{"id":14891,"date":"2025-05-05T23:00:37","date_gmt":"2025-05-05T15:00:37","guid":{"rendered":"https:\/\/www.ztuzx.com\/?p=14891"},"modified":"2025-05-05T23:00:37","modified_gmt":"2025-05-05T15:00:37","slug":"%e8%8b%b1%e7%89%b9%e5%b0%94%e7%ac%ac%e4%b8%80%e6%97%b6%e9%97%b4%e6%b7%b1%e5%ba%a6%e4%bc%98%e5%8c%96qwen3%e5%a4%a7%e6%a8%a1%e5%9e%8b%ef%bc%8c%e5%8d%87%e7%ba%a7ai-pc%e8%83%bd%e5%8a%9b%e8%b5%8b%e8%83%bd","status":"publish","type":"post","link":"https:\/\/www.ztuzx.com\/?p=14891","title":{"rendered":"\u82f1\u7279\u5c14\u7b2c\u4e00\u65f6\u95f4\u6df1\u5ea6\u4f18\u5316Qwen3\u5927\u6a21\u578b\uff0c\u5347\u7ea7AI PC\u80fd\u529b\u8d4b\u80fd\u591a\u6837\u5316\u573a\u666f"},"content":{"rendered":"

\u4e3a\u6301\u7eed\u63a8\u52a8\u5c16\u7aefAI\u6a21\u578b\u7684\u521b\u65b0\u53d1\u5c55\uff0c\u82f1\u7279\u5c14\u59cb\u7ec8\u4e0e\u4e1a\u754c\u9886\u5148\u7684\u521b\u65b0\u529b\u91cf\u4fdd\u6301\u6df1\u5ea6\u534f\u4f5c\u3002\u6211\u4eec\u6b23\u7136\u5ba3\u5e03\uff0c\u5168\u65b0\u5347\u7ea7\u7684\u82f1\u7279\u5c14AI\u89e3\u51b3\u65b9\u6848\u5df2\u5168\u9762\u8986\u76d6PC\u5ba2\u6237\u7aef\u3001\u8fb9\u7f18\u8ba1\u7b97\u3001\u667a\u80fd\u9a7e\u8231\u7b49\u573a\u666f\uff0c\u5728\u7b2c\u4e00\u65f6\u95f4\u4e3aQwen3\u7cfb\u5217\u5927\u6a21\u578b\u7684\u53d1\u5e03\u63d0\u4f9b\u6280\u672f\u652f\u6491\u3002<\/p>\n

\u5168\u65b0Qwen3\u7cfb\u5217\u5927\u6a21\u578b\u7684\u4e94\u5927\u4eae\u70b9\uff1a<\/p>\n

\u00b7\u7a00\u758fMoE\u6a21\u578b\u6df1\u5ea6\u4f18\u5316\uff0c\u7aef\u4fa7\u9ad8\u6548\u90e8\u7f72Qwen3<\/p>\n

\u00b7\u9996\u6b21\u5728NPU\u4e0aDay 0\u652f\u6301\u5927\u6a21\u578b\uff0c\u63d0\u4f9b\u66f4\u597d\u7684\u6027\u80fd\u548c\u529f\u8017\u8868\u73b0<\/p>\n

\u00b7\u7aef\u4fa7\u5fae\u8c03\uff0c\u63d0\u5347\u6a21\u578b\u667a\u80fd\uff0c\u4f18\u5316\u7528\u6237\u4f53\u9a8c<\/p>\n

\u00b7\u52a8\u6001\u7a00\u758f\u6ce8\u610f\u529b\u8d4b\u80fdQwen3\u957f\u4e0a\u4e0b\u6587\u7a97\u53e3\uff0c\u89e3\u9501\u7aef\u4fa7Agent\u65b0\u5e94\u7528<\/p>\n

\u00b7\u62e5\u62b1\u5f00\u6e90\u751f\u6001\uff0cDay 0\u652f\u6301\u9b54\u642d\u793e\u533aOllama<\/p>\n

\u5728\u672c\u6b21\u963f\u91cc\u5df4\u5df4\u5f00\u6e90\u7684\u65b0\u4e00\u4ee3\u901a\u4e49\u5343\u95eeQwen3\u7cfb\u5217\u6a21\u578b\u4e2d\uff0c\u6700\u4e3a\u5f15\u4eba\u6ce8\u76ee\u7684\u662f30B\u53c2\u6570\u89c4\u6a21\u7684MoE\u6df7\u5408\u4e13\u5bb6\u6a21\u578b\uff08Mixtureof Experts\uff09\uff08Qwen3-30B-MOE-A3B\uff09\u3002\u8be5\u6a21\u578b\u51ed\u501f\u5176\u5148\u8fdb\u7684\u52a8\u6001\u53ef\u8c03\u6df7\u5408\u4e13\u5bb6\u67b6\u6784\uff0c\u5728\u6709\u6548\u63d0\u5347\u8ba1\u7b97\u6548\u7387\u65b9\u9762\u8868\u73b0\u51fa\u8272,\u4f7f\u5176\u5728\u672c\u5730\u8bbe\u5907\uff08\u5ba2\u6237\u7aef\u548c\u8fb9\u7f18\u8bbe\u5907\uff09\u4e0a\u5177\u5907\u5e7f\u9614\u7684\u5e94\u7528\u524d\u666f\u3002\u7136\u800c\uff0c\u5176\u5728\u90e8\u7f72\u73af\u8282\u5b58\u5728\u8f83\u5927\u96be\u5ea6\uff0c\u4e14\u5bf9\u7cfb\u7edf\u6574\u4f53\u8d44\u6e90\u7684\u6d88\u8017\u4f9d\u7136\u8f83\u9ad8\u3002\u4e3a\u5e94\u5bf9\u8fd9\u4e9b\u6311\u6218\uff0c\u82f1\u7279\u5c14\u4e0e\u963f\u91cc\u7d27\u5bc6\u5408\u4f5c\uff0c\u9488\u5bf9MoE\u6a21\u578b\u5c55\u5f00\u4e86\u5168\u9762\u7684\u6280\u672f\u9002\u914d\u5de5\u4f5c\u3002\u901a\u8fc7\u5b9e\u65bd\u591a\u79cd\u4f18\u5316\u7b56\u7565\uff0c\u57fa\u4e8eOpenVINO\u5de5\u5177\u5957\u4ef6\u6210\u529f\u5730\u5c06Qwen\u6a21\u578b\u9ad8\u6548\u90e8\u7f72\u4e8e\u82f1\u7279\u5c14\u786c\u4ef6\u5e73\u53f0\u4e4b\u4e0a\u3002\u5177\u4f53\u800c\u8a00\uff0c\u5728ARL-H64G\u5185\u5b58\u7684\u7cfb\u7edf\u4e0a\u90e8\u7f72\u768430B\u53c2\u6570\u89c4\u6a21MoE\u6a21\u578b\uff0c\u5b9e\u73b0\u4e8633.97token\/s\u7684\u541e\u5410\u91cf\uff0c\u800c\u4e14\u76f8\u8f83\u4e8e\u540c\u7b49\u53c2\u6570\u89c4\u6a21\u7684\u7a20\u5bc6\u6a21\u578b\uff0c\u53d6\u5f97\u4e86\u663e\u8457\u7684\u6027\u80fd\u63d0\u5347\u3002\u82f1\u7279\u5c14\u91c7\u7528\u7684\u8f6f\u4ef6\u4f18\u5316\u7b56\u7565\u6db5\u76d6\u4e86\u9488\u5bf9\u7a00\u758f\u6df7\u5408\u4e13\u5bb6\u6a21\u578b\u67b6\u6784\uff08SparseMoE\uff09\u7684\u7b97\u5b50\u878d\u5408\uff0c\u9488\u5bf93B\u6fc0\u6d3bMOE\u6a21\u578b\u5b9a\u5236\u5316\u7684\u8c03\u5ea6\u548c\u8bbf\u5b58\u4f18\u5316\u4ee5\u53ca\u9488\u5bf9\u4e0d\u540c\u4e13\u5bb6\u4e4b\u95f4\u7684\u8d1f\u8f7d\u5747\u8861,\u8fd9\u4e9b\u6280\u672f\u80fd\u591f\u52a9\u529b\u66f4\u591aMOE\u6a21\u578b\u5728\u82f1\u7279\u5c14\u5e73\u53f0\u4e0a\u5b9e\u73b0\u9ad8\u6548\u90e8\u7f72\u3002<\/p>\n

\u6b64\u6b21\u53d1\u5e03\u7684Qwen3\u7cfb\u5217\u6a21\u578b\u4e3b\u8981\u96c6\u4e2d\u4e8e\u4e2d\u5c0f\u53c2\u6570\u91cf\u7684\u7a20\u5bc6\u67b6\u6784LLM\uff0c\u53c2\u6570\u89c4\u6a21\u6db5\u76d60.6B\u81f332B\uff0c\u80fd\u591f\u9002\u914d\u66f4\u5e7f\u6cdb\u7684\u786c\u4ef6\u8d44\u6e90\uff0c\u6ee1\u8db3\u591a\u6837\u5316\u7684\u4f7f\u7528\u573a\u666f\u9700\u6c42\u3002\u82f1\u7279\u5c14\u7684CPU\u3001GPU\u3001NPU\u67b6\u6784\u5168\u9762\u9002\u914dQwen\u7cfb\u5217\u6a21\u578b\uff0c\u4e3a\u6a21\u578b\u7684\u90e8\u7f72\u8fdb\u884c\u4f18\u5316\uff0c\u4f7f\u7528\u82f1\u7279\u5c14OpenVINOTM\u5de5\u5177\u5957\u4ef6\u548cPyTorch\u793e\u533a\u5de5\u5177\u90fd\u53ef\u4ee5\u4e3a\u5168\u7cfb\u5217Qwen\u6a21\u578b\u5728\u82f1\u7279\u5c14\u9177\u777fUltra\u5e73\u53f0(\u9177\u777fUltra100\u7cfb\u5217\/200\u7cfb\u5217)\u548c\u82f1\u7279\u5c14\u9510\u70abA\u7cfb\u5217\u663e\u5361\u548cB\u7cfb\u5217\u663e\u5361\u4e0a\u7684\u90e8\u7f72\u5b9e\u73b0\u5353\u8d8a\u6027\u80fd\u8868\u73b0\u3002<\/p>\n

\u503c\u5f97\u4e00\u63d0\u7684\u662f\uff0c\u82f1\u7279\u5c14\u9996\u6b21\u5728NPU\u4e0a\u5bf9\u6a21\u578b\u53d1\u5e03\u63d0\u4f9b\u7b2c\u96f6\u65e5\uff08Day0\uff09\u652f\u6301\uff0c\u4f53\u73b0\u4e86\u82f1\u7279\u5c14\u548c\u5f00\u6e90\u751f\u6001\u66f4\u52a0\u6df1\u5165\u7684\u5408\u4f5c\uff0c\u9488\u5bf9\u4e0d\u540c\u6a21\u578b\u53c2\u6570\u91cf\u548c\u5e94\u7528\u573a\u666f\u63d0\u4f9b\u66f4\u591a\u6837\u5316\u7684\u3001\u66f4\u6709\u9488\u5bf9\u6027\u7684\u5e73\u53f0\u652f\u6301\u3002\u9488\u5bf9\u4ece0.6B\u5230\u9ad8\u8fbe8B\u7684\u4e2d\u5c0f\u53c2\u6570\u91cf\u6a21\u578b\u5168\u9762\u652f\u6301\uff0c\u541e\u5410\u91cf\u6700\u9ad8\u8fbe\u523036.68token\/s\uff0c\u501f\u52a9\u82f1\u7279\u5c14Lunar Lake NPU\u5e73\u53f0\u548c\u82f1\u7279\u5c14OpenVINO\u5de5\u5177\u5957\u4ef6\uff0c\u53ef\u4ee5\u5728\u4fdd\u6301\u4f4e\u80fd\u8017\u7684\u540c\u65f6\uff0c\u53c8\u53ef\u4ee5\u83b7\u5f97\u4f18\u5f02\u6027\u80fd\u3002<\/p>\n

\u540c\u65f6\uff0c\u5728\u9177\u777fUltra\u7684iGPU\u5e73\u53f0\u4e0a\uff0c\u82f1\u7279\u5c14\u6301\u7eed\u4e3a\u6a21\u578b\u5e26\u6765\u5353\u8d8a\u7684\u6027\u80fd\u3002\u9488\u5bf9\u5c0f\u5c3a\u5bf8\u7684\u6a21\u578b\uff0c\u5728FP16\u7cbe\u5ea6\u4e0b\uff0c\u6700\u9ad8\u8fbe\u523066token\/s\uff0c\u9488\u5bf9\u4e2d\u5c0f\u5c3a\u5bf8\u6a21\u578b\uff0c\u5728INT4\u7cbe\u5ea6\u4e0b\uff0c\u6700\u9ad8\u8fbe\u523035.83token\/s\u3002\u5f00\u53d1\u8005\u53ef\u4ee5\u6839\u636e\u9002\u5408\u7684\u4f7f\u7528\u573a\u666f\uff0c\u627e\u5230\u7cbe\u5ea6\u548c\u6027\u80fd\u7684\u6700\u4f73\u7ec4\u5408\u3002\u5728\u82f1\u7279\u5c14\u9510\u70abB\u7cfb\u5217\u663e\u5361\u66f4\u5f3a\u5927\u7b97\u529b\u52a0\u6301\u4e0b\uff0cQwen3-8B\u6a21\u578b\u53ef\u5b9e\u73b070.67token\/s \uff0c\u5f00\u53d1\u8005\u80fd\u591f\u5373\u523b\u9886\u7565\u5230\u6700\u65b0\u6a21\u578b\u4e0e\u82f1\u7279\u5c14\u5353\u8d8a\u5e73\u53f0\u80fd\u529b\u7684\u8d85\u5f3a\u7ec4\u5408\uff0c\u7545\u4eab\u524d\u6cbf\u79d1\u6280\u5e26\u6765\u7684\u9ad8\u6548\u4e0e\u4fbf\u6377\u3002<\/p>\n

\u4f5c\u4e3a\u751f\u6210\u5f0fAI\u6a21\u578b\u4e2d\u7684\u8f7b\u91cf\u7ea7\u9009\u624b\uff0c0.6B\u53c2\u6570\u91cf\u7684\u5c0f\u578bLLM\u5177\u5907\u7075\u6d3b\u9ad8\u6548\u7684\u90e8\u7f72\u4f18\u52bf\u4ee5\u53ca\u5feb\u901f\u8fed\u4ee3\u66f4\u65b0\u7684\u80fd\u529b\u3002\u7136\u800c\uff0c\u5728\u5b9e\u9645\u5e94\u7528\u843d\u5730\u8fc7\u7a0b\u4e2d\uff0c\u4eba\u4eec\u5e38\u5e38\u5bf9\u5176\u77e5\u8bc6\u7684\u6df1\u5ea6\u4e0e\u5e7f\u5ea6\u4ee5\u53ca\u5904\u7406\u590d\u6742\u4efb\u52a1\u7684\u80fd\u529b\u5b58\u5728\u987e\u8651\u3002\u901a\u8fc7\u501f\u52a9\u7279\u5b9a\u7684\u6570\u636e\u96c6\u5bf9\u8fd9\u4e9b\u5c0f\u578bLLM\u8fdb\u884c\u7aef\u4fa7\u5fae\u8c03\uff08Fine-tune\uff09\uff0c\u53ef\u4ee5\u63d0\u5347\u6a21\u578b\u667a\u80fd\uff0c\u4f18\u5316\u7528\u6237\u4f53\u9a8c\u3002\u4e3a\u6b64\uff0c\u57fa\u4e8eUnsloth\u548cHuggingFace\u53c2\u6570\u9ad8\u6548\u5fae\u8c03\u6846\u67b6\uff08Parameter-EfficientFine-Tuning, PEFT\uff09\uff0c\u82f1\u7279\u5c14\u6784\u5efa\u4e86\u4e00\u5957\u5b8c\u6574\u7684\u7aef\u4fa7\u89e3\u51b3\u65b9\u6848\uff0c\u4f7f\u6a21\u578b\u53d8\u5f97\u66f4\u52a0\u667a\u80fd\uff0cAI PC\u5e94\u7528\u4e5f\u771f\u6b63\u53d8\u6210\u7528\u6237\u7684\u8d34\u5fc3\u667a\u80fd\u5c0f\u52a9\u624b\u3002<\/p>\n

\u501f\u52a9\u8fd9\u4e2a\u65b9\u6848\uff0c\u5728Qwen3 0.6B\u6a21\u578b\u7684\u591a\u8bed\u8a00\u80fd\u529b\uff0c\u53ef\u4ee5\u66f4\u597d\u7684\u53d1\u6325\u4f5c\u7528\uff0c\u8fd9\u91cc\u5c31\u662f\u4e00\u4e2a\u7528\u591a\u8bed\u8a00\u5bf9\u56fe\u7247\u8fdb\u884c\u67e5\u8be2\u7684\u5b9e\u4f8b<\/p>\n

\u5728\u672c\u6b21\u53d1\u5e03\u7684Qwen3\u6a21\u578b\u4e2d\uff0c\u6211\u4eec\u6ce8\u610f\u5230\u5176LLM\u957f\u4e0a\u4e0b\u6587\u80fd\u529b\u5f97\u5230\u4e86\u663e\u8457\u63d0\u5347\u3002\u9762\u5bf9\u6709\u9650\u7684\u7b97\u529b\u8d44\u6e90\uff0c\u5982\u4f55\u6709\u6548\u5229\u7528\u6a21\u578b\u7684\u957f\u4e0a\u4e0b\u6587\u80fd\u529b\uff0c\u907f\u514d\u8ba1\u7b97\u8d44\u6e90\u6d88\u8017\u5448\u6307\u6570\u7ea7\u589e\u957f\uff0c\u8fdb\u800c\u62d3\u5c55LLM\u5728\u5ba2\u6237\u7aef\u7684\u5e94\u7528\u573a\u666f\uff0c\u82f1\u7279\u5c14\u5df2\u7ed9\u51fa\u89e3\u51b3\u65b9\u6848\u3002\u57fa\u4e8e\u52a8\u6001\u7a00\u758f\u6ce8\u610f\u529b\u5341\u5927\u54c1\u724c\u7f51<\/a>\uff0c\u5728\u4fdd\u8bc1\u8fd1\u4e4e\u65e0\u635f\u7cbe\u5ea6\u7684\u524d\u63d0\u4e0b\uff0c\u6211\u4eec\u53ef\u4ee5\u4f7f\u957f\u4e0a\u4e0b\u6587\u7a97\u53e3\u5904\u7406\u901f\u5ea6\u6210\u500d\u63d0\u5347\u3002\u91c7\u7528\u8be5\u65b9\u6848\u540e\uff0cQwen3-8B\u6a21\u578b\u5728\u82f1\u7279\u5c14LNL\u5e73\u53f0\u4e0a\u53ef\u5b9e\u73b032K\u7684\u4e0a\u4e0b\u6587\u957f\u5ea6\u3002<\/p>\n

\u8fd9\u6837\u7684\u957f\u4e0a\u4e0b\u6587\u80fd\u529b\uff0c\u89e3\u9501\u4e86\u66f4\u591a\u7aef\u4fa7Agent\u65b0\u5e94\u7528\u3002\u7ed3\u5408Qwen3\u66f4\u5f3a\u7684Agent\u548c\u4ee3\u7801\u80fd\u529b\uff0c\u4ee5\u53ca\u5bf9MCP\u534f\u8bae\u7684\u52a0\u5f3a\u652f\u6301\uff0c\u4f7f\u5f97\u57fa\u4e8e\u7aef\u4fa7\u5927\u6a21\u578b\u8c03\u7528MCP\u670d\u52a1\u6765\u5f00\u53d1\u5404\u79cdAI PC Agent\u9996\u6b21\u6210\u4e3a\u53ef\u80fd\u3002\u89c6\u9891\u5c55\u793a\u4e86\uff0c\u5728\u82f1\u7279\u5c14AI PC\u4e0a\uff0c\u57fa\u4e8eQwen3-8B\u6a21\u578b\u8c03\u7528\u5fc5\u4f18\u79d1\u6280MCP\u670d\u52a1\u81ea\u52a8\u751f\u6210PPT\u7684\u8fc7\u7a0b\u3002<\/p>\n

\u82f1\u7279\u5c14\u6301\u7eed\u62e5\u62b1\u5f00\u6e90\u751f\u6001\uff0c\u82f1\u7279\u5c14\u4f18\u5316\u7248Ollama\u4e5f\u7b2c\u4e00\u65f6\u95f4\u652f\u6301\u4e86\u5305\u62ecMoE\u6a21\u578b\u5728\u5185\u7684Qwen3\u7cfb\u5217\u6a21\u578b\uff0c\u4f7f\u5f00\u53d1\u8005\u53ef\u4ee5\u5728\u82f1\u7279\u5c14\u5ba2\u6237\u7aef\u5e73\u53f0(\u5982\u82f1\u7279\u5c14\u9177\u777fUltraAI PC\uff0c\u82f1\u7279\u5c14\u9510\u70abA\/B\u7cfb\u5217\u663e\u5361)\u4e0a\uff0c\u5229\u7528Ollama\u6846\u67b6\u6765\u642d\u5efa\u57fa\u4e8eQwen3\u7cfb\u5217\u6a21\u578b\u7684\u667a\u80fd\u5e94\u7528\u3002\u4f18\u5316\u7248Ollama\u8fd8\u63d0\u4f9b\u4e86\u57fa\u4e8e\u9b54\u642d\u7684\u6a21\u578b\u4e0b\u8f7d\u6e90\u8bbe\u7f6e\uff0c\u4f7f\u5f97\u672c\u5730\u5f00\u53d1\u8005\u53ef\u4ee5\u4ece\u9b54\u642d\u793e\u533a\u66f4\u9ad8\u6548\u4e0b\u8f7d\u548c\u90e8\u7f72\u5b89\u88c5\u5305\u53ca\u6a21\u578b\u3002<\/p>\n

\u82f1\u7279\u5c14\u5728\u8f66\u7aef\u8231\u5185\u5bf9\u65b0\u53d1\u5e03\u7684Qwen3\u7cfb\u5217\u6a21\u578b\uff0c\u4e5f\u5df2\u7ecf\u987a\u6ed1\u5339\u914d\u3002\u57fa\u4e8e\u82f1\u7279\u5c14\u8f66\u8f7d\u8f6f\u786c\u4ef6\u89e3\u51b3\u65b9\u6848\uff08\u542b\u7b2c\u4e00\u4ee3\u82f1\u7279\u5c14AI\u589e\u5f3a\u8f6f\u4ef6\u5b9a\u4e49\u6c7d\u8f66\uff08SDV\uff09SOC\u3001\u521a\u521a\u5728\u4e0a\u6d77\u8f66\u5c55\u53d1\u5e03\u7684\u7b2c\u4e8c\u4ee3SDVSoC NPU\uff0c\u4ee5\u53ca\u82f1\u7279\u5c14\u9510\u70ab\u2122\u8f66\u8f7d\u72ec\u7acb\u663e\u5361\uff09\uff0c\u82f1\u7279\u5c14\u80fd\u8ba9Qwen3\u7cfb\u5217\u6a21\u578b\u5f88\u5feb\u5c31\u6709\u673a\u4f1a\u4e0a\u8f66\u90e8\u7f72\uff0c\u4ece\u800c\u5145\u5206\u53d1\u6325\u8f66\u7aef\u7684\u672c\u5730\u7b97\u529b\u3002\u5176\u4e2d\uff0c\u7387\u5148\u5728\u6c7d\u8f66\u884c\u4e1a\u5185\u91c7\u7528\u4e86\u591a\u8282\u70b9\u82af\u7c92\u67b6\u6784\u7684\u7b2c\u4e8c\u4ee3SDVSOC\uff0c\u5176\u751f\u6210\u5f0f\u548c\u591a\u6a21\u6001AI\u6027\u80fd\uff0c\u76f8\u6bd4\u4e0a\u4e00\u4ee3\uff0c\u6700\u9ad8\u53ef\u63d0\u5347\u5341\u500d\uff0c\u8fd9\u8ba9\u6c7d\u8f66AI\u4f53\u9a8c\uff0c\u5982\u8231\u5185\u7684\u5b9e\u65f6\u5bf9\u8bdd\u3001\u81ea\u7136\u8bed\u8a00\u4ea4\u4e92\u548c\u590d\u6742\u6307\u4ee4\u7684\u54cd\u5e94\u7b49\uff0c\u90fd\u5145\u6ee1AI\u7684\u7075\u6027\u3002<\/p>\n

\u5feb\u901f\u4e0a\u624b\u6307\u5357(Get Started)<\/p>\n

\u4e0b\u9762\u5c06\u4ee5Qwen3-8B\u4e3a\u4f8b\uff0c\u4ecb\u7ecd\u5982\u4f55\u5229\u7528OpenVINO\u7684PythonAPI\u5728\u82f1\u7279\u5c14\u5e73\u53f0\uff08GPU,NPU\uff09\u4e0a\u8fd0\u884cQwen3\u7cfb\u5217\u6a21\u578b\u3002 \u4e5f\u53ef\u4ee5\u53c2\u8003\u4e0b\u9762\u7684\u5b8c\u6574\u793a\u4f8b\uff1a<\/p>\n

\u7b2c\u4e00\u6b65\uff0c\u73af\u5883\u51c6\u5907<\/p>\n

\u57fa\u4e8e\u4ee5\u4e0b\u547d\u4ee4\u53ef\u4ee5\u5b8c\u6210\u6a21\u578b\u90e8\u7f72\u4efb\u52a1\u5728Python\u4e0a\u7684\u73af\u5883\u5b89\u88c5\u3002<\/p>\n

python -m venv py_venv<\/p>\n

.\/py_venv\/Scripts\/activate.bat<\/p>\n

pip install –pre -U openvino-genai openvino openvino-tokenizers –extra-index-url<\/p>\n

pip install nncf<\/p>\n

pip install git+@2aebd4441023d3c003b27c87fff5312254ac<\/p>\n

pip install transformers >=4.51.3<\/p>\n

\u7b2c\u4e8c\u6b65\uff0c\u6a21\u578b\u4e0b\u8f7d\u548c\u8f6c\u6362<\/p>\n

\u5728\u90e8\u7f72\u6a21\u578b\u4e4b\u524d\uff0c\u6211\u4eec\u9996\u5148\u9700\u8981\u5c06\u539f\u59cb\u7684PyTorch\u6a21\u578b\u8f6c\u6362\u4e3aOpenVINOTM\u7684IR\u9759\u6001\u56fe\u683c\u5f0f\uff0c\u5e76\u5bf9\u5176\u8fdb\u884c\u538b\u7f29\uff0c\u4ee5\u5b9e\u73b0\u66f4\u8f7b\u91cf\u5316\u7684\u90e8\u7f72\u548c\u6700\u4f73\u7684\u6027\u80fd\u8868\u73b0\u3002\u901a\u8fc7Optimum\u63d0\u4f9b\u7684\u547d\u4ee4\u884c\u5de5\u5177optimum-cli\uff0c\u6211\u4eec\u53ef\u4ee5\u4e00\u952e\u5b8c\u6210\u6a21\u578b\u7684\u683c\u5f0f\u8f6c\u6362\u548c\u6743\u91cd\u91cf\u5316\u4efb\u52a1\u3002<\/p>\n

optimum-cli export openvino –model Qwen\/Qwen3-8B –task text-generation-with-past –weight-format int4 –group-size 128 –ratio 0.8 Qwen3-8B-int4-ov<\/p>\n

optimum-cli\u4f7f\u7528\u65b9\u6cd5\u53ef\u4ee5\u53c2\u8003\uff1a<\/p>\n

\u6b64\u5916\u6211\u4eec\u5efa\u8bae\u4f7f\u7528\u4ee5\u4e0b\u53c2\u6570\u5bf9\u8fd0\u884c\u5728NPU\u4e0a\u7684\u6a21\u578b\u8fdb\u884c\u91cf\u5316\uff0c\u4ee5\u8fbe\u5230\u6027\u80fd\u548c\u7cbe\u5ea6\u7684\u5e73\u8861\u3002<\/p>\n

optimum-cliexport openvino –model Qwen\/Qwen3-8B –tasktext-generation-with-past –weight-format nf4 –sym –group-size -1Qwen3-8B-nf4-ov –backup-precision int8_sym<\/p>\n

\u7b2c\u4e09\u6b65\uff0c\u6a21\u578b\u90e8\u7f72<\/p>\n

OpenVINOTM\u76ee\u524d\u63d0\u4f9b\u4e24\u79cd\u9488\u5bf9\u5927\u8bed\u8a00\u6a21\u578b\u7684\u90e8\u7f72\u65b9\u6848\uff0c\u5982\u679c\u60a8\u4e60\u60ef\u4e8eTransformers\u5e93\u7684\u63a5\u53e3\u6765\u90e8\u7f72\u6a21\u578b\uff0c\u5e76\u60f3\u4f53\u9a8c\u76f8\u5bf9\u66f4\u4e30\u5bcc\u7684\u529f\u80fd\uff0c\u63a8\u8350\u4f7f\u7528\u57fa\u4e8ePython\u63a5\u53e3\u7684Optimum-intel\u5de5\u5177\u6765\u8fdb\u884c\u4efb\u52a1\u642d\u5efa\u3002\u5982\u679c\u60a8\u60f3\u5c1d\u8bd5\u66f4\u6781\u81f4\u7684\u6027\u80fd\u6216\u662f\u8f7b\u91cf\u5316\u7684\u90e8\u7f72\u65b9\u5f0f\uff0cGenAIAPI\u5219\u662f\u4e0d\u4e8c\u7684\u9009\u62e9\uff0c\u5b83\u540c\u65f6\u652f\u6301Python\u548cC++\u4e24\u79cd\u7f16\u7a0b\u8bed\u8a00\uff0c\u5b89\u88c5\u5bb9\u91cf\u4e0d\u5230200MB\u3002<\/p>\n

Optimum-intel\u90e8\u7f72\u793a\u4f8b<\/p>\n

from optimum.intel.openvino import OVModelForCausalLM<\/p>\n

from transformers import AutoConfig, AutoTokenizer<\/p>\n

ov_model = OVModelForCausalLM.from_pretrained(<\/p>\n

llm_model_path,<\/p>\n

device=’GPU’,<\/p>\n

tokenizer = AutoTokenizer.from_pretrained(llm_model_path)<\/p>\n

prompt = “Give me a short introduction to large language model.”<\/p>\n

messages = [{“role”: “user”, “content”: prompt}]<\/p>\n

text = tokenizer.apply_chat_template(<\/p>\n

messages,<\/p>\n

tokenize=False,<\/p>\n

add_generation_prompt=True,<\/p>\n

enable_thinking=True<\/p>\n

model_inputs = tokenizer([text], return_tensors=”pt”)<\/p>\n

generated_ids = ov_model.generate(**model_inputs, max_new_tokens=1024)<\/p>\n

output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist()<\/p>\n

try:<\/p>\n

index = len(output_ids) – output_ids[::-1].index(151668)<\/p>\n

except ValueError:<\/p>\n

index = 0<\/p>\n

thinking_content = tokenizer.decode(output_ids[:index], skip_special_tokens=True).strip(“n”)<\/p>\n

content = tokenizer.decode(output_ids[index:], skip_special_tokens=True).strip(“n”)<\/p>\n

print(“thinking content:”, thinking_content)<\/p>\n

print(“content:”, content)<\/p>\n

GenAI API\u90e8\u7f72\u793a\u4f8b<\/p>\n

import openvino_genai as ov_genai<\/p>\n

generation_config = ov_genai.GenerationConfig()<\/p>\n

generation_config.max_new_tokens = 128<\/p>\n

generation_config.apply_chat_template = False<\/p>\n

pipe = ov_genai.LLMPipeline(llm_model_path, “GPU”)<\/p>\n

result = pipe.generate(prompt, generation_config)<\/p>\n

\u8fd9\u91cc\u53ef\u4ee5\u4fee\u6539devicename\u7684\u65b9\u5f0f\u5c06\u6a21\u578b\u8f7b\u677e\u90e8\u7f72\u5230NPU\u4e0a\u3002<\/p>\n

pipe=ov_genai.LLMPipeline(llm_model_path,”NPU”)<\/p>\n

\u6b64\u5916Ollama package on Modelscope\uff08\uff09 is ready to download now<\/p>\n

———————————————————————<\/p>\n

\u6027\u80fd\u56e0\u4f7f\u7528\u60c5\u51b5\u3001\u914d\u7f6e\u548c\u5176\u4ed6\u56e0\u7d20\u800c\u5f02\u3002 \u6b32\u4e86\u89e3\u66f4\u591a\u4fe1\u606f\uff0c\u8bf7\u8bbf\u95ee\/performanceindex\u3002\u57fa\u4e8e\u5bf9\u7b2c\u4e8c\u4ee3\u82f1\u7279\u5c14AI\u589e\u5f3aSDV SoC GPU+NPU\u4e0eMBL i7-13800HAQ CPU+GPU\uff08\u5173\u95ed\u777f\u9891\uff09\u7684\u5185\u90e8\u9884\u6d4b\uff0cAI\u6027\u80fd\u6700\u9ad8\u53ef\u63d0\u5347\u5341\u500d\u3002<\/p>\n

\u672c\u6587\u5c5e\u4e8e\u539f\u521b\u6587\u7ae0\uff0c\u5982\u82e5\u8f6c\u8f7d\uff0c\u8bf7\u6ce8\u660e\u6765\u6e90\uff1a\u82f1\u7279\u5c14\u7b2c\u4e00\u65f6\u95f4\u6df1\u5ea6\u4f18\u5316Qwen3\u5927\u6a21\u578b\uff0c\u5347\u7ea7AI PC\u80fd\u529b\u8d4b\u80fd\u591a\u6837\u5316\u573a\u666f<\/p>\n","protected":false},"excerpt":{"rendered":"

\u503c\u5f97\u4e00\u63d0\u7684\u662f\uff0c\u82f1\u7279\u5c14\u9996\u6b21\u5728NPU\u4e0a\u5bf9\u6a21\u578b\u53d1\u5e03\u63d0\u4f9b\u7b2c\u96f6\u65e5\uff08Day0\uff09\u652f\u6301\uff0c\u4f53\u73b0\u4e86\u82f1\u7279\u5c14\u548c\u5f00\u6e90\u751f\u6001\u66f4\u52a0\u6df1\u5165\u7684\u5408\u4f5c\uff0c\u9488\u5bf9\u4e0d\u540c\u6a21\u578b\u53c2\u6570\u91cf\u548c\u5e94\u7528\u573a\u666f\u63d0\u4f9b\u66f4\u591a\u6837\u5316\u7684\u3001\u66f4\u6709\u9488\u5bf9\u6027\u7684\u5e73\u53f0\u652f\u6301\u3002<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[7],"tags":[],"class_list":["post-14891","post","type-post","status-publish","format-standard","hentry","category-fs"],"_links":{"self":[{"href":"https:\/\/www.ztuzx.com\/index.php?rest_route=\/wp\/v2\/posts\/14891","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.ztuzx.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.ztuzx.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.ztuzx.com\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.ztuzx.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=14891"}],"version-history":[{"count":0,"href":"https:\/\/www.ztuzx.com\/index.php?rest_route=\/wp\/v2\/posts\/14891\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.ztuzx.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=14891"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.ztuzx.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=14891"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.ztuzx.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=14891"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}