{"id":818,"date":"2026-07-20T17:04:02","date_gmt":"2026-07-20T09:04:02","guid":{"rendered":"https:\/\/blog.heartwarming.online\/?p=818"},"modified":"2026-07-20T17:04:04","modified_gmt":"2026-07-20T09:04:04","slug":"%e4%bb%8e%e4%ba%91%e7%ab%af%e8%af%ad%e9%9f%b3%e5%8a%a9%e6%89%8b%e5%88%b0%e6%9c%ac%e5%9c%b0%e5%8f%af%e6%9b%bf%e6%8d%a2%e6%a8%a1%e5%9e%8b%e6%9c%8d%e5%8a%a1%ef%bc%9aqwen3-asr-vllm-%e9%83%a8%e7%bd%b2","status":"publish","type":"post","link":"https:\/\/blog.heartwarming.online\/index.php\/2026\/07\/20\/%e4%bb%8e%e4%ba%91%e7%ab%af%e8%af%ad%e9%9f%b3%e5%8a%a9%e6%89%8b%e5%88%b0%e6%9c%ac%e5%9c%b0%e5%8f%af%e6%9b%bf%e6%8d%a2%e6%a8%a1%e5%9e%8b%e6%9c%8d%e5%8a%a1%ef%bc%9aqwen3-asr-vllm-%e9%83%a8%e7%bd%b2\/","title":{"rendered":"\u4ece\u4e91\u7aef\u8bed\u97f3\u52a9\u624b\u5230\u672c\u5730\u53ef\u66ff\u6362\u6a21\u578b\u670d\u52a1\uff1aQwen3-ASR + vLLM \u90e8\u7f72\u5b9e\u6218\u8bb0\u5f55"},"content":{"rendered":"\n
\n

\u65e5\u671f\uff1a2026-07-20
\u76ee\u6807\uff1a\u5728\u5b66\u6821 GPU \u670d\u52a1\u5668\u4e0a\u5b8c\u6210\u672c\u5730 ASR \u4e0e\u5c0f\u53c2\u6570 LLM \u7684\u90e8\u7f72\u9a8c\u8bc1\uff0c\u5e76\u8ba9 Windows \u4e1a\u52a1\u540e\u7aef\u901a\u8fc7\u53ef\u914d\u7f6e URL \u8c03\u7528\u3002\u540e\u7eed\u5230\u7532\u65b9\u670d\u52a1\u5668\u65f6\uff0c\u53ea\u66ff\u6362\u6a21\u578b\u670d\u52a1\u5730\u5740\u3001Token \u548c\u5fc5\u8981\u7684\u6a21\u578b\u522b\u540d\uff0c\u4e0d\u518d\u5927\u6539\u4e1a\u52a1\u4ee3\u7801\u3002<\/p>\n<\/blockquote>\n\n\n\n

\n\n\n\n

\u4e00\u3001\u9879\u76ee\u80cc\u666f<\/h2>\n\n\n\n

\u73b0\u6709\u7cfb\u7edf\u662f\u4e00\u5957 AI \u8bed\u97f3\u52a9\u624b\uff0c\u4e3b\u8981\u6d41\u7a0b\u5982\u4e0b\uff1a<\/p>\n\n\n\n

\u6d4f\u89c8\u5668\u5f55\u97f3\n\u2192 Windows FastAPI \u540e\u7aef\n\u2192 \u8bed\u97f3\u8bc6\u522b ASR\n\u2192 LLM \u5224\u65ad\u610f\u56fe\n\u2192 \u8c03\u7528\u4e1a\u52a1 API\/\u5de5\u5177\n\u2192 LLM \u603b\u7ed3\u7ed3\u679c\n\u2192 \u524d\u7aef\u663e\u793a\u56de\u7b54<\/code><\/pre>\n\n\n\n
\u539f\u59cb\u5b9e\u73b0\u4e2d\uff1a<\/p>\n\n\n\n
    \n
  • \u5b9e\u65f6\u8bed\u97f3\u8bc6\u522b\u4f7f\u7528\u963f\u91cc\u4e91 DashScope Fun-ASR\uff1b<\/li>\n\n\n\n
  • \u4e2d\u95f4 LLM \u4f7f\u7528\u4e91\u7aef OpenAI \u517c\u5bb9 API\uff1b<\/li>\n\n\n\n
  • Windows FastAPI \u540e\u7aef\u6838\u5fc3\u63a5\u53e3\u4e3a\uff1a<\/li>\n\n\n\n
  • WebSocket \/ws\/asr<\/code><\/li>\n\n\n\n
  • POST \/api\/chat\/query<\/code><\/li>\n\n\n\n
  • GET \/health<\/code><\/li>\n\n\n\n
  • \u4e3a\u4e86\u65b9\u4fbf\u7532\u65b9\u79c1\u6709\u5316\u90e8\u7f72\uff0c\u5e0c\u671b\u6539\u6210\uff1a<\/li>\n\n\n\n
  • ASR \u53ef\u66ff\u6362\uff1b<\/li>\n\n\n\n
  • LLM \u53ef\u66ff\u6362\uff1b<\/li>\n\n\n\n
  • \u4e1a\u52a1\u540e\u7aef\u53ea\u8bfb\u53d6 URL\u3001Token \u548c\u6a21\u578b\u540d\uff1b<\/li>\n\n\n\n
  • \u5b66\u6821\u6d4b\u8bd5\u73af\u5883\u4e0e\u7532\u65b9\u751f\u4ea7\u73af\u5883\u4f7f\u7528\u540c\u4e00\u5957\u534f\u8bae\u3002<\/li>\n<\/ul>\n\n\n\n
    \n\n\n\n
    \u4e8c\u3001\u6700\u7ec8\u76ee\u6807\u67b6\u6784<\/h2>\n\n\n\n
    \u672c\u6b21\u786e\u5b9a\u7684\u76ee\u6807\u67b6\u6784\u4e3a\uff1a<\/p>\n\n\n\n
    \u6d4f\u89c8\u5668\n\u2502\n\u251c\u2500 WebSocket \/ws\/asr\n\u2502      \u2193\n\u2502  Windows voice_api.py\n\u2502      \u2193\n\u2502  ASR_REMOTE_WS_URL\n\u2502      \u2193\n\u2502  Qwen3-ASR \u670d\u52a1\n\u2502\n\u2514\u2500 POST \/api\/chat\/query\n       \u2193\n   Windows voice_api.py\n       \u2193\n   LLM_BASE_URL\n       \u2193\n   OpenAI \u517c\u5bb9 LLM \u670d\u52a1\n       \u2193\n   \u5de5\u5177\u8c03\u7528 \/ API \u67e5\u8be2\n       \u2193\n   \u603b\u7ed3\u56de\u7b54<\/code><\/pre>\n\n\n\n
    \u914d\u7f6e\u7edf\u4e00\u653e\u5728 .env<\/code>\uff1a<\/p>\n\n\n\n
    ASR_REMOTE_WS_URL=ws:\/\/127.0.0.1:18095\/v1\/asr\/stream\nASR_API_TOKEN=<ASR_TOKEN>\n\nLLM_BASE_URL=http:\/\/127.0.0.1:18096\/v1\nLLM_API_KEY=<LLM_TOKEN>\nLLM_MODEL=business-llm\nLLM_TIMEOUT=60<\/code><\/pre>\n\n\n\n
    \u4ee5\u540e\u7532\u65b9\u90e8\u7f72\u65f6\uff0c\u53ea\u9700\u8981\u6539\uff1a<\/p>\n\n\n\n
    ASR_REMOTE_WS_URL=ws:\/\/<CUSTOMER_ASR_HOST>:10095\/v1\/asr\/stream\nLLM_BASE_URL=http:\/\/<CUSTOMER_LLM_HOST>:10096\/v1\nASR_API_TOKEN=<CUSTOMER_ASR_TOKEN>\nLLM_API_KEY=<CUSTOMER_LLM_TOKEN><\/code><\/pre>\n\n\n\n
    \u5982\u679c\u6240\u6709\u73af\u5883\u7edf\u4e00\u66b4\u9732\u6a21\u578b\u540d business-llm<\/code>\uff0c\u5219 LLM_MODEL<\/code> \u4e5f\u4e0d\u9700\u8981\u4fee\u6539\u3002<\/p>\n\n\n\n
    \n\n\n\n
    \u4e09\u3001\u670d\u52a1\u5668\u78c1\u76d8\u95ee\u9898\u4e0e\u76ee\u5f55\u89c4\u5212<\/h2>\n\n\n\n
    \u5b66\u6821\u670d\u52a1\u5668\u6839\u5206\u533a\u5df2\u7ecf\u63a5\u8fd1\u6ee1\u8f7d\uff0c\u6570\u636e\u76d8 \/app\/hd\/data01<\/code> \u4ecd\u6709\u7ea6 11 TB \u7a7a\u95f4\u3002\u56e0\u6b64\u6240\u6709\u4e2a\u4eba\u73af\u5883\u3001\u6a21\u578b\u3001\u7f13\u5b58\u548c\u9879\u76ee\u7edf\u4e00\u8fc1\u79fb\u5230\uff1a<\/p>\n\n\n\n
    \/app\/hd\/data01\/username<\/code><\/pre>\n\n\n\n
    \u76ee\u5f55\u89c4\u5212\uff1a<\/p>\n\n\n\n
    \/app\/hd\/data01\/username\/\n\u251c\u2500\u2500 conda_envs\/\n\u2502   \u2514\u2500\u2500 qwen3-asr\/\n\u251c\u2500\u2500 conda_pkgs\/\n\u251c\u2500\u2500 models\/\n\u2502   \u251c\u2500\u2500 Qwen3-ASR-0.6B\/\n\u2502   \u2514\u2500\u2500 Qwen2.5-1.5B-Instruct-GPTQ-Int4\/\n\u251c\u2500\u2500 projects\/\n\u2502   \u251c\u2500\u2500 qwen-asr-service\/\n\u2502   \u2514\u2500\u2500 qwen-llm-service\/\n\u251c\u2500\u2500 huggingface\/\n\u251c\u2500\u2500 modelscope\/\n\u251c\u2500\u2500 pip_cache\/\n\u251c\u2500\u2500 tmp\/\n\u2514\u2500\u2500 torch\/<\/code><\/pre>\n\n\n\n
    \u5199\u5165 ~\/.bashrc<\/code>\uff1a<\/p>\n\n\n\n
    export AI_DATA_ROOT=\"\/app\/hd\/data01\/username\"\nexport TMPDIR=\"$AI_DATA_ROOT\/tmp\"\nexport PIP_CACHE_DIR=\"$AI_DATA_ROOT\/pip_cache\"\nexport HF_HOME=\"$AI_DATA_ROOT\/huggingface\"\nexport HUGGINGFACE_HUB_CACHE=\"$AI_DATA_ROOT\/huggingface\/hub\"\nexport TRANSFORMERS_CACHE=\"$AI_DATA_ROOT\/huggingface\/transformers\"\nexport MODELSCOPE_CACHE=\"$AI_DATA_ROOT\/modelscope\"\nexport TORCH_HOME=\"$AI_DATA_ROOT\/torch\"<\/code><\/pre>\n\n\n\n
    \u6267\u884c\uff1a<\/p>\n\n\n\n
    source ~\/.bashrc<\/code><\/pre>\n\n\n\n
    \u6ce8\u610f\u4e8b\u9879<\/h3>\n\n\n\n
      \n
    • \u4e0d\u8981\u5728\u6839\u5206\u533a\u7ee7\u7eed\u5b89\u88c5\u5927\u578b\u6a21\u578b\u6216 Python \u73af\u5883\uff1b<\/li>\n\n\n\n
    • \u4e0d\u8981\u6267\u884c sudo pip install<\/code>\uff1b<\/li>\n\n\n\n
    • \u4e0d\u8981\u5bf9 \/app\/hd\/data01<\/code> \u6574\u76d8\u9012\u5f52 chown<\/code>\uff1b<\/li>\n\n\n\n
    • \u53ea\u5904\u7406\u4e2a\u4eba\u76ee\u5f55\uff1a<\/li>\n<\/ul>\n\n\n\n
      sudo chown -R username:username \/app\/hd\/data01\/username<\/code><\/pre>\n\n\n\n
      \n\n\n\n
      \u56db\u3001VS Code Remote-SSH \u9879\u76ee\u76ee\u5f55\u5207\u6362<\/h2>\n\n\n\n
      \u9879\u76ee\u76ee\u5f55\u521b\u5efa\u5728\u6570\u636e\u76d8\uff1a<\/p>\n\n\n\n
      mkdir -p \/app\/hd\/data01\/username\/projects\/qwen-asr-service<\/code><\/pre>\n\n\n\n
      VS Code \u4e2d\u6267\u884c\uff1a<\/p>\n\n\n\n
      File \u2192 Open Folder...<\/code><\/pre>\n\n\n\n
      \u9009\u62e9\uff1a<\/p>\n\n\n\n
      \/app\/hd\/data01\/username\/projects\/qwen-asr-service<\/code><\/pre>\n\n\n\n
      \u6216\u8005\u5728\u8fdc\u7a0b\u7ec8\u7aef\u6267\u884c\uff1a<\/p>\n\n\n\n
      code -r \/app\/hd\/data01\/username\/projects\/qwen-asr-service<\/code><\/pre>\n\n\n\n
      \u5efa\u8bae\u7ed3\u6784\uff1a<\/p>\n\n\n\n
      qwen-asr-service\/\n\u251c\u2500\u2500 asr_service.py\n\u251c\u2500\u2500 requirements.txt\n\u251c\u2500\u2500 .env\n\u251c\u2500\u2500 .env.example\n\u251c\u2500\u2500 .gitignore\n\u251c\u2500\u2500 logs\/\n\u2514\u2500\u2500 scripts\/\n    \u251c\u2500\u2500 start_asr.sh\n    \u2514\u2500\u2500 test_ws.py<\/code><\/pre>\n\n\n\n
      \n\n\n\n
      \u4e94\u3001Qwen3-ASR \u90e8\u7f72<\/h2>\n\n\n\n
      1. \u73af\u5883\u4e0e\u6a21\u578b\u76ee\u5f55<\/h3>\n\n\n\n
      Conda \u73af\u5883\uff1a<\/p>\n\n\n\n
      \/app\/hd\/data01\/username\/conda_envs\/qwen3-asr<\/code><\/pre>\n\n\n\n
      \u6a21\u578b\u76ee\u5f55\uff1a<\/p>\n\n\n\n
      \/app\/hd\/data01\/username\/models\/Qwen3-ASR-0.6B<\/code><\/pre>\n\n\n\n
      \u6fc0\u6d3b\u73af\u5883\uff1a<\/p>\n\n\n\n
      source ~\/miniconda3\/etc\/profile.d\/conda.sh\nconda activate \/app\/hd\/data01\/username\/conda_envs\/qwen3-asr<\/code><\/pre>\n\n\n\n
      2. GPU \u9009\u62e9<\/h3>\n\n\n\n
      \u4f7f\u7528\u524d\u68c0\u67e5\uff1a<\/p>\n\n\n\n
      nvidia-smi\nnvidia-smi pmon -i <GPU_ID> -c 1<\/code><\/pre>\n\n\n\n
      \u786e\u8ba4\u8fdb\u7a0b\u6240\u5c5e\u7528\u6237\uff1a<\/p>\n\n\n\n
      ps -o user,pid,etime,cmd -p <PID><\/code><\/pre>\n\n\n\n
      \u4e0d\u8981\u7ed3\u675f\u5176\u4ed6\u7528\u6237\u7684\u4efb\u52a1\u3002\u672c\u6b21 ASR \u6700\u7ec8\u4f7f\u7528\u7269\u7406 GPU 2\u3002<\/p>\n\n\n\n
      \u6ce8\u610f\uff1a<\/p>\n\n\n\n
      CUDA_VISIBLE_DEVICES=2<\/code><\/pre>\n\n\n\n
      \u540e\uff0c\u7a0b\u5e8f\u65e5\u5fd7\u91cc\u663e\u793a cuda:0<\/code> \u662f\u6b63\u5e38\u7684\uff1a<\/p>\n\n\n\n
      \u7269\u7406 GPU 2 \u2192 \u8fdb\u7a0b\u5185\u903b\u8f91 cuda:0<\/code><\/pre>\n\n\n\n
      3. ASR \u670d\u52a1\u914d\u7f6e<\/h3>\n\n\n\n
      .env<\/code> \u793a\u4f8b\uff1a<\/p>\n\n\n\n
      ASR_MODEL_PATH=\/app\/hd\/data01\/username\/models\/Qwen3-ASR-0.6B\nASR_GPU_MEMORY_UTILIZATION=0.60\nASR_CHUNK_SIZE_SEC=1.0\n\nASR_HOST=127.0.0.1\nASR_PORT=10095\nCUDA_VISIBLE_DEVICES=2\n\nASR_API_TOKEN=<RANDOM_TOKEN>\nPYTHONUNBUFFERED=1<\/code><\/pre>\n\n\n\n
      Token \u751f\u6210\uff1a<\/p>\n\n\n\n
      openssl rand -hex 32<\/code><\/pre>\n\n\n\n
      \u8bbe\u7f6e\u6743\u9650\uff1a<\/p>\n\n\n\n
      chmod 600 .env<\/code><\/pre>\n\n\n\n
      4. \u542f\u52a8\u811a\u672c<\/h3>\n\n\n\n
      scripts\/start_asr.sh<\/code>\uff1a<\/p>\n\n\n\n
      #!\/usr\/bin\/env bash\n\nset -euo pipefail\n\nPROJECT_ROOT=\"\/app\/hd\/data01\/username\/projects\/qwen-asr-service\"\nCONDA_ENV=\"\/app\/hd\/data01\/username\/conda_envs\/qwen3-asr\"\n\ncd \"$PROJECT_ROOT\"\nsource \"\/home\/username\/miniconda3\/etc\/profile.d\/conda.sh\"\nconda activate \"$CONDA_ENV\"\n\nif [[ ! -f \"$PROJECT_ROOT\/.env\" ]]; then\n    echo \"\u9519\u8bef\uff1a\u6ca1\u6709\u627e\u5230 $PROJECT_ROOT\/.env\"\n    exit 1\nfi\n\nset -a\nsource \"$PROJECT_ROOT\/.env\"\nset +a\n\nmkdir -p \"$PROJECT_ROOT\/logs\"\n\nexec env \\\n    CUDA_VISIBLE_DEVICES=\"$CUDA_VISIBLE_DEVICES\" \\\n    PYTHONUNBUFFERED=\"${PYTHONUNBUFFERED:-1}\" \\\n    python -m uvicorn asr_service:app \\\n        --host \"$ASR_HOST\" \\\n        --port \"$ASR_PORT\" \\\n        --workers 1<\/code><\/pre>\n\n\n\n
      \u8d4b\u6743\u5e76\u542f\u52a8\uff1a<\/p>\n\n\n\n
      chmod +x scripts\/start_asr.sh\n.\/scripts\/start_asr.sh<\/code><\/pre>\n\n\n\n
      5. \u5065\u5eb7\u68c0\u67e5<\/h3>\n\n\n\n
      curl -sS http:\/\/127.0.0.1:10095\/health | python -m json.tool<\/code><\/pre>\n\n\n\n
      \u6210\u529f\u7ed3\u679c\uff1a<\/p>\n\n\n\n
      {\n  \"success\": true,\n  \"service\": \"qwen3-asr\",\n  \"version\": \"1.0.0\",\n  \"model_loaded\": true,\n  \"stream_endpoint\": \"\/v1\/asr\/stream\"\n}<\/code><\/pre>\n\n\n\n
      \n\n\n\n
      \u516d\u3001Windows \u901a\u8fc7 SSH \u96a7\u9053\u8bbf\u95ee ASR<\/h2>\n\n\n\n
      \u8fdc\u7a0b\u670d\u52a1\u76d1\u542c\uff1a<\/p>\n\n\n\n
      127.0.0.1:10095<\/code><\/pre>\n\n\n\n
      Windows \u672c\u5730\u6620\u5c04\u5230\uff1a<\/p>\n\n\n\n
      127.0.0.1:18095<\/code><\/pre>\n\n\n\n
      \u547d\u4ee4\uff1a<\/p>\n\n\n\n
      ssh -N `\n  -o ExitOnForwardFailure=yes `\n  -o ServerAliveInterval=30 `\n  -o ServerAliveCountMax=3 `\n  -L 127.0.0.1:18095:127.0.0.1:10095 `\n  username@<GPU_SERVER_IP><\/code><\/pre>\n\n\n\n
      \u6d4b\u8bd5\uff1a<\/p>\n\n\n\n
      Test-NetConnection 127.0.0.1 -Port 18095\ncurl.exe http:\/\/127.0.0.1:18095\/health<\/code><\/pre>\n\n\n\n
      \u9047\u5230\u7684\u95ee\u9898\uff1aCould not resolve hostname app01<\/code><\/h3>\n\n\n\n
      \u539f\u56e0\uff1a<\/p>\n\n\n\n
        \n
      • app01<\/code> \u662f\u670d\u52a1\u5668\u5185\u90e8\u4e3b\u673a\u540d\uff1b<\/li>\n\n\n\n
      • Windows DNS \u65e0\u6cd5\u89e3\u6790\uff1b<\/li>\n\n\n\n
      • \u4e0e\u9879\u76ee\u76ee\u5f55\u65e0\u5173\u3002<\/li>\n<\/ul>\n\n\n\n
        \u89e3\u51b3\uff1a<\/p>\n\n\n\n
          \n
        • \u4f7f\u7528\u771f\u5b9e IP\uff1b<\/li>\n\n\n\n
        • \u6216\u4f7f\u7528 Windows SSH \u914d\u7f6e\u4e2d\u7684\u771f\u5b9e Host<\/code> \u522b\u540d\uff1b<\/li>\n\n\n\n
        • \u4e0d\u8981\u628a\u793a\u4f8b\u540d school-gpu<\/code> \u5f53\u6210\u771f\u5b9e\u5730\u5740\u3002<\/li>\n<\/ul>\n\n\n\n
          \n\n\n\n
          \u4e03\u3001Windows ASR \u6d4b\u8bd5<\/h2>\n\n\n\n
          \u73af\u5883\u53d8\u91cf\uff1a<\/p>\n\n\n\n
          $env:TEST_ASR_URL = \"ws:\/\/127.0.0.1:18095\/v1\/asr\/stream\"\n$env:ASR_API_TOKEN = \"<ASR_TOKEN>\"<\/code><\/pre>\n\n\n\n
          \u6d4b\u8bd5 WAV\uff1a<\/p>\n\n\n\n
          python .\\test_remote_asr.py .\\audio\\1.wav<\/code><\/pre>\n\n\n\n
          \u6210\u529f\u8f93\u51fa\uff1a<\/p>\n\n\n\n
          READY: {\"type\":\"session.ready\", ...}\ntranscript.partial : ...\ntranscript.completed : ...<\/code><\/pre>\n\n\n\n
          \u5b8c\u6574\u94fe\u8def\uff1a<\/p>\n\n\n\n
          Windows WAV\n\u2192 SSH \u96a7\u9053\n\u2192 Qwen3-ASR\n\u2192 partial\n\u2192 completed<\/code><\/pre>\n\n\n\n
          \u9047\u5230\u7684\u95ee\u9898\uff1a\u8f93\u51fa\u7e41\u4f53\u5b57<\/h3>\n\n\n\n
          \u793a\u4f8b\uff1a<\/p>\n\n\n\n
          \u67e5\u8a62\u4e8c\u96f6\u4e8c\u516d\u5e74\u4e03\u6708\u5341\u516d\u65e5\u7684\u751f\u5316\u9032\u6c34\u6c28\u6c2e<\/code><\/pre>\n\n\n\n
          \u5904\u7406\u601d\u8def\uff1a<\/p>\n\n\n\n
            \n
          1. \u63d0\u793a\u8bcd\u8981\u6c42\u8f93\u51fa\u7b80\u4f53\u4e2d\u6587\uff1b<\/li>\n\n\n\n
          2. \u670d\u52a1\u7aef\u4f7f\u7528 OpenCC \u8fdb\u884c\u7e41\u8f6c\u7b80\uff1b<\/li>\n\n\n\n
          3. \u672f\u8bed\u8bc6\u522b\u9519\u8bef\u4e0e\u7e41\u7b80\u95ee\u9898\u5206\u5f00\u5904\u7406\u3002<\/li>\n<\/ol>\n\n\n\n
            \u4f8b\u5982\uff1a<\/p>\n\n\n\n
            \u6de8\u6c34 \u2192 \u51c0\u6c34<\/code><\/pre>\n\n\n\n
            \u5c5e\u4e8e\u7e41\u7b80\u8f6c\u6362\uff1b<\/p>\n\n\n\n
            \u51c0\u6c34 \u2192 \u8fdb\u6c34<\/code><\/pre>\n\n\n\n
            \u5c5e\u4e8e ASR \u8bc6\u522b\u9519\u8bef\uff0c\u9700\u8981\u4e0a\u4e0b\u6587\u8bcd\u8868\u6216\u66f4\u597d\u7684\u6a21\u578b\u5904\u7406\u3002<\/p>\n\n\n\n
            \n\n\n\n
            \u516b\u3001Windows voice_api.py<\/code> \u6539\u9020\u6210\u8fdc\u7a0b ASR \u4ee3\u7406<\/h2>\n\n\n\n
            \u539f\u59cb \/ws\/asr<\/code>\uff1a<\/p>\n\n\n\n
            \u6d4f\u89c8\u5668 PCM\n\u2192 DashScope Recognition\n\u2192 \u963f\u91cc\u4e91 Fun-ASR<\/code><\/pre>\n\n\n\n
            \u4fee\u6539\u540e\uff1a<\/p>\n\n\n\n
            \u6d4f\u89c8\u5668 PCM\n\u2192 Windows \/ws\/asr\n\u2192 ASR_REMOTE_WS_URL\n\u2192 \u8fdc\u7a0b Qwen3-ASR<\/code><\/pre>\n\n\n\n
            \u914d\u7f6e\uff1a<\/p>\n\n\n\n
            ASR_REMOTE_WS_URL=ws:\/\/127.0.0.1:18095\/v1\/asr\/stream\nASR_API_TOKEN=<ASR_TOKEN><\/code><\/pre>\n\n\n\n
            WebSocket \u53c2\u6570\uff1a<\/p>\n\n\n\n
            ASR_REMOTE_OPEN_TIMEOUT = 10\nASR_REMOTE_PING_INTERVAL = 20\nASR_REMOTE_PING_TIMEOUT = 20\nASR_REMOTE_CLOSE_TIMEOUT = 5<\/code><\/pre>\n\n\n\n
            \u5355\u4f4d\u5747\u4e3a\u79d2\uff1a<\/p>\n\n\n\n
              \n
            • \u8fde\u63a5\u5efa\u7acb\u8d85\u65f6 10 \u79d2\uff1b<\/li>\n\n\n\n
            • \u6bcf 20 \u79d2\u53d1\u9001\u5fc3\u8df3\uff1b<\/li>\n\n\n\n
            • \u5fc3\u8df3\u54cd\u5e94\u7b49\u5f85 20 \u79d2\uff1b<\/li>\n\n\n\n
            • \u5173\u95ed\u63e1\u624b\u7b49\u5f85 5 \u79d2\u3002<\/li>\n<\/ul>\n\n\n\n
              \u8fd9\u4e9b\u4e0d\u662f\u4e00\u6b21\u8bed\u97f3\u8bc6\u522b\u7684\u603b\u65f6\u957f\u9650\u5236\u3002<\/p>\n\n\n\n
              \n\n\n\n
              \u4e5d\u3001\u4e1a\u52a1\u6838\u5fc3\u63a5\u53e3<\/h2>\n\n\n\n
              \u771f\u6b63\u9700\u8981\u4fdd\u7559\uff1a<\/p>\n\n\n\n
              GET       \/health\nWebSocket \/ws\/asr\nPOST      \/api\/chat\/query<\/code><\/pre>\n\n\n\n
              \/ws\/asr<\/code><\/h3>\n\n\n\n
              \u63a5\u6536\uff1a<\/p>\n\n\n\n
              session.start\nPCM16 \u4e8c\u8fdb\u5236\u97f3\u9891\nsession.stop<\/code><\/pre>\n\n\n\n
              \u8fd4\u56de\uff1a<\/p>\n\n\n\n
              session.ready\ntranscript.partial\ntranscript.completed\nerror<\/code><\/pre>\n\n\n\n
              \/api\/chat\/query<\/code><\/h3>\n\n\n\n
              \u8d1f\u8d23\uff1a<\/p>\n\n\n\n
              \u8bc6\u522b\u6587\u5b57\n+ \u9875\u9762\u4e0a\u4e0b\u6587\n+ \u5bf9\u8bdd\u5386\u53f2\n\u2192 Agent\n\u2192 \u5de5\u5177\u8c03\u7528\n\u2192 \u603b\u7ed3\u7ed3\u679c\n\u2192 answer_text<\/code><\/pre>\n\n\n\n
              \u65e7\u63a5\u53e3\u53ef\u5728\u5b8c\u6574\u6d4b\u8bd5\u7a33\u5b9a\u540e\u9010\u6b65\u5220\u9664\uff1a<\/p>\n\n\n\n
              \/api\/asr\/transcribe\n\/api\/tts\/synthesize\n\/api\/voice\/query\n\/audio<\/code><\/pre>\n\n\n\n
              \n\n\n\n
              \u5341\u3001\u5c0f\u53c2\u6570 LLM \u9009\u578b\u4e0e\u90e8\u7f72<\/h2>\n\n\n\n
              \u7531\u4e8e\u5b66\u6821\u670d\u52a1\u5668\u4e0b\u8f7d\u9700\u8981\u6821\u56ed\u7f51\u6d41\u91cf\uff0c\u6700\u7ec8\u9009\u62e9\uff1a<\/p>\n\n\n\n
              Qwen\/Qwen2.5-1.5B-Instruct-GPTQ-Int4<\/code><\/pre>\n\n\n\n
              \u6a21\u578b\u76ee\u5f55\u7ea6 1.1G\uff0c\u6743\u91cd\u7ea6 1.07 GiB\uff1a<\/p>\n\n\n\n
              \/app\/hd\/data01\/username\/models\/Qwen2.5-1.5B-Instruct-GPTQ-Int4<\/code><\/pre>\n\n\n\n
              \u9009\u578b\u5b9a\u4f4d<\/h3>\n\n\n\n
                \n
              • \u4e0b\u8f7d\u91cf\u4f4e\uff1b<\/li>\n\n\n\n
              • \u53ef\u590d\u7528\u73b0\u6709 vLLM \u73af\u5883\uff1b<\/li>\n\n\n\n
              • \u9002\u5408\u4f5c\u4e3a\u94fe\u8def\u6d4b\u8bd5\u6a21\u578b\uff1b<\/li>\n\n\n\n
              • \u4e0d\u9002\u5408\u4f5c\u4e3a\u6700\u7ec8\u751f\u4ea7\u6a21\u578b\u3002<\/li>\n<\/ul>\n\n\n\n
                \u4e0d\u9700\u8981\u91cd\u65b0\u5b89\u88c5\u5341\u51e0 GB \u6846\u67b6<\/h3>\n\n\n\n
                \u76f4\u63a5\u590d\u7528\uff1a<\/p>\n\n\n\n
                \/app\/hd\/data01\/username\/conda_envs\/qwen3-asr<\/code><\/pre>\n\n\n\n
                \u8be5\u73af\u5883\u5df2\u6709\uff1a<\/p>\n\n\n\n
                PyTorch\nCUDA \u4f9d\u8d56\nvLLM\nTransformers<\/code><\/pre>\n\n\n\n
                \u56e0\u6b64\u53ea\u65b0\u589e\u6a21\u578b\u6743\u91cd\uff0c\u4e0d\u518d\u91cd\u590d\u4e0b\u8f7d\u5927\u578b\u6846\u67b6\u3002<\/p>\n\n\n\n
                \n\n\n\n
                \u5341\u4e00\u3001\u6a21\u578b\u4e0b\u8f7d\u95ee\u9898<\/h2>\n\n\n\n
                1. hf --dry-run<\/code> \u4e0d\u652f\u6301<\/h3>\n\n\n\n
                \u9519\u8bef\uff1a<\/p>\n\n\n\n
                hf: error: unrecognized arguments: --dry-run<\/code><\/pre>\n\n\n\n
                \u539f\u56e0\uff1a\u670d\u52a1\u5668\u4e0a\u7684 Hugging Face CLI \u7248\u672c\u8f83\u65e7\u3002<\/p>\n\n\n\n
                \u89e3\u51b3\uff1a\u53bb\u6389 --dry-run<\/code>\uff0c\u4e0d\u4e3a\u4e86\u8be5\u53c2\u6570\u5347\u7ea7\u6574\u4e2a\u73af\u5883\uff0c\u907f\u514d\u7834\u574f\u5df2\u8fd0\u884c\u7684 ASR\u3002<\/p>\n\n\n\n
                2. Hugging Face \u65e0\u6cd5\u8bbf\u95ee<\/h3>\n\n\n\n
                \u9519\u8bef\uff1a<\/p>\n\n\n\n
                OSError: [Errno 101] Network is unreachable<\/code><\/pre>\n\n\n\n
                \u539f\u56e0\uff1a\u5b66\u6821\u670d\u52a1\u5668\u65e0\u6cd5\u8bbf\u95ee huggingface.co:443<\/code>\u3002<\/p>\n\n\n\n
                \u89e3\u51b3\u8def\u5f84\uff1a<\/p>\n\n\n\n
                \u4f18\u5148 ModelScope\n\u2192 \u82e5\u4ecd\u4e0d\u53ef\u7528\n\u2192 Windows \u4e0b\u8f7d\n\u2192 SCP \u4e0a\u4f20<\/code><\/pre>\n\n\n\n
                3. \u7f3a\u5c11 quantize_config.json<\/code><\/h3>\n\n\n\n
                \u68c0\u67e5\u811a\u672c\u8bef\u5224\u6a21\u578b\u4e0d\u5b8c\u6574\u3002\u5b9e\u9645\u4e0a GPTQ \u914d\u7f6e\u4f4d\u4e8e\uff1a<\/p>\n\n\n\n
                config.json \u2192 quantization_config<\/code><\/pre>\n\n\n\n
                \u56e0\u6b64\u4e0d\u9700\u8981\u5355\u72ec\u7684 quantize_config.json<\/code>\u3002<\/p>\n\n\n\n
                \n\n\n\n
                \u5341\u4e8c\u3001vLLM LLM \u670d\u52a1\u90e8\u7f72<\/h2>\n\n\n\n
                LLM \u670d\u52a1\u76d1\u542c\uff1a<\/p>\n\n\n\n
                127.0.0.1:10096<\/code><\/pre>\n\n\n\n
                Token \u4fdd\u5b58\uff1a<\/p>\n\n\n\n
                mkdir -p \/app\/hd\/data01\/username\/projects\/qwen-llm-service\n\nopenssl rand -hex 32 \\\n  > \/app\/hd\/data01\/username\/projects\/qwen-llm-service\/.llm_token\n\nchmod 600 \\\n  \/app\/hd\/data01\/username\/projects\/qwen-llm-service\/.llm_token<\/code><\/pre>\n\n\n\n
                \u8f7d\u5165 Token\uff1a<\/p>\n\n\n\n
                export LLM_API_TOKEN=\"$(\n  cat \/app\/hd\/data01\/username\/projects\/qwen-llm-service\/.llm_token\n)\"<\/code><\/pre>\n\n\n\n
                \u542f\u52a8\uff1a<\/p>\n\n\n\n
                MODEL_DIR=\"\/app\/hd\/data01\/username\/models\/Qwen2.5-1.5B-Instruct-GPTQ-Int4\"\n\nCUDA_VISIBLE_DEVICES=<LLM_GPU_ID> \\\nvllm serve \"$MODEL_DIR\" \\\n  --served-model-name business-llm \\\n  --host 127.0.0.1 \\\n  --port 10096 \\\n  --max-model-len 4096 \\\n  --gpu-memory-utilization 0.25 \\\n  --api-key \"$LLM_API_TOKEN\" \\\n  --enable-auto-tool-choice \\\n  --tool-call-parser hermes<\/code><\/pre>\n\n\n\n
                \u5173\u952e\u53c2\u6570\uff1a<\/p>\n\n\n\n
                --served-model-name business-llm<\/code><\/pre>\n\n\n\n
                \u65e0\u8bba\u5b9e\u9645\u90e8\u7f72\u4ec0\u4e48\u6a21\u578b\uff0c\u5bf9\u5916\u7edf\u4e00\u53eb business-llm<\/code>\u3002<\/p>\n\n\n\n
                \u95ee\u9898\uff1a\u81ea\u52a8\u5de5\u5177\u8c03\u7528\u672a\u542f\u7528<\/h3>\n\n\n\n
                \u9519\u8bef\uff1a<\/p>\n\n\n\n
                \"auto\" tool choice requires\n--enable-auto-tool-choice\nand\n--tool-call-parser<\/code><\/pre>\n\n\n\n
                \u89e3\u51b3\uff1a<\/p>\n\n\n\n
                --enable-auto-tool-choice\n--tool-call-parser hermes<\/code><\/pre>\n\n\n\n
                \u95ee\u9898\uff1a\u6a21\u578b\u522b\u540d\u4e0d\u5b58\u5728<\/h3>\n\n\n\n
                \u9519\u8bef\uff1a<\/p>\n\n\n\n
                The model `business-llm` does not exist.<\/code><\/pre>\n\n\n\n
                \u539f\u56e0\uff1aWindows \u8bf7\u6c42 business-llm<\/code>\uff0c\u4f46 vLLM \u66b4\u9732\u7684\u662f\u539f\u59cb\u6a21\u578b\u540d\u3002<\/p>\n\n\n\n
                \u63a8\u8350\u89e3\u51b3\uff1a\u91cd\u65b0\u542f\u52a8 vLLM\uff0c\u5e76\u52a0\u5165\uff1a<\/p>\n\n\n\n
                --served-model-name business-llm<\/code><\/pre>\n\n\n\n
                \n\n\n\n
                \u5341\u4e09\u3001Windows \u540c\u65f6\u5efa\u7acb ASR \u4e0e LLM \u96a7\u9053<\/h2>\n\n\n\n
                \u670d\u52a1\u5668\u7aef\u53e3\uff1a<\/p>\n\n\n\n
                ASR\uff1a10095\nLLM\uff1a10096<\/code><\/pre>\n\n\n\n
                Windows \u672c\u5730\u7aef\u53e3\uff1a<\/p>\n\n\n\n
                ASR\uff1a18095\nLLM\uff1a18096<\/code><\/pre>\n\n\n\n
                \u7edf\u4e00 SSH \u547d\u4ee4\uff1a<\/p>\n\n\n\n
                ssh -N `\n  -o ExitOnForwardFailure=yes `\n  -o ServerAliveInterval=30 `\n  -o ServerAliveCountMax=3 `\n  -L 127.0.0.1:18095:127.0.0.1:10095 `\n  -L 127.0.0.1:18096:127.0.0.1:10096 `\n  username@<GPU_SERVER_IP><\/code><\/pre>\n\n\n\n
                \u6620\u5c04\u5173\u7cfb\uff1a<\/p>\n\n\n\n
                Windows 18095 \u2192 \u670d\u52a1\u5668 10095 \u2192 ASR\nWindows 18096 \u2192 \u670d\u52a1\u5668 10096 \u2192 LLM<\/code><\/pre>\n\n\n\n
                \u6d4b\u8bd5\uff1a<\/p>\n\n\n\n
                Test-NetConnection 127.0.0.1 -Port 18095\nTest-NetConnection 127.0.0.1 -Port 18096<\/code><\/pre>\n\n\n\n
                \n\n\n\n
                \u5341\u56db\u3001Windows LLM \u914d\u7f6e<\/h2>\n\n\n\n
                Windows \u9879\u76ee .env<\/code>\uff1a<\/p>\n\n\n\n
                LLM_BASE_URL=http:\/\/127.0.0.1:18096\/v1\nLLM_API_KEY=<LLM_TOKEN>\nLLM_MODEL=business-llm\nLLM_TIMEOUT=60<\/code><\/pre>\n\n\n\n
                \u901a\u7528\u5ba2\u6237\u7aef\uff1a<\/p>\n\n\n\n
                def get_chat_llm_base_url() -> str:\n    base_url = (\n        os.environ.get(\"LLM_BASE_URL\")\n        or \"\"\n    ).strip()\n\n    if not base_url:\n        raise RuntimeError(\n            \"LLM_BASE_URL is not configured.\"\n        )\n\n    return base_url.rstrip(\"\/\")\n\n\ndef get_chat_llm_api_key() -> str:\n    return (\n        os.environ.get(\"LLM_API_KEY\")\n        or \"local-llm\"\n    ).strip()\n\n\ndef get_chat_llm_model() -> str:\n    model = (\n        os.environ.get(\"LLM_MODEL\")\n        or \"\"\n    ).strip()\n\n    if not model:\n        raise RuntimeError(\n            \"LLM_MODEL is not configured.\"\n        )\n\n    return model\n\n\ndef create_chat_llm_client() -> OpenAI:\n    return OpenAI(\n        api_key=get_chat_llm_api_key(),\n        base_url=get_chat_llm_base_url(),\n        timeout=float(\n            os.environ.get(\n                \"LLM_TIMEOUT\",\n                \"60\",\n            )\n        ),\n    )<\/code><\/pre>\n\n\n\n
                \/api\/chat\/query<\/code> \u4e2d\u4f7f\u7528\uff1a<\/p>\n\n\n\n
                model_client = create_chat_llm_client()\n\nanswer_text = await run_in_threadpool(\n    run_agent,\n    model_client,\n    user_text,\n    page_context,\n    get_chat_llm_model(),\n)<\/code><\/pre>\n\n\n\n
                \n\n\n\n
                \u5341\u4e94\u3001.env<\/code> \u6ca1\u6709\u8bfb\u53d6\u7684\u95ee\u9898<\/h2>\n\n\n\n
                \u9519\u8bef\uff1a<\/p>\n\n\n\n
                RuntimeError: LLM_BASE_URL is not configured.<\/code><\/pre>\n\n\n\n
                \u539f\u56e0\uff1a<\/p>\n\n\n\n
                  \n
                • Uvicorn \u8fdb\u7a0b\u542f\u52a8\u524d\u6ca1\u6709\u8bbe\u7f6e\u73af\u5883\u53d8\u91cf\uff1b<\/li>\n\n\n\n
                • .env<\/code> \u6ca1\u6709\u4e3b\u52a8\u52a0\u8f7d\uff1b<\/li>\n\n\n\n
                • PowerShell \u4e2d\u8bbe\u7f6e\u53d8\u91cf\u540e\u6ca1\u6709\u91cd\u542f Uvicorn\u3002<\/li>\n<\/ul>\n\n\n\n
                  \u6b63\u786e\u505a\u6cd5\uff1a<\/p>\n\n\n\n
                  from pathlib import Path\nfrom dotenv import load_dotenv\n\nPROJECT_ROOT = Path(__file__).resolve().parent\n\nload_dotenv(\n    PROJECT_ROOT \/ \".env\"\n)<\/code><\/pre>\n\n\n\n
                  \u6ce8\u610f import \u5fc5\u987b\u5728\u8c03\u7528\u4e4b\u524d\u3002<\/p>\n\n\n\n
                  \u7c98\u8d34\u4ee3\u7801\u5bfc\u81f4\u7684\u9519\u8bef<\/h3>\n\n\n\n
                  \u9519\u8bef\uff1a<\/p>\n\n\n\n
                  FiePROJECT_ROOT = Path(__file__).resolve().parentld,<\/code><\/pre>\n\n\n\n
                  \u6b63\u786e\uff1a<\/p>\n\n\n\n
                  from pydantic import BaseModel, Field\n\nPROJECT_ROOT = Path(__file__).resolve().parent<\/code><\/pre>\n\n\n\n
                  \u8bed\u6cd5\u68c0\u67e5\uff1a<\/p>\n\n\n\n
                  python -m py_compile .\\voice_api.py<\/code><\/pre>\n\n\n\n
                  \u914d\u7f6e\u68c0\u67e5\uff1a<\/p>\n\n\n\n
                  python -c \"\nfrom voice_api import get_chat_llm_base_url, get_chat_llm_model\nprint(get_chat_llm_base_url())\nprint(get_chat_llm_model())\n\"<\/code><\/pre>\n\n\n\n
                  \n\n\n\n
                  \u5341\u516d\u3001Windows \u4e2d\u6587\u4e71\u7801<\/h2>\n\n\n\n
                  \u73b0\u8c61\uff1a<\/p>\n\n\n\n
                  answer_text = \u00e6\u00a8\u00e7...<\/code><\/pre>\n\n\n\n
                  \u8bf4\u660e\u8bf7\u6c42\u5df2\u7ecf\u6210\u529f\uff0c\u53ea\u662f Windows PowerShell 5.1 \u89e3\u7801 UTF-8 \u65f6\u51fa\u73b0\u4e71\u7801\u3002<\/p>\n\n\n\n
                  \u7ec8\u7aef\u8bbe\u7f6e\uff1a<\/p>\n\n\n\n
                  chcp 65001 > $null\n\n$env:PYTHONUTF8 = \"1\"\n$env:PYTHONIOENCODING = \"utf-8\"\n\n[Console]::InputEncoding = [System.Text.UTF8Encoding]::new()\n[Console]::OutputEncoding = [System.Text.UTF8Encoding]::new()\n$OutputEncoding = [System.Text.UTF8Encoding]::new()<\/code><\/pre>\n\n\n\n
                  \u53d1\u9001 UTF-8 \u8bf7\u6c42\u4f53\uff1a<\/p>\n\n\n\n
                  $json = $payload | ConvertTo-Json -Depth 10\n\n$utf8Body = [System.Text.Encoding]::UTF8.GetBytes(\n    $json\n)\n\nInvoke-RestMethod `\n  -Uri \"http:\/\/127.0.0.1:8001\/api\/chat\/query\" `\n  -Method Post `\n  -ContentType \"application\/json; charset=utf-8\" `\n  -Body $utf8Body<\/code><\/pre>\n\n\n\n
                  \u66f4\u7a33\u5b9a\u7684\u65b9\u5f0f\u662f\u4f7f\u7528 Python \u8bf7\u6c42\u5e76\u663e\u5f0f decode(\"utf-8\")<\/code>\u3002<\/p>\n\n\n\n
                  \n\n\n\n
                  \u5341\u4e03\u3001\u5c0f\u6a21\u578b\u6548\u679c\u5dee\u7684\u539f\u56e0<\/h2>\n\n\n\n
                  \u5b9e\u9645\u524d\u7aef\u6d4b\u8bd5\u4e2d\uff0c\u6a21\u578b\u51fa\u73b0\uff1a<\/p>\n\n\n\n
                    \n
                  • \u7528\u6237\u5df2\u7ecf\u63d0\u4f9b\u5b8c\u6574\u67e5\u8be2\u6761\u4ef6\uff0c\u6a21\u578b\u4ecd\u53cd\u590d\u8ffd\u95ee\uff1b<\/li>\n\n\n\n
                  • \u53ea\u4f1a\u590d\u8ff0\u9700\u6c42\uff1b<\/li>\n\n\n\n
                  • \u591a\u8f6e\u7406\u89e3\u8f83\u5dee\uff1b<\/li>\n\n\n\n
                  • \u5de5\u5177\u53c2\u6570\u5bb9\u6613\u4e22\u5931\uff1b<\/li>\n\n\n\n
                  • \u201c\u751f\u5316\u8fdb\u6c34\u6c28\u6c2e\u201d\u7f29\u6210\u201c\u6c28\u6c2e\u201d\uff1b<\/li>\n\n\n\n
                  • \u5355\u65e5\u67e5\u8be2\u672a\u8865\u9f50 end_date<\/code>\u3002<\/li>\n<\/ul>\n\n\n\n
                    \u539f\u56e0\u4e3b\u8981\u6709\u4e24\u7c7b\u3002<\/p>\n\n\n\n
                    1. \u6a21\u578b\u592a\u5c0f<\/h3>\n\n\n\n
                    \u5f53\u524d\u6a21\u578b\uff1a<\/p>\n\n\n\n
                    Qwen2.5-1.5B-Instruct-GPTQ-Int4<\/code><\/pre>\n\n\n\n
                    \u5b9a\u4f4d\u5e94\u662f\uff1a<\/p>\n\n\n\n
                    \u94fe\u8def\u6d4b\u8bd5\u6a21\u578b<\/code><\/pre>\n\n\n\n
                    \u800c\u4e0d\u662f\u6700\u7ec8\u751f\u4ea7\u6a21\u578b\u3002<\/p>\n\n\n\n
                    2. \u5bf9\u8bdd\u4e0a\u4e0b\u6587\u53ef\u80fd\u672a\u6301\u7eed\u4f20\u56de<\/h3>\n\n\n\n
                    \u540e\u7aef\u7684 append_conversation(...)<\/code> \u53ea\u66f4\u65b0\u672c\u6b21\u54cd\u5e94\u4e2d\u7684 context<\/code>\uff0c\u4e0d\u4f1a\u81ea\u52a8\u6c38\u4e45\u4fdd\u5b58\u3002<\/p>\n\n\n\n
                    \u524d\u7aef\u5fc5\u987b\u4fdd\u5b58\u540e\u7aef\u8fd4\u56de\u7684 result.context<\/code>\uff0c\u5e76\u5728\u4e0b\u4e00\u8f6e\u7ee7\u7eed\u53d1\u9001\u3002\u4e0d\u80fd\u6bcf\u6b21\u90fd\u91cd\u65b0\u53d1\u9001\uff1a<\/p>\n\n\n\n
                    {\n  \"conversation\": []\n}<\/code><\/pre>\n\n\n\n
                    \n\n\n\n
                    \u5341\u516b\u3001\u5207\u56de\u4e91\u7aef LLM API<\/h2>\n\n\n\n
                    \u5f53\u524d\u67b6\u6784\u5df2\u7ecf\u662f OpenAI \u517c\u5bb9\u5ba2\u6237\u7aef\uff0c\u56e0\u6b64\u4ece\u672c\u5730 vLLM \u5207\u56de\u4e91\u7aef API\uff0c\u53ea\u6539 .env<\/code>\u3002<\/p>\n\n\n\n
                    \u672c\u5730\uff1a<\/p>\n\n\n\n
                    LLM_BASE_URL=http:\/\/127.0.0.1:18096\/v1\nLLM_API_KEY=<LOCAL_TOKEN>\nLLM_MODEL=business-llm<\/code><\/pre>\n\n\n\n
                    \u4e91\u7aef\u793a\u4f8b\uff1a<\/p>\n\n\n\n
                    LLM_BASE_URL=https:\/\/<WORKSPACE>.cn-beijing.maas.aliyuncs.com\/compatible-mode\/v1\nLLM_API_KEY=<CLOUD_API_KEY>\nLLM_MODEL=qwen-plus\nLLM_TIMEOUT=60<\/code><\/pre>\n\n\n\n
                    \u5207\u56de\u4e91\u7aef\u540e\uff1a<\/p>\n\n\n\n
                    LLM SSH \u96a7\u9053\u4e0d\u518d\u9700\u8981\nASR SSH \u96a7\u9053\u7ee7\u7eed\u4fdd\u7559<\/code><\/pre>\n\n\n\n
                    \n\n\n\n
                    \u5341\u4e5d\u3001\u7532\u65b9\u751f\u4ea7\u73af\u5883\u5efa\u8bae<\/h2>\n\n\n\n
                    \u5b66\u6821\u6d4b\u8bd5\u6a21\u578b\u53ea\u9a8c\u8bc1\uff1a<\/p>\n\n\n\n
                    URL \u53ef\u66ff\u6362\nOpenAI \u534f\u8bae\u53ef\u66ff\u6362\nSSH\/\u5185\u7f51\u94fe\u8def\u53ef\u66ff\u6362\n\u6a21\u578b\u522b\u540d\u53ef\u7edf\u4e00<\/code><\/pre>\n\n\n\n
                    \u7532\u65b9\u6b63\u5f0f\u90e8\u7f72\u5efa\u8bae\uff1a<\/p>\n\n\n\n
                      \n
                    • \u81f3\u5c11 4B\uff5e8B \u7ea7\u522b\u6307\u4ee4\u6a21\u578b\uff1b<\/li>\n\n\n\n
                    • \u4f18\u5148\u9009\u62e9\u5de5\u5177\u8c03\u7528\u7a33\u5b9a\u7684\u6a21\u578b\uff1b<\/li>\n\n\n\n
                    • \u751f\u4ea7\u73af\u5883\u7edf\u4e00 served-model-name=business-llm<\/code>\uff1b<\/li>\n\n\n\n
                    • \u4e1a\u52a1\u540e\u7aef\u53ea\u8bfb\u53d6 URL\u3001Token \u548c\u6a21\u578b\u540d\uff1b<\/li>\n\n\n\n
                    • \u751f\u4ea7\u6a21\u578b\u4e0d\u8981\u76f4\u63a5\u88f8\u9732\u516c\u7f51\u3002<\/li>\n<\/ul>\n\n\n\n
                      \u793a\u4f8b\uff1a<\/p>\n\n\n\n
                      CUDA_VISIBLE_DEVICES=0 \\\nvllm serve \/models\/<BETTER_MODEL> \\\n  --served-model-name business-llm \\\n  --host 0.0.0.0 \\\n  --port 10096 \\\n  --max-model-len 8192 \\\n  --gpu-memory-utilization 0.60 \\\n  --enable-auto-tool-choice \\\n  --tool-call-parser hermes \\\n  --api-key \"$LLM_API_TOKEN\"<\/code><\/pre>\n\n\n\n
                      \u7f51\u7edc\u5efa\u8bae\uff1a<\/p>\n\n\n\n
                      \u4e1a\u52a1\u540e\u7aef\u4e0e\u6a21\u578b\u540c\u673a\uff1a127.0.0.1\n\u4e1a\u52a1\u540e\u7aef\u4e0e\u6a21\u578b\u5206\u673a\uff1a\u5185\u7f51 IP + \u9632\u706b\u5899\u767d\u540d\u5355\n\u8de8\u7f51\u7edc\uff1aNginx \/ \u7f51\u5173 + TLS + WSS\/HTTPS<\/code><\/pre>\n\n\n\n
                      \n\n\n\n
                      \u4e8c\u5341\u3001\u7532\u65b9\u90e8\u7f72\u68c0\u67e5\u6e05\u5355<\/h2>\n\n\n\n
                      A. \u786c\u4ef6\u68c0\u67e5<\/h3>\n\n\n\n
                      nvidia-smi\ndf -h\nfree -h<\/code><\/pre>\n\n\n\n
                      \u786e\u8ba4\uff1a<\/p>\n\n\n\n
                        \n
                      • GPU \u578b\u53f7\u4e0e\u663e\u5b58\uff1b<\/li>\n\n\n\n
                      • CUDA \u9a71\u52a8\u53ef\u7528\uff1b<\/li>\n\n\n\n
                      • \u6a21\u578b\u76d8\u7a7a\u95f4\u5145\u8db3\uff1b<\/li>\n\n\n\n
                      • \u6839\u5206\u533a\u4e0d\u4f1a\u88ab\u6a21\u578b\u5199\u6ee1\u3002<\/li>\n<\/ul>\n\n\n\n
                        B. \u76ee\u5f55\u89c4\u5212<\/h3>\n\n\n\n
                        \/data\/ai\/\n\u251c\u2500\u2500 conda_envs\/\n\u251c\u2500\u2500 models\/\n\u251c\u2500\u2500 projects\/\n\u251c\u2500\u2500 logs\/\n\u2514\u2500\u2500 cache\/<\/code><\/pre>\n\n\n\n
                        C. ASR \u670d\u52a1<\/h3>\n\n\n\n
                        \u7aef\u53e3\uff1a10095\n\u534f\u8bae\uff1aWebSocket\n\u8def\u5f84\uff1a\/v1\/asr\/stream\n\u5065\u5eb7\u68c0\u67e5\uff1a\/health<\/code><\/pre>\n\n\n\n
                        \u9a8c\u8bc1\uff1a<\/p>\n\n\n\n
                        curl http:\/\/127.0.0.1:10095\/health<\/code><\/pre>\n\n\n\n
                        D. LLM \u670d\u52a1<\/h3>\n\n\n\n
                        \u7aef\u53e3\uff1a10096\n\u534f\u8bae\uff1aOpenAI compatible\n\u6a21\u578b\u540d\uff1abusiness-llm<\/code><\/pre>\n\n\n\n
                        \u9a8c\u8bc1\uff1a<\/p>\n\n\n\n
                        curl \\\n  http:\/\/127.0.0.1:10096\/v1\/models \\\n  -H \"Authorization: Bearer $LLM_API_TOKEN\"<\/code><\/pre>\n\n\n\n
                        E. \u4e1a\u52a1\u540e\u7aef .env<\/code><\/h3>\n\n\n\n
                        ASR_REMOTE_WS_URL=ws:\/\/<ASR_HOST>:10095\/v1\/asr\/stream\nASR_API_TOKEN=<ASR_TOKEN>\n\nLLM_BASE_URL=http:\/\/<LLM_HOST>:10096\/v1\nLLM_API_KEY=<LLM_TOKEN>\nLLM_MODEL=business-llm\nLLM_TIMEOUT=60<\/code><\/pre>\n\n\n\n
                        F. \u7aef\u5230\u7aef\u6d4b\u8bd5\u987a\u5e8f<\/h3>\n\n\n\n
                        1. ASR \/health\n2. LLM \/v1\/models\n3. LLM \/v1\/chat\/completions\n4. \u4e1a\u52a1\u540e\u7aef \/health\n5. \/api\/chat\/query\n6. WebSocket \/ws\/asr\n7. \u6d4f\u89c8\u5668\u9ea6\u514b\u98ce\n8. \u5de5\u5177\u8c03\u7528\n9. \u6700\u7ec8\u56de\u7b54<\/code><\/pre>\n\n\n\n
                        \n\n\n\n
                        \u4e8c\u5341\u4e00\u3001\u5efa\u8bae\u8865\u5145\u7684 LLM \u542f\u52a8\u811a\u672c<\/h2>\n\n\n\n
                        \u5efa\u8bae\u521b\u5efa\uff1a<\/p>\n\n\n\n
                        \/app\/hd\/data01\/username\/projects\/qwen-llm-service\/start_llm.sh<\/code><\/pre>\n\n\n\n
                        \u793a\u4f8b\uff1a<\/p>\n\n\n\n
                        #!\/usr\/bin\/env bash\n\nset -euo pipefail\n\nCONDA_ENV=\"\/app\/hd\/data01\/username\/conda_envs\/qwen3-asr\"\nMODEL_DIR=\"\/app\/hd\/data01\/username\/models\/Qwen2.5-1.5B-Instruct-GPTQ-Int4\"\nTOKEN_FILE=\"\/app\/hd\/data01\/username\/projects\/qwen-llm-service\/.llm_token\"\n\nsource \"\/home\/username\/miniconda3\/etc\/profile.d\/conda.sh\"\nconda activate \"$CONDA_ENV\"\n\nexport LLM_API_TOKEN=\"$(cat \"$TOKEN_FILE\")\"\n\nexec env CUDA_VISIBLE_DEVICES=\"<LLM_GPU_ID>\" \\\nvllm serve \"$MODEL_DIR\" \\\n  --served-model-name business-llm \\\n  --host 127.0.0.1 \\\n  --port 10096 \\\n  --max-model-len 4096 \\\n  --gpu-memory-utilization 0.25 \\\n  --api-key \"$LLM_API_TOKEN\" \\\n  --enable-auto-tool-choice \\\n  --tool-call-parser hermes<\/code><\/pre>\n\n\n\n
                        \u8d4b\u6743\uff1a<\/p>\n\n\n\n
                        chmod +x start_llm.sh<\/code><\/pre>\n\n\n\n
                        \n\n\n\n
                        \u4e8c\u5341\u4e8c\u3001\u4eca\u65e5\u6700\u7ec8\u6210\u679c<\/h2>\n\n\n\n
                        \u4eca\u5929\u5df2\u7ecf\u5b8c\u6210\uff1a<\/p>\n\n\n\n
                          \n
                        • \u6570\u636e\u76d8\u76ee\u5f55\u89c4\u5212\uff1b<\/li>\n\n\n\n
                        • Conda \u73af\u5883\u3001\u6a21\u578b\u548c\u7f13\u5b58\u8fc1\u79fb\u5230\u5927\u5bb9\u91cf\u6570\u636e\u76d8\uff1b<\/li>\n\n\n\n
                        • Qwen3-ASR \u672c\u5730\u670d\u52a1\u90e8\u7f72\uff1b<\/li>\n\n\n\n
                        • \u81ea\u5b9a\u4e49 ASR WebSocket \u534f\u8bae\uff1b<\/li>\n\n\n\n
                        • Windows SSH \u96a7\u9053\u8fde\u63a5\uff1b<\/li>\n\n\n\n
                        • Windows WAV \u8fdc\u7a0b\u8bc6\u522b\u6d4b\u8bd5\uff1b<\/li>\n\n\n\n
                        • voice_api.py<\/code> \u4ece DashScope \u5b9e\u65f6 ASR \u6539\u4e3a\u8fdc\u7a0b ASR \u4ee3\u7406\uff1b<\/li>\n\n\n\n
                        • Qwen2.5 1.5B GPTQ Int4 \u6a21\u578b\u4e0b\u8f7d\uff1b<\/li>\n\n\n\n
                        • vLLM OpenAI \u517c\u5bb9 LLM \u670d\u52a1\u90e8\u7f72\uff1b<\/li>\n\n\n\n
                        • \u81ea\u52a8\u5de5\u5177\u8c03\u7528\u542f\u7528\uff1b<\/li>\n\n\n\n
                        • business-llm<\/code> \u7edf\u4e00\u6a21\u578b\u522b\u540d\uff1b<\/li>\n\n\n\n
                        • Windows LLM \u96a7\u9053\uff1b<\/li>\n\n\n\n
                        • \/api\/chat\/query<\/code> \u5b8c\u6574\u94fe\u8def\u9a8c\u8bc1\uff1b<\/li>\n\n\n\n
                        • \u6d4f\u89c8\u5668\u7aef ASR \u2192 LLM \u2192 Agent \u2192 \u56de\u7b54\u94fe\u8def\u9a8c\u8bc1\uff1b<\/li>\n\n\n\n
                        • \u4e91\u7aef API \u4e0e\u672c\u5730 LLM \u7684\u53ef\u5207\u6362\u914d\u7f6e\u3002<\/li>\n<\/ul>\n\n\n\n
                          \u6700\u7ec8\u9a8c\u8bc1\u7684\u6838\u5fc3\u67b6\u6784\uff1a<\/p>\n\n\n\n
                          \u6d4f\u89c8\u5668\n\u2192 Windows FastAPI\n\u2192 \u8fdc\u7a0b ASR URL\n\u2192 \u8fdc\u7a0b LLM URL\n\u2192 \u5de5\u5177\u8c03\u7528\n\u2192 \u524d\u7aef\u56de\u7b54<\/code><\/pre>\n\n\n\n
                          \u540e\u7eed\u5230\u7532\u65b9\u670d\u52a1\u5668\u65f6\uff0c\u4e0d\u9700\u8981\u91cd\u65b0\u8bbe\u8ba1\u4e1a\u52a1\u7cfb\u7edf\uff0c\u53ea\u9700\u8981\uff1a<\/p>\n\n\n\n
                          \u90e8\u7f72\u66f4\u597d\u7684 ASR\/LLM \u6a21\u578b\n\u2192 \u4fdd\u6301\u76f8\u540c\u534f\u8bae\n\u2192 \u4fee\u6539 .env\n\u2192 \u91cd\u542f\u670d\u52a1\n\u2192 \u6267\u884c\u9a8c\u6536\u6d4b\u8bd5<\/code><\/pre>\n\n\n\n
                          \n\n\n\n
                          \u4e8c\u5341\u4e09\u3001\u4e0b\u4e00\u6b21\u90e8\u7f72\u65f6\u6700\u5173\u952e\u7684\u4e94\u6761\u7ecf\u9a8c<\/h2>\n\n\n\n
                            \n
                          1. \u6a21\u578b\u3001\u73af\u5883\u3001\u7f13\u5b58\u5168\u90e8\u653e\u6570\u636e\u76d8\uff0c\u4e0d\u8981\u653e\u6839\u5206\u533a\u3002<\/strong><\/li>\n\n\n\n
                          2. ASR \u4e0e LLM \u5bf9\u5916\u534f\u8bae\u56fa\u5b9a\uff0c\u4e1a\u52a1\u4ee3\u7801\u53ea\u8bfb\u53d6 URL\u3002<\/strong><\/li>\n\n\n\n
                          3. vLLM \u6a21\u578b\u7edf\u4e00\u66b4\u9732\u4e3a business-llm<\/code>\u3002<\/strong><\/li>\n\n\n\n
                          4. \u5c0f\u6a21\u578b\u53ea\u505a\u94fe\u8def\u6d4b\u8bd5\uff0c\u7532\u65b9\u9a8c\u6536\u5fc5\u987b\u6362\u66f4\u597d\u7684\u6a21\u578b\u3002<\/strong><\/li>\n\n\n\n
                          5. \u5148\u9010\u5c42\u6d4b\u8bd5\uff0c\u518d\u6d4b\u6d4f\u89c8\u5668\u5b8c\u6574\u94fe\u8def\u3002<\/strong><\/li>\n<\/ol>\n\n\n\n
                            \u9010\u5c42\u987a\u5e8f\uff1a<\/p>\n\n\n\n
                            \u7aef\u53e3\n\u2192 \u5065\u5eb7\u68c0\u67e5\n\u2192 \u6a21\u578b\u63a5\u53e3\n\u2192 \u540e\u7aef\u63a5\u53e3\n\u2192 WebSocket\n\u2192 \u6d4f\u89c8\u5668\n\u2192 \u5de5\u5177\u8c03\u7528\n\u2192 \u6700\u7ec8\u56de\u7b54<\/code><\/pre>\n","protected":false},"excerpt":{"rendered":"
                            \u65e5\u671f\uff1a2026-07-20\u76ee\u6807\uff1a\u5728\u5b66\u6821 GPU \u670d\u52a1\u5668\u4e0a\u5b8c\u6210\u672c\u5730 ASR \u4e0e\u5c0f\u53c2\u6570 LLM \u7684\u90e8\u7f72\u9a8c\u8bc1\uff0c\u5e76\u8ba9  […]<\/p>\n","protected":false},"author":1,"featured_media":821,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_jetpack_memberships_contains_paid_content":false,"footnotes":""},"categories":[57,4],"tags":[68,72,71,73],"class_list":["post-818","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-python","category-blog","tag-linux","tag-llm","tag-qwen","tag-73"],"jetpack_sharing_enabled":true,"jetpack_featured_media_url":"https:\/\/blog.heartwarming.online\/wp-content\/uploads\/2026\/07\/pexels-photo-30855423.jpeg","_links":{"self":[{"href":"https:\/\/blog.heartwarming.online\/index.php\/wp-json\/wp\/v2\/posts\/818","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/blog.heartwarming.online\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/blog.heartwarming.online\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/blog.heartwarming.online\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/blog.heartwarming.online\/index.php\/wp-json\/wp\/v2\/comments?post=818"}],"version-history":[{"count":2,"href":"https:\/\/blog.heartwarming.online\/index.php\/wp-json\/wp\/v2\/posts\/818\/revisions"}],"predecessor-version":[{"id":820,"href":"https:\/\/blog.heartwarming.online\/index.php\/wp-json\/wp\/v2\/posts\/818\/revisions\/820"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/blog.heartwarming.online\/index.php\/wp-json\/wp\/v2\/media\/821"}],"wp:attachment":[{"href":"https:\/\/blog.heartwarming.online\/index.php\/wp-json\/wp\/v2\/media?parent=818"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/blog.heartwarming.online\/index.php\/wp-json\/wp\/v2\/categories?post=818"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/blog.heartwarming.online\/index.php\/wp-json\/wp\/v2\/tags?post=818"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}