOCR-VL-1.5-0.9B \u6a21\u578b\u7684\u52a0\u8f7d\u548c\u63a8\u7406\u3002\u6bcf\u4e2a\u5b9e\u4f8b\u76d1\u542c\u4e00\u4e2a\u72ec\u7acb\u7aef\u53e3\uff0c\u901a\u8fc7 OpenAI \u517c\u5bb9\u7684 `\/v1\/chat\/completions` \u63a5\u53e3\u4e0e\u4e0a\u5c42 paddlex \u4ea4\u4e92\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5355\u5361\u53cc\u5b9e\u4f8b\u7684\u5173\u952e\u53c2\u6570 `--gpu-memory-utilization 0.45` \u786e\u4fdd\u4e24\u4e2avllm\u5b9e\u4f8b\u5e73\u5206\u4e00\u5f20\u5361\u7684\u663e\u5b58\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">OCR\u5904\u7406\u5c42\uff1apaddlex<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">paddlex \u662f PaddleOCR \u7684\u5b8c\u6574 Pipeline \u670d\u52a1\uff0c\u8d1f\u8d23\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>**\u6587\u6863\u9884\u5904\u7406**\uff08\u65b9\u5411\u6821\u6b63\u3001\u53bb\u626d\u66f2\uff09<\/li>\n\n\n\n<li>**\u7248\u9762\u68c0\u6d4b**\uff08PP-DocLayoutV3\uff09<\/li>\n\n\n\n<li>**VL\u8bc6\u522b**\uff08\u8c03\u7528vllm\u5b8c\u6210\u5b9e\u9645OCR\uff09<\/li>\n\n\n\n<li>**Markdown\u683c\u5f0f\u5316**\uff08\u7248\u9762\u91cd\u6392\u3001\u8868\u683c\/\u516c\u5f0f\u5904\u7406\uff09<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">paddlex \u901a\u8fc7 `server_url: http:\/\/127.0.0.1:{vllm\u7aef\u53e3}\/v1` \u6307\u5411\u5bf9\u5e94\u7684vllm\u5b9e\u4f8b\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u8d1f\u8f7d\u5747\u8861\u5c42\uff1auvicorn\u7f51\u5173<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">FastAPI \u670d\u52a1\u8fd0\u884c\u5728 `paddle-10020-npu-2` \u5bb9\u5668\u5185\u7684 9101 \u7aef\u53e3\uff0c\u5bf9\u6240\u6709 48 \u4e2a paddlex \u5b9e\u4f8b\u505a **Round-Robin \u8f6e\u8be2** + **failover\u5bb9\u9519**\uff0c\u7edf\u4e00\u5bf9\u5916\u63d0\u4f9b `\/ocr` \u548c `\/ocr-file` \u4e24\u4e2aAPI\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u56db\u3001\u5b8c\u6574\u7aef\u53e3\u89c4\u5212<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">vllm-ascend\uff08\u6bcf\u53612\u4e2a\uff0c\u517116\u4e2a\uff09<\/h3>\n\n\n\n<figure class=\"wp-block-table\"><table>\n<thead><tr><th>NPU\u5361<\/th><th>vllm\u5b9e\u4f8b1<\/th><th>vllm\u5b9e\u4f8b2<\/th><\/tr><\/thead>\n<tbody>\n<tr><td>\u53610<\/td><td>9031 (\u65b0\u589e)<\/td><td>9041 (\u65b0\u589e)<\/td><\/tr>\n<tr><td>\u53611<\/td><td>9032 (\u65b0\u589e)<\/td><td>9042 (\u65b0\u589e)<\/td><\/tr>\n<tr><td>\u53612<\/td><td>9025 (\u5df2\u6709)<\/td><td>9035 (\u65b0\u589e)<\/td><\/tr>\n<tr><td>\u53613<\/td><td>9026 (\u5df2\u6709)<\/td><td>9036 (\u65b0\u589e)<\/td><\/tr>\n<tr><td>\u53614<\/td><td>9027 (\u5df2\u6709)<\/td><td>9037 (\u65b0\u589e)<\/td><\/tr>\n<tr><td>\u53615<\/td><td>9028 (\u5df2\u6709)<\/td><td>9038 (\u65b0\u589e)<\/td><\/tr>\n<tr><td>\u53616<\/td><td>9029 (\u5df2\u6709)<\/td><td>9039 (\u65b0\u589e)<\/td><\/tr>\n<tr><td>\u53617<\/td><td>9030 (\u5df2\u6709)<\/td><td>9040 (\u65b0\u589e)<\/td><\/tr>\n<\/tbody>\n<\/table><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\">paddlex\uff08\u6bcf\u53616\u4e2a\uff0c\u517148\u4e2a\uff09<\/h3>\n\n\n\n<figure class=\"wp-block-table\"><table>\n<thead><tr><th>NPU\u5361<\/th><th>\u5b9e\u4f8b1<\/th><th>\u5b9e\u4f8b2<\/th><th>\u5b9e\u4f8b3<\/th><th>\u5b9e\u4f8b4<\/th><th>\u5b9e\u4f8b5<\/th><th>\u5b9e\u4f8b6<\/th><\/tr><\/thead>\n<tbody>\n<tr><td>\u53610<\/td><td>9126<\/td><td>9136<\/td><td>9146<\/td><td>9156<\/td><td>9166<\/td><td>9176<\/td><\/tr>\n<tr><td>\u53611<\/td><td>9127<\/td><td>9137<\/td><td>9147<\/td><td>9157<\/td><td>9167<\/td><td>9177<\/td><\/tr>\n<tr><td>\u53612<\/td><td>9120<\/td><td>9130<\/td><td>9140<\/td><td>9150<\/td><td>9160<\/td><td>9170<\/td><\/tr>\n<tr><td>\u53613<\/td><td>9121<\/td><td>9131<\/td><td>9141<\/td><td>9151<\/td><td>9161<\/td><td>9171<\/td><\/tr>\n<tr><td>\u53614<\/td><td>9122<\/td><td>9132<\/td><td>9142<\/td><td>9152<\/td><td>9162<\/td><td>9172<\/td><\/tr>\n<tr><td>\u53615<\/td><td>9123<\/td><td>9133<\/td><td>9143<\/td><td>9153<\/td><td>9163<\/td><td>9173<\/td><\/tr>\n<tr><td>\u53616<\/td><td>9124<\/td><td>9134<\/td><td>9144<\/td><td>9154<\/td><td>9164<\/td><td>9174<\/td><\/tr>\n<tr><td>\u53617<\/td><td>9125<\/td><td>9135<\/td><td>9145<\/td><td>9155<\/td><td>9165<\/td><td>9175<\/td><\/tr>\n<\/tbody>\n<\/table><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\">paddlex \u2192 vllm \u8def\u7531\u89c4\u5219<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u6bcf\u5361\u524d3\u4e2apaddlex\u5b9e\u4f8b\u6307\u5411vllm\u5b9e\u4f8b1\uff0c\u540e3\u4e2a\u6307\u5411\u5b9e\u4f8b2\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table>\n<thead><tr><th>NPU\u5361<\/th><th>\u2192 vllm\u5b9e\u4f8b1<\/th><th>\u2192 vllm\u5b9e\u4f8b2<\/th><\/tr><\/thead>\n<tbody>\n<tr><td>\u53610<\/td><td>9126\/9136\/9146 \u2192 9031<\/td><td>9156\/9166\/9176 \u2192 9041<\/td><\/tr>\n<tr><td>\u53611<\/td><td>9127\/9137\/9147 \u2192 9032<\/td><td>9157\/9167\/9177 \u2192 9042<\/td><\/tr>\n<tr><td>\u53612<\/td><td>9120\/9130\/9140 \u2192 9025<\/td><td>9150\/9160\/9170 \u2192 9035<\/td><\/tr>\n<tr><td>\u53613<\/td><td>9121\/9131\/9141 \u2192 9026<\/td><td>9151\/9161\/9171 \u2192 9036<\/td><\/tr>\n<tr><td>\u53614<\/td><td>9122\/9132\/9142 \u2192 9027<\/td><td>9152\/9162\/9172 \u2192 9037<\/td><\/tr>\n<tr><td>\u53615<\/td><td>9123\/9133\/9143 \u2192 9028<\/td><td>9153\/9163\/9173 \u2192 9038<\/td><\/tr>\n<tr><td>\u53616<\/td><td>9124\/9134\/9144 \u2192 9029<\/td><td>9154\/9164\/9174 \u2192 9039<\/td><\/tr>\n<tr><td>\u53617<\/td><td>9125\/9135\/9145 \u2192 9030<\/td><td>9155\/9165\/9175 \u2192 9040<\/td><\/tr>\n<\/tbody>\n<\/table><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e94\u3001vllm-ascend \u63a8\u7406\u5f15\u64ce\u90e8\u7f72<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">5.1 \u542f\u52a8\u811a\u672c<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u6bcf\u4e2avllm\u5b9e\u4f8b\u5bf9\u5e94\u4e00\u4e2a\u542f\u52a8\u811a\u672c `vllm-start-{\u7aef\u53e3}.sh`\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>#!\/bin\/bash\nexport TASK_QUEUE_ENABLE=1\nexport CPU_AFFINITY_CONF=1\nexport PYTORCH_NPU_ALLOC_CONF=\"expandable_segments:True\"\nexport VLLM_USE_MODELSCOPE=true\nexport MODEL_PATH=\"\/data\/PaddleOCR-VL-1.5-0.9B\"\nvllm serve ${MODEL_PATH} \\\n  --max-num-batched-tokens 16384 \\\n  --served-model-name PaddleOCR-VL-1.5-0.9B \\\n  --port __PORT__ \\\n  --trust-remote-code \\\n  --no-enable-prefix-caching \\\n  --mm-processor-cache-gb 0 \\\n  --gpu-memory-utilization 0.45 \\\n  --compilation-config '{\"cudagraph_mode\":\"FULL_DECODE_ONLY\"}' \\\n  --additional_config '{\"enable_cpu_binding\":true}'<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u5173\u952e\u53c2\u6570\u8bf4\u660e\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table>\n<thead><tr><th>\u53c2\u6570<\/th><th>\u503c<\/th><th>\u4f5c\u7528<\/th><\/tr><\/thead>\n<tbody>\n<tr><td>`--port`<\/td><td>9025~9042<\/td><td>\u6bcf\u4e2a\u5b9e\u4f8b\u72ec\u7acb\u7aef\u53e3<\/td><\/tr>\n<tr><td>`--gpu-memory-utilization`<\/td><td>0.45<\/td><td>\u5355\u5361\u53cc\u5b9e\u4f8b\uff0c\u5e73\u5206\u663e\u5b58<\/td><\/tr>\n<tr><td>`--max-num-batched-tokens`<\/td><td>16384<\/td><td>\u6700\u5927\u6279\u5904\u7406token\u6570<\/td><\/tr>\n<tr><td>`--mm-processor-cache-gb`<\/td><td>0<\/td><td>\u5173\u95ed\u591a\u6a21\u6001\u5904\u7406\u5668\u7f13\u5b58<\/td><\/tr>\n<tr><td>`--no-enable-prefix-caching`<\/td><td>\u2014<\/td><td>\u5173\u95ed\u524d\u7f00\u7f13\u5b58\uff08OCR\u573a\u666f\u65e0\u76ca\uff09<\/td><\/tr>\n<tr><td>`expandable_segments`<\/td><td>True<\/td><td>\u52a8\u6001\u663e\u5b58\u5206\u914d\uff08PyTorch NPU\u7279\u6027\uff09<\/td><\/tr>\n<\/tbody>\n<\/table><\/figure>\n\n\n\n<h3 class=\"wp-block-heading\">5.2 \u5bb9\u5668\u521b\u5efa<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u901a\u8fc7 `setup.sh config` \u81ea\u52a8\u6279\u91cf\u751f\u6210\u6240\u6709\u7aef\u53e3\u7684\u542f\u52a8\u811a\u672c\uff0c\u5bb9\u5668\u521b\u5efa\u7531 Docker Compose \u63a5\u7ba1\uff08\u89c1\u7b2c8\u8282\uff09\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">5.3 \u6a21\u578b\u52a0\u8f7d\u9a8c\u8bc1<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">vllm\u542f\u52a8\u540e\u7b49\u5f852-5\u5206\u949f\u6a21\u578b\u52a0\u8f7d\uff0c\u53ef\u901a\u8fc7API\u6d4b\u8bd5\u9a8c\u8bc1\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>curl http:\/\/127.0.0.1:9031\/v1\/models | python3 -m json.tool<\/code><\/pre>\n\n\n\n<h2 class=\"wp-block-heading\">\u516d\u3001paddle + paddlex OCR\u5904\u7406\u5c42\u90e8\u7f72<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">6.1 Pipeline\u914d\u7f6e\u6587\u4ef6<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u6bcf\u4e2apaddlex\u5b9e\u4f8b\u5bf9\u5e94\u4e00\u4e2a `pipeline_config_vllm-{\u7aef\u53e3}.yaml`\uff0c\u6838\u5fc3\u914d\u7f6e\u5982\u4e0b\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>pipeline_name: PaddleOCR-VL-1.5\nbatch_size: 128\nuse_queues: True\nuse_doc_preprocessor: True        # \u6587\u6863\u9884\u5904\u7406\nuse_layout_detection: True        # \u7248\u9762\u68c0\u6d4b\nuse_chart_recognition: False\nuse_seal_recognition: False\nformat_block_content: False\nmerge_layout_blocks: True\nSubModules:\n  LayoutDetection:\n    module_name: layout_detection\n    model_name: PP-DocLayoutV3\n    model_dir: \/root\/PP-DocLayoutV3\n    batch_size: 32\n  VLRecognition:\n    module_name: vl_recognition\n    model_name: PaddleOCR-VL-1.5-0.9B\n    batch_size: 16384\n    genai_config:\n      backend: vllm-server\n      server_url: http:\/\/127.0.0.1:9031\/v1  # \u2190 \u6307\u5411\u5bf9\u5e94vllm\u5b9e\u4f8b\nSubPipelines:\n  DocPreprocessor:\n    pipeline_name: doc_preprocessor\n    batch_size: 16\n    use_doc_orientation_classify: True\n    use_doc_unwarping: True\n    SubModules:\n      DocOrientationClassify:\n        model_name: PP-LCNet_x1_0_doc_ori\n        model_dir: \/data\/PP-LCNet_x1_0_doc_ori\n      DocUnwarping:\n        model_name: UVDoc\n        model_dir: \/data\/UVDoc<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u6ce8\u610f `server_url` \u5fc5\u987b\u6307\u5411\u540c\u5361\u5bf9\u5e94\u7684vllm\u5b9e\u4f8b\uff0c\u8fd9\u662fpaddlex\u4e0evllm\u4e4b\u95f4\u7684\u552f\u4e00\u7ebd\u5e26\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">6.2 Batch Size\u4f18\u5316<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u539f\u6709\u914d\u7f6e batch_size \u8f83\u5c0f\uff0c\u6269\u5bb9\u540e\u6839\u636e\u6bcf\u5361\u5b9e\u4f8b\u6570\u8fdb\u884c\u4e86\u8c03\u6574\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>sed -i \"s\/batch_size: 8\/batch_size: 16\/\" pipeline_config_vllm*\nsed -i \"s\/batch_size: 64\/batch_size: 128\/\" pipeline_config_vllm*\nsed -i \"s\/batch_size: 4096\/batch_size: 8192\/\" pipeline_config_vllm*<\/code><\/pre>\n\n\n\n<h3 class=\"wp-block-heading\">6.3 \u670d\u52a1\u542f\u52a8<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">paddlex\u901a\u8fc7 `docker exec` \u5728\u5bb9\u5668\u5185\u540e\u53f0\u542f\u52a8\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>docker exec -d paddle-10026-npu-0 bash -c \\\n  \"source ~\/.bashrc &amp;&amp; paddlex --serve --device npu --port 9126 \\\n   --pipeline \/data\/pipeline_config_vllm-9126.yaml \\\n   &gt; \/data\/paddle-9126.log 2&gt;&amp;1 &amp; wait\"<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u5173\u952e\u53c2\u6570\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table>\n<thead><tr><th>\u53c2\u6570<\/th><th>\u8bf4\u660e<\/th><\/tr><\/thead>\n<tbody>\n<tr><td>`--serve`<\/td><td>\u4ee5HTTP\u670d\u52a1\u6a21\u5f0f\u8fd0\u884c<\/td><\/tr>\n<tr><td>`--device npu`<\/td><td>\u4f7f\u7528NPU\u8bbe\u5907<\/td><\/tr>\n<tr><td>`--port`<\/td><td>paddlex\u670d\u52a1\u7aef\u53e3\uff0c\u4e0e\u5bb9\u5668\u540d\u4e2d\u7684\u7aef\u53e3\u4e00\u81f4<\/td><\/tr>\n<tr><td>`--pipeline`<\/td><td>\u6307\u5411\u5bf9\u5e94\u7684pipeline\u914d\u7f6e\u6587\u4ef6<\/td><\/tr>\n<\/tbody>\n<\/table><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e03\u3001\u8d1f\u8f7d\u5747\u8861\u7f51\u5173<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u6240\u6709 paddlex \u5b9e\u4f8b\u524d\u9762\u52a0\u4e86\u4e00\u5c42\u57fa\u4e8e FastAPI + uvicorn \u7684\u8d1f\u8f7d\u5747\u8861\u7f51\u5173\uff0c\u8d1f\u8d23\u5c06\u5916\u90e8\u8bf7\u6c42\u5747\u5300\u5206\u53d1\u523048\u4e2a\u4e0a\u6e38\u5b9e\u4f8b\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u6838\u5fc3\u80fd\u529b<\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Round-Robin \u8f6e\u8be2<\/strong>\uff1a\u5bf948\u4e2a\u4e0a\u6e38\u5b9e\u4f8b\u9010\u4e00\u5faa\u73af\u5206\u53d1\uff0c\u4fdd\u8bc1\u8d1f\u8f7d\u5747\u5300<\/li>\n\n\n\n<li><strong>Failover \u5bb9\u9519<\/strong>\uff1a\u5355\u6b21\u8bf7\u6c42\u6700\u591a\u5c1d\u8bd512\u4e2a\u4e0a\u6e38\uff0c\u5f53\u67d0\u5b9e\u4f8b\u6545\u969c\u65f6\u81ea\u52a8\u5207\u6362\u5230\u4e0b\u4e00\u4e2a<\/li>\n\n\n\n<li><strong>\u7edf\u4e00API<\/strong>\uff1a\u63d0\u4f9b <code>POST \/ocr<\/code>\uff08JSON\u5165\u53c2\uff0c\u652f\u6301base64\/URL\/local_path\uff09\u548c <code>POST \/ocr-file<\/code>\uff08multipart\u6587\u4ef6\u4e0a\u4f20\uff09<\/li>\n\n\n\n<li><strong>OCR\u540e\u5904\u7406<\/strong>\uff1a\u53bb\u9664\u56fe\u7247\u6807\u7b7e\u3001\u6e05\u7406\u8bc6\u522b\u566a\u70b9\u3001\u5408\u5e76\u7a7a\u767d\u884c<\/li>\n\n\n\n<li><strong>\u8fde\u63a5\u6c60\u590d\u7528<\/strong>\uff1a\u57fa\u4e8e httpx.AsyncClient \u7684 HTTP \u8fde\u63a5\u6c60\uff0c\u907f\u514d\u53cd\u590d\u5efa\u8fde\u5f00\u9500<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">\u914d\u7f6e\u9a71\u52a8<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u7f51\u5173\u901a\u8fc7 <code>config.json<\/code> \u9a71\u52a8\uff0c\u6838\u5fc3\u914d\u7f6e\u9879\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>{\n  \"upstreams\": [\n    \"http:\/\/127.0.0.1:9120\", \"http:\/\/127.0.0.1:9121\",\n    \/\/ ... \u517148\u4e2a paddlex \u5b9e\u4f8b\n    \"http:\/\/127.0.0.1:9177\"\n  ],\n  \"load_balance\": {\n    \"mode\": \"round_robin\",    \/\/ round_robin \u6216 random\n    \"max_attempts\": 12        \/\/ \u5355\u6b21\u8bf7\u6c42\u6700\u591a\u5c1d\u8bd5\u7684\u4e0a\u6e38\u6570\n  },\n  \"http\": {\n    \"timeout_s\": 300,         \/\/ OCR\u5355\u6b21\u8bf7\u6c42\u53ef\u80fd\u8017\u65f6\u8f83\u957f\n    \"pool_max_connections\": 200,\n    \"pool_max_keepalive_connections\": 100\n  },\n  \"postprocess\": {\n    \"strip_images\": true,     \/\/ \u53bb\u9664\u7ed3\u679c\u4e2d\u7684\u56fe\u7247Markdown\u6807\u7b7e\n    \"remove_repeated_hui\": {  \/\/ \u6e05\u7406\u7b7e\u540d\u533a\u8bc6\u522b\u566a\u70b9\uff08\u8fde\u7eed\"\u56de\"\u5b57\uff09\n      \"enabled\": true,\n      \"min_repeats\": 2\n    }\n  }\n}<\/code><\/pre>\n\n\n\n<h3 class=\"wp-block-heading\">Failover \u6d41\u7a0b<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u7f51\u5173\u4e0e\u6bcf\u4e2a paddlex \u5b9e\u4f8b\u7684\u4ea4\u4e92\u9700\u8981\u4e24\u7ea7\u8c03\u7528\uff1a<code>\/layout-parsing<\/code> \u5b8c\u6210\u7248\u9762\u89e3\u6790\uff0c<code>\/restructure-pages<\/code> \u5b8c\u6210 Markdown \u91cd\u7ec4\u3002\u5f53\u67d0\u4e2a\u4e0a\u6e38\u53d1\u751f\u8d85\u65f6\u6216\u9519\u8bef\u65f6\uff0c\u7f51\u5173\u81ea\u52a8\u8df3\u8fc7\u8be5\u5b9e\u4f8b\uff0c\u5c1d\u8bd5\u5e8f\u5217\u4e2d\u7684\u4e0b\u4e00\u4e2a\u3002\u53ea\u6709\u6240\u6709 max_attempts \u6b21\u5c1d\u8bd5\u5747\u5931\u8d25\u540e\u624d\u8fd4\u56de 502\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u5ba2\u6237\u7aef\u8c03\u7528\u793a\u4f8b<\/h3>\n\n\n\n<pre class=\"wp-block-code\"><code># \u65b9\u5f0f1: base64 \u7f16\u7801\nimport requests, base64\nwith open(\"doc.png\", \"rb\") as f:\n    b64 = base64.b64encode(f.read()).decode()\nr = requests.post(\"http:\/\/127.0.0.1:9101\/ocr\",\n                  json={\"base64\": b64}, timeout=300)\nprint(r.json()[\"text\"])\n# \u65b9\u5f0f2: \u6587\u4ef6\u4e0a\u4f20\nr = requests.post(\"http:\/\/127.0.0.1:9101\/ocr-file\",\n                  files={\"file\": open(\"doc.png\", \"rb\")}, timeout=300)\nprint(r.json()[\"text\"])<\/code><\/pre>\n\n\n\n<h3 class=\"wp-block-heading\">\u9a8c\u8bc1\u8d1f\u8f7d\u5747\u8861<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fde\u7eed\u8bf7\u6c42\u89c2\u5bdf\u8fd4\u56de\u7684 <code>upstream<\/code> \u5b57\u6bb5\u662f\u5426\u8f6e\u8be2\u53d8\u5316\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>for i in {1..8}; do\n  curl -s -X POST http:\/\/127.0.0.1:9101\/ocr     -H \"Content-Type: application\/json\"     -d '{\"local_path\":\".\/test.png\"}'   | python -c 'import sys,json; d=json.load(sys.stdin); print(d.get(\"upstream\"))'\ndone<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u7f51\u5173\u6838\u5fc3\u4ee3\u7801\u57fa\u4e8e FastAPI + httpx \u5b9e\u73b0\uff0c\u4e0d\u5230400\u884c\u3002\u65b0\u589e\u5b9e\u4f8b\u53ea\u9700\u52a0\u5230 config.json \u7684 upstreams \u5217\u8868\u4e2d\u5373\u53ef\u81ea\u52a8\u7eb3\u5165\u8f6e\u8be2\uff0c\u65e0\u9700\u91cd\u542f\u5176\u4ed6\u670d\u52a1\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u516b\u3001Docker Compose \u7f16\u6392<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u7531\u4e8e\u670d\u52a1\u5668\u65e0\u6cd5\u4f7f\u7528K8s\uff0c\u6574\u4e2a\u96c6\u7fa4\u901a\u8fc7 Docker Compose \u7f16\u6392\u7ba1\u7406\u3002\u5206\u4e3a\u4e24\u7ec4compose\u6587\u4ef6\uff1a<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">8.1 all-docker-compose.vllm.yml<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">16\u4e2avllm\u5bb9\u5668\u5b9a\u4e49\u3002\u4ee5\u53610\u53cc\u5b9e\u4f8b\u4e3a\u4f8b\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>version: \"3.7\"\nx-vllm-common: &amp;vllm-common\n  image: quay.io\/ascend\/vllm-ascend:v0.19.1rc1\n  restart: unless-stopped\n  network_mode: host          # \u76f4\u63a5\u4f7f\u7528\u5bbf\u4e3b\u673a\u7f51\u7edc\n  privileged: true            # NPU\u8bbe\u5907\u8bbf\u95ee\u9700\u8981\u7279\u6743\u6a21\u5f0f\n  user: \"0\"\n  shm_size: \"100g\"\n  devices:                    # NPU\u7ba1\u7406\u8bbe\u5907\n    - \/dev\/davinci_manager\n    - \/dev\/devmm_svm\n    - \/dev\/hisi_hdc\n  volumes:                    # NPU\u9a71\u52a8 + \u6570\u636e\u6302\u8f7d\n    - \/usr\/local\/dcmi:\/usr\/local\/dcmi\n    - \/usr\/local\/bin\/npu-smi:\/usr\/local\/bin\/npu-smi\n    - \/usr\/local\/Ascend\/driver\/lib64\/:\/usr\/local\/Ascend\/driver\/lib64\/\n    - \/usr\/local\/Ascend\/driver\/version.info:\/usr\/local\/Ascend\/driver\/version.info\n    - \/etc\/ascend_install.info:\/etc\/ascend_install.info\n    - \/opt\/its\/model\/PaddleOCR:\/data\/\nservices:\n  vllm-9031-npu-0:\n    &lt;&lt;: *vllm-common\n    container_name: vllm-ascend-paddleocr-9031-npu-0\n    devices:\n      - \/dev\/davinci0              # \u2190 \u7ed1\u5b9a\u53610\n    environment:\n      ASCEND_RT_VISIBLE_DEVICES: \"0\"\n    command: bash -c \"\/data\/vllm-start-9031.sh &amp;&gt; \/data\/vllm-9031.log &amp; wait\"\n  vllm-9041-npu-0:\n    &lt;&lt;: *vllm-common\n    container_name: vllm-ascend-paddleocr-9041-npu-0\n    devices:\n      - \/dev\/davinci0              # \u2190 \u4e0e9031\u5171\u4eab\u53610\n    environment:\n      ASCEND_RT_VISIBLE_DEVICES: \"0\"\n    command: bash -c \"\/data\/vllm-start-9041.sh &amp;&gt; \/data\/vllm-9041.log &amp; wait\"<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u5173\u952e\u8bbe\u8ba1\u70b9\uff1a<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">1. **YAML\u951a\u70b9\u590d\u7528**\uff1a`x-vllm-common: &vllm-common` \u5b9a\u4e49\u516c\u5171\u914d\u7f6e\uff0c\u6240\u6709\u670d\u52a1\u901a\u8fc7 `<<: *vllm-common` \u7ee7\u627f\uff0c\u5927\u5e45\u51cf\u5c11\u91cd\u590d\u4ee3\u7801 2. **network_mode: host**\uff1aNPU\u63a8\u7406\u670d\u52a1\u76f4\u63a5\u4f7f\u7528\u5bbf\u4e3b\u673a\u7aef\u53e3\uff0c\u65e0\u9700\u7aef\u53e3\u6620\u5c04 3. **\u5355\u5361\u53cc\u5b9e\u4f8b**\uff1a\u4e24\u4e2a\u5bb9\u5668\u90fd\u7ed1\u5b9a\u540c\u4e00\u5f20NPU\u5361\uff08\u5982 `\/dev\/davinci0`\uff09\uff0c\u901a\u8fc7 `--gpu-memory-utilization 0.45` \u5b9e\u73b0\u663e\u5b58\u5e73\u5206 4. **command**\uff1a\u5bb9\u5668\u542f\u52a8\u65f6\u81ea\u52a8\u6267\u884c\u542f\u52a8\u811a\u672c\uff0cvllm\u670d\u52a1\u4f5c\u4e3a\u5bb9\u5668\u4e3b\u8fdb\u7a0b\u8fd0\u884c<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">8.2 all-docker-compose.paddle.yml<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">48\u4e2apaddle\u5bb9\u5668\u5b9a\u4e49\uff0c\u7ed3\u6784\u4e0evllm\u7c7b\u4f3c\u3002\u5dee\u5f02\u70b9\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>x-paddle-common: &amp;paddle-common\n  image: paddleocrvl:20260529\n  restart: unless-stopped\n  network_mode: host\n  privileged: true\n  shm_size: \"128G\"                # paddle\u9700\u8981\u66f4\u5927\u5171\u4eab\u5185\u5b58\n  volumes:\n    - \/usr\/local\/Ascend\/driver:\/usr\/local\/Ascend\/driver\n    - \/opt\/its\/model\/PaddleOCR:\/data\/  # \u914d\u7f6e\u6587\u4ef6 + \u6a21\u578b + \u811a\u672c\n  command: tail -f \/dev\/null      # \u5bb9\u5668\u4fdd\u6301\u8fd0\u884c\uff0cpaddlex\u901a\u8fc7exec\u542f\u52a8<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u6ce8\u610f paddle \u5bb9\u5668\u7684 `command: tail -f \/dev\/null` \u2014\u2014 \u5bb9\u5668\u4fdd\u6301\u5b58\u6d3b\u4f46\u4e0d\u81ea\u52a8\u542f\u52a8paddlex\u670d\u52a1\uff0cpaddlex\u901a\u8fc7\u5916\u90e8 `docker exec` \u547d\u4ee4\u5728\u5bb9\u5668\u5185\u624b\u52a8\u542f\u52a8\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u662f\u56e0\u4e3a paddle \u5bb9\u5668\u5185\u9700\u8981\u5148 `source ~\/.bashrc` \u521d\u59cb\u5316NPU\u73af\u5883\u53d8\u91cf\uff0c\u4e14paddlex\u670d\u52a1\u7684\u542f\u52a8\u4f9d\u8d56\u4e8e\u5bf9\u5e94vllm\u5b9e\u4f8b\u5df2\u5c31\u7eea\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u4e5d\u3001\u4e00\u952e\u90e8\u7f72\u4e0e\u8fd0\u7ef4<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">9.1 setup.sh \u81ea\u52a8\u5316\u90e8\u7f72<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u90e8\u7f72\u811a\u672c `setup.sh` \u63d0\u4f9b\u5b8c\u6574\u7684\u81ea\u52a8\u5316\u90e8\u7f72\u6d41\u7a0b\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>bash setup.sh all  # \u4e00\u952e\u5b8c\u6574\u90e8\u7f72<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u6267\u884c\u8fc7\u7a0b\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>\u68c0\u67e5\u7aef\u53e3\u5360\u7528\n  \u2193\n\u751f\u6210 vllm \u542f\u52a8\u811a\u672c\uff0816\u4e2a\uff09\n  \u2193\n\u751f\u6210 paddle \u914d\u7f6e\u6587\u4ef6\uff0848\u4e2a\uff09\n  \u2193\nDocker Compose \u542f\u52a8 vllm \u5bb9\u5668\n  \u2193\n\u7b49\u5f85\u6a21\u578b\u52a0\u8f7d\u5b8c\u6210\uff08\u4eba\u5de5\u786e\u8ba4\uff09\n  \u2193\nDocker Compose \u542f\u52a8 paddle \u5bb9\u5668\n  \u2193\ndocker exec \u5728\u6bcf\u4e2a\u5bb9\u5668\u5185\u542f\u52a8 paddlex \u670d\u52a1\n  \u2193\n\u66f4\u65b0\u8d1f\u8f7d\u5747\u8861\u914d\u7f6e \u2192 \u91cd\u542f uvicorn<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e5f\u652f\u6301\u5206\u6b65\u6267\u884c\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>bash setup.sh check     # \u68c0\u67e5\u7aef\u53e3\nbash setup.sh config    # \u751f\u6210\u914d\u7f6e\nbash setup.sh vllm      # \u542f\u52a8vllm\nbash setup.sh paddle    # \u542f\u52a8paddle\nbash setup.sh lb        # \u66f4\u65b0\u8d1f\u8f7d\u5747\u8861\nbash setup.sh verify    # \u9a8c\u8bc1\u6240\u6709\u670d\u52a1<\/code><\/pre>\n\n\n\n<h3 class=\"wp-block-heading\">9.2 Pipeline\u914d\u7f6e\u6587\u4ef6\u81ea\u52a8\u751f\u6210<\/h3>\n\n\n\n<pre class=\"wp-block-code\"><code># \u6838\u5fc3\u903b\u8f91\uff1a\u590d\u5236\u6a21\u677f + \u66ff\u6362server_url\u6307\u5411\u5bf9\u5e94vllm\ngen_one() {\n    local P_PORT=$1 V_PORT=$2\n    cp pipeline_config_vllm.yaml pipeline_config_vllm-${P_PORT}.yaml\n    sed -i \"s\/127.0.0.1:9000\/127.0.0.1:${V_PORT}\/g\" \\\n        pipeline_config_vllm-${P_PORT}.yaml\n}\n# \u53610\uff1a\u524d3\u21929031\uff0c\u540e3\u21929041\ngen_one 9126 9031; gen_one 9136 9031; gen_one 9146 9031\ngen_one 9156 9041; gen_one 9166 9041; gen_one 9176 9041\n# ... \u4f9d\u6b64\u7c7b\u63a8\uff0c\u517148\u4e2a\u914d\u7f6e<\/code><\/pre>\n\n\n\n<h3 class=\"wp-block-heading\">9.3 \u8d1f\u8f7d\u5747\u8861\u914d\u7f6e\u81ea\u52a8\u66f4\u65b0<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">`setup.sh lb` \u901a\u8fc7 `docker exec` \u5728\u7f51\u5173\u5bb9\u5668\u5185\u66f4\u65b0\u914d\u7f6e\u5e76\u91cd\u542f\u670d\u52a1\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code># 1. \u5199\u5165\u65b0\u7684 config.json\uff0848\u4e2aupstream\uff09\n# 2. \u81ea\u52a8\u5907\u4efd\u65e7\u914d\u7f6e\u4e3a config.json.bak.{\u65f6\u95f4\u6233}\n# 3. \u67e5\u627e\u5e76\u6740\u6389\u65e7\u7684 uvicorn \u8fdb\u7a0b\n# 4. \u91cd\u65b0\u542f\u52a8 uvicorn<\/code><\/pre>\n\n\n\n<h3 class=\"wp-block-heading\">9.4 ops.sh \u8fd0\u7ef4\u5de5\u5177<\/h3>\n\n\n\n<pre class=\"wp-block-code\"><code>bash ops.sh status              # \u67e5\u770b\u6240\u6709\u5bb9\u5668\u72b6\u6001\nbash ops.sh ports               # \u6279\u91cf\u68c0\u67e5\u6240\u6709\u7aef\u53e3\nbash ops.sh restart-vllm 9031   # \u91cd\u542f\u6307\u5b9avllm\nbash ops.sh restart-paddle 9126 # \u91cd\u542f\u6307\u5b9apaddle\nbash ops.sh log-vllm 9031 100   # \u67e5\u770bvllm\u65e5\u5fd7\nbash ops.sh log-paddle 9126 100 # \u67e5\u770bpaddle\u65e5\u5fd7\nbash ops.sh test-vllm 9035      # \u6d4b\u8bd5vllm\u63a5\u53e3\nbash ops.sh test-paddle 9126    # \u6d4b\u8bd5paddle\u63a5\u53e3\nbash ops.sh npu                 # \u67e5\u770bNPU\u72b6\u6001\nbash ops.sh start-all           # \u4e00\u952e\u542f\u52a8\u6240\u6709\u5df2\u505c\u6b62\u7684\u5bb9\u5668\nbash ops.sh up-vllm             # Compose\u6279\u91cf\u64cd\u4f5c\nbash ops.sh down-paddle<\/code><\/pre>\n\n\n\n<h2 class=\"wp-block-heading\">\u5341\u3001\u8bbe\u8ba1\u51b3\u7b56\u4e0e\u7ecf\u9a8c\u603b\u7ed3<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">\u4e3a\u4ec0\u4e48\u4e0d\u80fd\u76f4\u63a5\u7528Nginx\u505a\u8d1f\u8f7d\u5747\u8861\uff1f<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">Nginx\u786e\u5b9e\u80fd\u505aHTTP\u53cd\u5411\u4ee3\u7406\u548c\u8f6e\u8be2\uff0c\u4f46\u5728OCR\u573a\u666f\u6709\u51e0\u4e2a\u95ee\u9898\uff1a<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">1. **\u4e24\u7ea7API\u8c03\u7528**\uff1aOCR\u9700\u8981\u5148\u8c03 `\/layout-parsing` \u518d\u8c03 `\/restructure-pages`\uff0cNginx\u65e0\u6cd5\u4e32\u8054 2. **\u540e\u5904\u7406\u9700\u6c42**\uff1aOCR\u7ed3\u679c\u9700\u8981strip images\u3001\u53bb\u9664\"\u56de\"\u5b57\u566a\u70b9\u7b49\u5b9a\u5236\u5316\u540e\u5904\u7406 3. **\u7075\u6d3b\u7684\u5165\u53c2\u652f\u6301**\uff1a\u9700\u8981\u540c\u65f6\u652f\u6301URL\/base64\/data_url\/local_path\u591a\u79cd\u8f93\u5165\u65b9\u5f0f 4. **\u7cbe\u7ec6\u7684failover\u63a7\u5236**\uff1a\u9700\u8981\u8fd4\u56de\u5177\u4f53\u54ea\u4e2a\u4e0a\u6e38\u5931\u8d25\u3001\u91cd\u8bd5\u4e86\u591a\u5c11\u6b21<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7528FastAPI\u81ea\u7814\u7f51\u5173\uff0c\u4ee3\u7801\u4e0d\u5230400\u884c\u5c31\u5b8c\u7f8e\u89e3\u51b3\u4e86\u8fd9\u4e9b\u95ee\u9898\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u4e3a\u4ec0\u4e48paddlex\u4e0d\u968f\u5bb9\u5668\u81ea\u52a8\u542f\u52a8\uff1f<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">paddlex\u9700\u8981\u5148\u52a0\u8f7d\u6a21\u578b\uff0c\u800c\u6a21\u578b\u52a0\u8f7d\u4f9d\u8d56vllm\u5df2\u5c31\u7eea\u3002\u5982\u679c\u968f\u5bb9\u5668\u81ea\u52a8\u542f\u52a8\uff0c\u65f6\u5e8f\u4e0d\u53ef\u63a7\u3002\u901a\u8fc7 `docker exec` \u5916\u90e8\u63a7\u5236\uff0c\u53ef\u4ee5\u7b49vllm\u5168\u90e8\u5c31\u7eea\u540e\u518d\u9010\u6279\u542f\u52a8paddlex\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u5355\u5361\u591a\u5b9e\u4f8b\u7684GPU\u5185\u5b58\u89c4\u5212<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">vllm\u5355\u5361\u53cc\u5b9e\u4f8b\u7684\u6838\u5fc3\u53c2\u6570\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>--gpu-memory-utilization 0.45<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u6bcf\u4e2a\u5b9e\u4f8b\u6700\u591a\u4f7f\u752845%\u7684NPU\u663e\u5b58\uff0c\u4e24\u4e2a\u5b9e\u4f8b\u5408\u8ba190%\uff0c\u755910%\u7ed9\u7cfb\u7edf\u5f00\u9500\u3002\u8fd9\u662f\u7ecf\u8fc7\u5b9e\u9645\u538b\u6d4b\u9a8c\u8bc1\u7684\u503c\u2014\u2014\u592a\u4f4e\u6d6a\u8d39\u7b97\u529b\uff0c\u592a\u9ad8\u4f1aOOM\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">paddlex\u7684batch_size\u4e5f\u505a\u4e86\u5bf9\u5e94\u8c03\u6574\uff1a\u4ece\u539f\u672c8\/64\/4096\u63d0\u5347\u523016\/128\/8192\uff0c\u5145\u5206\u5229\u7528\u5355\u5361\u8dd1\u591a\u4e2a\u5b9e\u4f8b\u7684\u5269\u4f59\u7b97\u529b\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u547d\u540d\u89c4\u8303\u7684\u5de5\u7a0b\u4ef7\u503c<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">\u6574\u4e2a\u9879\u76ee\u6709\u7edf\u4e00\u7684\u547d\u540d\u89c4\u8303\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>\u89c4\u5219                    \u793a\u4f8b\nvllm\u5bb9\u5668: vllm-ascend-paddleocr-{\u7aef\u53e3}-npu-{\u5361\u53f7}   vllm-ascend-paddleocr-9031-npu-0\npaddle\u5bb9\u5668: paddle-{\u7aef\u53e3}-npu-{\u5361\u53f7}               paddle-10026-npu-0\n\u542f\u52a8\u811a\u672c: vllm-start-{\u7aef\u53e3}.sh                     vllm-start-9031.sh\n\u914d\u7f6e\u6587\u4ef6: pipeline_config_vllm-{\u7aef\u53e3}.yaml         pipeline_config_vllm-9126.yaml<\/code><\/pre>\n\n\n\n<p class=\"wp-block-paragraph\">\u7aef\u53e3\u53f7\u672c\u8eab\u5c31\u7f16\u7801\u4e86\u7c7b\u578b\u548c\u5f52\u5c5e\u4fe1\u606f\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>81xx \u2192 vllm\u7aef\u53e3<\/li>\n\n\n\n<li>82xx\/83xx\/84xx\/85xx\/86xx\/87xx \u2192 paddle\u7aef\u53e3\uff0c\u524d\u4e24\u4f4d\u533a\u5206\u5b9e\u4f8b\u7f16\u53f7<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u8fd9\u8ba9\u81ea\u52a8\u5316\u811a\u672c\u53ef\u4ee5\u7b80\u5355\u5730\u901a\u8fc7\u7aef\u53e3\u53f7\u63a8\u5bfc\u51fa\u6240\u6709\u5173\u8054\u6587\u4ef6\u7684\u8def\u5f84\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u7ed3\u8bed<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">\u4ece\u6700\u521d\u76846\u4e2avllm+12\u4e2apaddle\uff0c\u5230\u6700\u7ec8\u768416\u4e2avllm+48\u4e2apaddle\u96c6\u7fa4\uff0c\u8fd9\u5957\u65b9\u6848\u7ecf\u5386\u4e86\u4ece\u624b\u52a8 `docker run` \u5230 Docker Compose \u7f16\u6392\uff0c\u4ece\u9759\u6001\u8f6e\u8be2\u5230\u81ea\u7814uvicorn\u667a\u80fd\u7f51\u5173\u7684\u5b8c\u6574\u6f14\u8fdb\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6838\u5fc3\u4ee3\u7801 `ocr_backend.py` \u4e0d\u5230400\u884c\uff0c\u914d\u5408 `config.json` \u9a71\u52a8\uff0c\u5b9e\u73b0\u4e86\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>48\u4e2a\u4e0a\u6e38\u5b9e\u4f8b\u7684Round-Robin\u8d1f\u8f7d\u5747\u8861<\/li>\n\n\n\n<li>\u8bf7\u6c42\u7ea7failover\u5bb9\u9519<\/li>\n\n\n\n<li>\u7075\u6d3b\u7684\u56fe\u7247\u8f93\u5165\u65b9\u5f0f\uff08URL\/base64\/local_path\/multipart\uff09<\/li>\n\n\n\n<li>\u5b8c\u5584\u7684OCR\u540e\u5904\u7406Pipeline<\/li>\n\n\n\n<li>\u53ef\u52a8\u6001\u6269\u5c55\uff08\u65b0\u589e\u5b9e\u4f8b\u53ea\u9700\u52a0\u5230config.json\uff09<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u6574\u5957\u90e8\u7f72\u5b9e\u73b0\u4e86\u5168\u81ea\u52a8\u5316\uff08`bash setup.sh all`\uff09\uff0c\u4ece\u96f6\u5230\u53ef\u7528\u53ea\u9700\u4e00\u6761\u547d\u4ee4\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7b80\u8bb0\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u4e00\u3001\u80cc\u666f\u4e0e\u52a8\u673a \u5728\u5927\u89c4\u6a21\u4f01\u4e1a\u6587\u6863\u5904\u7406\u7684OCR\u573a\u666f\u4e2d\uff0c\u9700\u8981\u5bf9\u6d77\u91cf\u626b\u63cf\u4ef6\/\u6587\u6863\u56fe\u7247\u8fdb\u884c\u9ad8\u5e76\u53d1\u8bc6\u522b\u3002\u5355\u5b9e\u4f8bPaddleOCR\u9762\u5bf9\u5927\u91cf\u8bf7\u6c42 &#8230;<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-1698","post","type-post","status-publish","format-standard","hentry","category-uncategorized"],"_links":{"self":[{"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/posts\/1698","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/comments?post=1698"}],"version-history":[{"count":3,"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/posts\/1698\/revisions"}],"predecessor-version":[{"id":1701,"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/posts\/1698\/revisions\/1701"}],"wp:attachment":[{"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/media?parent=1698"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/categories?post=1698"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.izhuhn.cn\/index.php\/wp-json\/wp\/v2\/tags?post=1698"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}