\u5982\u679c\u4f60\u6b63\u5728\u5bfb\u627e\u4e00\u4e2a\u6027\u80fd\u5f3a\u52b2\u3001\u90e8\u7f72\u6210\u672c\u53ef\u63a7\u7684\u5f00\u6e90\u5927\u8bed\u8a00\u6a21\u578b，\u90a3\u4e48Google\u7684Gemma-3-12B-IT\u7edd\u5bf9\u503c\u5f97\u4f60\u82b1\u65f6\u95f4\u4e86\u89e3\u4e00\u4e0b\u3002\u6211\u6700\u8fd1\u5728\u4e00\u53f032GB\u5185\u5b58\u7684\u670d\u52a1\u5668\u4e0a\u5b8c\u6574\u90e8\u7f72\u5e76\u6d4b\u8bd5\u4e86\u8fd9\u4e2a\u6a21\u578b，\u6574\u4e2a\u8fc7\u7a0b\u6bd4\u60f3\u8c61\u4e2d\u8981\u987a\u5229\u5f97\u591a\u3002<\/p>\n
\u4f60\u53ef\u80fd\u542c\u8bf4\u8fc7\u52a8\u8f84\u51e0\u767eGB\u751a\u81f3\u4e0aTB\u7684\u5927\u6a21\u578b，\u90e8\u7f72\u8d77\u6765\u9700\u8981\u6602\u8d35\u7684\u4e13\u4e1a\u786c\u4ef6\u3002\u4f46Gemma-3-12B-IT\u53ea\u670923GB\u5927\u5c0f，\u5728\u666e\u901a\u670d\u52a1\u5668\u4e0a\u5c31\u80fd\u8dd1\u8d77\u6765，\u800c\u4e14\u6548\u679c\u76f8\u5f53\u4e0d\u9519\u3002\u5b83\u4e13\u95e8\u9488\u5bf9\u4eba\u7c7b\u6307\u4ee4\u8fdb\u884c\u4e86\u4f18\u5316，\u8fd9\u610f\u5473\u7740\u4f60\u4e0d\u9700\u8981\u590d\u6742\u7684\u63d0\u793a\u8bcd\u5de5\u7a0b，\u7528\u65e5\u5e38\u5bf9\u8bdd\u7684\u65b9\u5f0f\u5c31\u80fd\u83b7\u5f97\u9ad8\u8d28\u91cf\u7684\u56de\u590d\u3002<\/p>\n
\u8fd9\u7bc7\u6587\u7ae0\u6211\u4f1a\u5206\u4eab\u5b8c\u6574\u7684\u90e8\u7f72\u8fc7\u7a0b\u3001\u5b9e\u9645\u6d4b\u8bd5\u6548\u679c，\u4ee5\u53ca\u4e00\u4e9b\u4f60\u53ef\u80fd\u9047\u5230\u7684\u5751\u548c\u89e3\u51b3\u65b9\u6cd5\u3002\u65e0\u8bba\u4f60\u662f\u60f3\u642d\u5efa\u4e00\u4e2a\u5185\u90e8\u7684\u77e5\u8bc6\u95ee\u7b54\u52a9\u624b，\u8fd8\u662f\u9700\u8981\u4e00\u4e2a\u4ee3\u7801\u751f\u6210\u5de5\u5177，\u8fd9\u4e2a\u65b9\u6848\u90fd\u80fd\u7ed9\u4f60\u63d0\u4f9b\u53c2\u8003\u3002<\/p>\n
\u5728\u5f00\u59cb\u90e8\u7f72\u4e4b\u524d，\u6211\u4eec\u5148\u7b80\u5355\u4e86\u89e3\u4e00\u4e0bGemma-3-12B-IT\u5230\u5e95\u662f\u4ec0\u4e48，\u4ee5\u53ca\u5b83\u76f8\u6bd4\u524d\u4ee3\u6709\u54ea\u4e9b\u6539\u8fdb\u3002<\/p>\n
Gemma-3-12B-IT\u662fGoogle\u6700\u65b0\u53d1\u5e03\u7684\u8f7b\u91cf\u7ea7\u5f00\u6e90\u8bed\u8a00\u6a21\u578b，\u5c5e\u4e8eGemma\u7cfb\u5217\u7684\u7b2c\u4e09\u4ee3\u4ea7\u54c1\u3002\u8fd9\u91cc\u7684\u201c12B\u201d\u6307\u7684\u662f120\u4ebf\u53c2\u6570，\u8fd9\u4e2a\u89c4\u6a21\u5728\u5f53\u524d\u7684\u6a21\u578b\u751f\u6001\u4e2d\u5c5e\u4e8e\u201c\u751c\u70b9\u7ea7\u201d\u2014\u2014\u65e2\u6709\u4e0d\u9519\u7684\u6027\u80fd，\u53c8\u4e0d\u4f1a\u5bf9\u786c\u4ef6\u63d0\u51fa\u8fc7\u5206\u7684\u8981\u6c42\u3002<\/p>\n
\u201cIT\u201d\u4ee3\u8868Instruction Tuned，\u4e5f\u5c31\u662f\u6307\u4ee4\u5fae\u8c03\u7248\u672c\u3002\u8fd9\u610f\u5473\u7740\u6a21\u578b\u7ecf\u8fc7\u4e86\u4e13\u95e8\u7684\u8bad\u7ec3，\u80fd\u591f\u66f4\u597d\u5730\u7406\u89e3\u548c\u6267\u884c\u4eba\u7c7b\u7684\u6307\u4ee4\u3002\u6bd4\u5982\u4f60\u95ee\u201c\u5199\u4e00\u4e2aPython\u51fd\u6570\u6765\u8ba1\u7b97\u6590\u6ce2\u90a3\u5951\u6570\u5217\u201d，\u5b83\u4e0d\u4f1a\u7ed9\u4f60\u8bb2\u4e00\u5806\u6570\u5b66\u7406\u8bba，\u800c\u662f\u76f4\u63a5\u7ed9\u51fa\u53ef\u8fd0\u884c\u7684\u4ee3\u7801\u3002<\/p>\n
\u5982\u679c\u4f60\u7528\u8fc7Gemma 1\u6216Gemma 2，\u53ef\u80fd\u4f1a\u597d\u5947\u7b2c\u4e09\u4ee3\u5230\u5e95\u5f3a\u5728\u54ea\u91cc\u3002\u6839\u636e\u6211\u7684\u5b9e\u6d4b\u548c\u5b98\u65b9\u6587\u6863，\u4e3b\u8981\u63d0\u5347\u5728\u4e09\u4e2a\u65b9\u9762：<\/p>\n
\u63a8\u7406\u80fd\u529b\u660e\u663e\u589e\u5f3a：\u5728\u5904\u7406\u903b\u8f91\u63a8\u7406\u3001\u6570\u5b66\u8ba1\u7b97\u8fd9\u7c7b\u4efb\u52a1\u65f6，\u56de\u7b54\u7684\u51c6\u786e\u6027\u548c\u6761\u7406\u6027\u90fd\u6709\u63d0\u5347\u3002\u6211\u6d4b\u8bd5\u4e86\u51e0\u4e2a\u7ecf\u5178\u7684\u903b\u8f91\u8c1c\u9898，Gemma-3\u7684\u89e3\u9898\u601d\u8def\u66f4\u52a0\u6e05\u6670\u3002<\/p>\n
\u591a\u8bed\u8a00\u652f\u6301\u66f4\u597d：\u867d\u7136\u4e3b\u8981\u8fd8\u662f\u4ee5\u82f1\u6587\u8bad\u7ec3\u4e3a\u4e3b，\u4f46\u5bf9\u4e2d\u6587\u3001\u65e5\u6587\u3001\u6cd5\u6587\u7b49\u5176\u4ed6\u8bed\u8a00\u7684\u7406\u89e3\u548c\u751f\u6210\u80fd\u529b\u90fd\u6709\u6240\u6539\u5584\u3002\u6211\u7528\u4e2d\u6587\u63d0\u95ee\u65f6，\u5f97\u5230\u7684\u56de\u590d\u8d28\u91cf\u76f8\u5f53\u4e0d\u9519\u3002<\/p>\n
\u6548\u7387\u4f18\u5316\u663e\u8457：\u8fd9\u662f\u90e8\u7f72\u65f6\u611f\u53d7\u6700\u660e\u663e\u7684\u4e00\u70b9\u3002\u6a21\u578b\u5728\u63a8\u7406\u65f6\u7684\u5185\u5b58\u5360\u7528\u66f4\u52a0\u5408\u7406，\u54cd\u5e94\u901f\u5ea6\u4e5f\u66f4\u5feb\u3002\u5728\u540c\u6837\u7684\u786c\u4ef6\u6761\u4ef6\u4e0b，Gemma-3\u6bd4\u524d\u4ee3\u80fd\u5904\u7406\u66f4\u957f\u7684\u4e0a\u4e0b\u6587\u3002<\/p>\n
\u57fa\u4e8e\u6211\u7684\u6d4b\u8bd5，\u8fd9\u4e2a\u6a21\u578b\u7279\u522b\u9002\u5408\u4ee5\u4e0b\u51e0\u79cd\u5e94\u7528：<\/p>\n
\u5982\u679c\u4f60\u9700\u8981\u4e00\u4e2a\u65e2\u5f3a\u5927\u53c8\u5bb9\u6613\u90e8\u7f72\u7684AI\u52a9\u624b，Gemma-3-12B-IT\u662f\u4e2a\u5f88\u4e0d\u9519\u7684\u9009\u62e9\u3002<\/p>\n
\u90e8\u7f72\u5927\u6a21\u578b\u542c\u8d77\u6765\u53ef\u80fd\u6709\u70b9\u5413\u4eba，\u4f46\u5176\u5b9e\u53ea\u8981\u786c\u4ef6\u8fbe\u6807，\u8fc7\u7a0b\u5e76\u4e0d\u590d\u6742\u3002\u4e0b\u9762\u662f\u6211\u5b9e\u6d4b\u7684\u786c\u4ef6\u8981\u6c42\u548c\u51c6\u5907\u5de5\u4f5c\u3002<\/p>\n
\u6211\u7528\u7684\u662f\u4e00\u53f0\u6807\u51c6\u7684\u4e91\u670d\u52a1\u5668，\u914d\u7f6e\u5982\u4e0b：<\/p>\n
\u4e3a\u4ec0\u4e48\u9009\u62e9\u8fd9\u4e2a\u914d\u7f6e？<\/p>\n
23GB\u7684\u6a21\u578b\u6587\u4ef6\u52a0\u8f7d\u5230\u5185\u5b58\u540e，\u52a0\u4e0a\u7cfb\u7edf\u5f00\u9500\u548c\u63a8\u7406\u65f6\u7684\u4e34\u65f6\u5185\u5b58，32GB\u5185\u5b58\u521a\u597d\u591f\u7528\u3002\u5982\u679c\u5185\u5b58\u518d\u5c0f\u4e00\u4e9b，\u6bd4\u598224GB，\u53ef\u80fd\u5c31\u9700\u8981\u4f7f\u7528\u91cf\u5316\u7248\u672c\u6216\u8005\u5916\u6302\u4ea4\u6362\u7a7a\u95f4，\u90a3\u6837\u4f1a\u5f71\u54cd\u63a8\u7406\u901f\u5ea6\u3002<\/p>\n
\u5982\u679c\u6ca1\u6709GPU\u600e\u4e48\u529e？<\/p>\n
\u5b8c\u5168\u6ca1\u95ee\u9898\u3002\u6211\u8fd9\u6b21\u6d4b\u8bd5\u5c31\u662f\u7eafCPU\u73af\u5883，\u867d\u7136\u63a8\u7406\u901f\u5ea6\u6bd4\u4e0d\u4e0aGPU（\u5927\u7ea6\u6bcf\u79d2\u751f\u62103-5\u4e2atoken），\u4f46\u5bf9\u4e8e\u5927\u591a\u6570\u5bf9\u8bdd\u573a\u666f\u6765\u8bf4\u5b8c\u5168\u591f\u7528\u3002\u5982\u679c\u4f60\u6709NVIDIA GPU（\u663e\u5b588GB\u4ee5\u4e0a），\u901f\u5ea6\u4f1a\u5feb\u5f88\u591a\u3002<\/p>\n
\u5728\u5f00\u59cb\u90e8\u7f72\u4e4b\u524d，\u9700\u8981\u786e\u4fdd\u7cfb\u7edf\u73af\u5883\u6b63\u786e\u3002\u4ee5\u4e0b\u662f\u5fc5\u987b\u7684\u51c6\u5907\u5de5\u4f5c：<\/p>\n
\u7b2c\u4e00\u6b65：\u66f4\u65b0\u7cfb\u7edf\u5e76\u5b89\u88c5\u57fa\u7840\u5de5\u5177<\/p>\n
# \u66f4\u65b0\u7cfb\u7edf\u5305
\nsudo apt update && sudo apt upgrade -y<\/p>\n
# \u5b89\u88c5\u5fc5\u8981\u7684\u5de5\u5177
\nsudo apt install -y python3-pip python3-venv git curl wget<\/p>\n
# \u68c0\u67e5Python\u7248\u672c（\u9700\u89813.10\u4ee5\u4e0a）
\npython3 –version<\/p>\n
\u7b2c\u4e8c\u6b65：\u521b\u5efa\u4e13\u7528\u73af\u5883<\/p>\n
\u4e3a\u4e86\u907f\u514d\u4f9d\u8d56\u51b2\u7a81，\u5efa\u8bae\u4e3a\u6a21\u578b\u90e8\u7f72\u521b\u5efa\u72ec\u7acb\u7684Python\u73af\u5883：<\/p>\n
# \u521b\u5efa\u9879\u76ee\u76ee\u5f55
\nmkdir -p ~\/gemma-3-deployment
\ncd ~\/gemma-3-deployment<\/p>\n
# \u521b\u5efa\u865a\u62df\u73af\u5883
\npython3 -m venv venv<\/p>\n
# \u6fc0\u6d3b\u73af\u5883
\nsource venv\/bin\/activate<\/p>\n
\u7b2c\u4e09\u6b65：\u5b89\u88c5PyTorch<\/p>\n
PyTorch\u662f\u8fd0\u884c\u6a21\u578b\u7684\u57fa\u7840\u6846\u67b6，\u9700\u8981\u6839\u636e\u4f60\u7684\u786c\u4ef6\u9009\u62e9\u6b63\u786e\u7684\u7248\u672c：<\/p>\n
# \u5982\u679c\u662f\u7eafCPU\u73af\u5883
\npip install torch torchvision torchaudio –index-url https:\/\/download.pytorch.org\/whl\/cpu<\/p>\n
# \u5982\u679c\u6709NVIDIA GPU（CUDA 12.1）
\npip install torch torchvision torchaudio –index-url https:\/\/download.pytorch.org\/whl\/cu121<\/p>\n
\u5b89\u88c5\u5b8c\u6210\u540e，\u53ef\u4ee5\u9a8c\u8bc1\u4e00\u4e0b：<\/p>\n
# \u6d4b\u8bd5PyTorch\u662f\u5426\u5b89\u88c5\u6210\u529f
\npython3 -c "import torch; print(f'PyTorch\u7248\u672c: {torch.__version__}')"<\/p>\n
\u51c6\u5907\u597d\u4e86\u73af\u5883，\u6211\u4eec\u5c31\u53ef\u4ee5\u5f00\u59cb\u6b63\u5f0f\u7684\u90e8\u7f72\u4e86\u3002\u6574\u4e2a\u8fc7\u7a0b\u6211\u628a\u5b83\u5206\u4e3a\u56db\u4e2a\u4e3b\u8981\u6b65\u9aa4，\u8ddf\u7740\u505a\u5e94\u8be5\u4e0d\u4f1a\u6709\u4ec0\u4e48\u95ee\u9898\u3002<\/p>\n
Gemma-3-12B-IT\u7684\u6a21\u578b\u6587\u4ef6\u5927\u7ea623GB，\u4e0b\u8f7d\u9700\u8981\u4e00\u4e9b\u65f6\u95f4\u548c\u8db3\u591f\u7684\u78c1\u76d8\u7a7a\u95f4\u3002<\/p>\n
\u65b9\u6cd5\u4e00：\u4eceHugging Face\u4e0b\u8f7d（\u63a8\u8350）<\/p>\n
# \u5b89\u88c5huggingface-cli
\npip install huggingface-hub<\/p>\n
# \u8bbe\u7f6e\u7f13\u5b58\u76ee\u5f55（\u786e\u4fdd\u6709\u8db3\u591f\u7a7a\u95f4）
\nexport HF_HOME=~\/hf_cache<\/p>\n
# \u4e0b\u8f7d\u6a21\u578b
\npython3 -c "
\nfrom huggingface_hub import snapshot_download
\nsnapshot_download(
\n repo_id='google\/gemma-3-12b-it',
\n local_dir='.\/gemma-3-12b-it',
\n ignore_patterns=['*.safetensors', '*.bin'], # \u53ea\u4e0b\u8f7d\u5fc5\u8981\u7684\u6587\u4ef6
\n local_dir_use_symlinks=False
\n)
\n"<\/p>\n
\u65b9\u6cd5\u4e8c：\u624b\u52a8\u4e0b\u8f7d（\u5982\u679c\u7f51\u7edc\u4e0d\u7a33\u5b9a）<\/p>\n
\u5982\u679c\u76f4\u63a5\u4eceHugging Face\u4e0b\u8f7d\u901f\u5ea6\u592a\u6162，\u53ef\u4ee5\u5c1d\u8bd5\u7528\u5176\u4ed6\u65b9\u5f0f：<\/p>\n
# \u4f7f\u7528wget\u5206\u6bb5\u4e0b\u8f7d（\u5982\u679c\u6709\u76f4\u94fe）
\n# \u6216\u8005\u4f7f\u7528\u56fd\u5185\u955c\u50cf\u6e90<\/p>\n
# \u521b\u5efa\u6a21\u578b\u76ee\u5f55
\nmkdir -p ~\/models\/gemma-3-12b-it
\ncd ~\/models\/gemma-3-12b-it<\/p>\n
# \u8fd9\u91cc\u9700\u8981\u6839\u636e\u5b9e\u9645\u53ef\u7528\u7684\u4e0b\u8f7d\u94fe\u63a5\u6765\u8c03\u6574
\n# \u901a\u5e38\u53ef\u4ee5\u5728Hugging Face\u9875\u9762\u627e\u5230\u4e0b\u8f7d\u94fe\u63a5<\/p>\n
\u4e0b\u8f7d\u5b8c\u6210\u540e，\u68c0\u67e5\u6587\u4ef6\u5927\u5c0f：<\/p>\n
# \u67e5\u770b\u6a21\u578b\u6587\u4ef6\u5927\u5c0f
\ndu -sh ~\/models\/gemma-3-12b-it\/<\/p>\n
# \u5e94\u8be5\u663e\u793a\u5927\u7ea623GB<\/p>\n
\u6211\u9009\u62e9\u4e86\u4e24\u4e2a\u6bd4\u8f83\u6d41\u884c\u7684\u6846\u67b6\u8fdb\u884c\u6d4b\u8bd5：Transformers\u548cllama.cpp\u3002\u524d\u8005\u529f\u80fd\u5168\u9762，\u540e\u8005\u5728CPU\u4e0a\u6548\u7387\u66f4\u9ad8\u3002<\/p>\n
\u5b89\u88c5Transformers\u548c\u76f8\u5173\u4f9d\u8d56<\/p>\n
# \u786e\u4fdd\u5728\u865a\u62df\u73af\u5883\u4e2d
\nsource venv\/bin\/activate<\/p>\n
# \u5b89\u88c5\u5fc5\u8981\u7684\u5e93
\npip install transformers accelerate sentencepiece protobuf<\/p>\n
# \u5982\u679c\u9700\u8981Web\u754c\u9762，\u53ef\u4ee5\u5b89\u88c5Gradio
\npip install gradio==4.19.2<\/p>\n
\u5b89\u88c5llama.cpp（\u53ef\u9009，\u7528\u4e8eCPU\u4f18\u5316）<\/p>\n
\u5982\u679c\u4f60\u4e3b\u8981\u5728CPU\u4e0a\u8fd0\u884c，llama.cpp\u80fd\u63d0\u4f9b\u66f4\u597d\u7684\u6027\u80fd：<\/p>\n
# \u514b\u9686\u4ed3\u5e93
\ngit clone https:\/\/github.com\/ggerganov\/llama.cpp
\ncd llama.cpp<\/p>\n
# \u7f16\u8bd1
\nmake -j4<\/p>\n
# \u5c06\u6a21\u578b\u8f6c\u6362\u4e3agguf\u683c\u5f0f（\u9700\u8981\u5148\u4e0b\u8f7d\u539f\u59cb\u6a21\u578b）
\npython3 convert.py ~\/models\/gemma-3-12b-it –outfile gemma-3-12b-it.gguf<\/p>\n
# \u91cf\u5316\u6a21\u578b\u4ee5\u51cf\u5c0f\u5927\u5c0f（\u53ef\u9009）
\n.\/quantize gemma-3-12b-it.gguf gemma-3-12b-it-q4_0.gguf q4_0<\/p>\n
\u4e3a\u4e86\u8ba9\u6a21\u578b\u66f4\u5bb9\u6613\u4f7f\u7528，\u6211\u642d\u5efa\u4e86\u4e00\u4e2a\u7b80\u5355\u7684Web\u754c\u9762\u3002\u8fd9\u91cc\u6211\u7528Gradio，\u5b83\u8db3\u591f\u8f7b\u91cf\u4e14\u5bb9\u6613\u914d\u7f6e\u3002<\/p>\n
\u521b\u5efaWeb\u5e94\u7528\u6587\u4ef6<\/p>\n
# app.py
\nimport gradio as gr
\nfrom transformers import AutoTokenizer, AutoModelForCausalLM
\nimport torch
\nimport time<\/p>\n
class GemmaChatbot:
\n def __init__(self, model_path):
\n print("\u6b63\u5728\u52a0\u8f7d\u6a21\u578b…")
\n start_time = time.time()<\/p>\n
# \u52a0\u8f7dtokenizer\u548c\u6a21\u578b
\n self.tokenizer = AutoTokenizer.from_pretrained(model_path)
\n self.model = AutoModelForCausalLM.from_pretrained(
\n model_path,
\n torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
\n device_map="auto" if torch.cuda.is_available() else "cpu",
\n low_cpu_mem_usage=True
\n )<\/p>\n
# \u5982\u679c\u662fCPU\u73af\u5883，\u4f7f\u7528\u66f4\u8282\u7701\u5185\u5b58\u7684\u6a21\u5f0f
\n if not torch.cuda.is_available():
\n self.model = self.model.to(torch.float32)<\/p>\n
load_time = time.time() – start_time
\n print(f"\u6a21\u578b\u52a0\u8f7d\u5b8c\u6210，\u8017\u65f6: {load_time:.2f}\u79d2")<\/p>\n
# \u521d\u59cb\u5316\u5bf9\u8bdd\u5386\u53f2
\n self.conversation_history = []<\/p>\n
def generate_response(self, message, history, temperature=0.7, max_tokens=512):
\n # \u6784\u5efa\u63d0\u793a
\n prompt = self._build_prompt(message, history)<\/p>\n
# \u7f16\u7801\u8f93\u5165
\n inputs = self.tokenizer(prompt, return_tensors="pt")<\/p>\n
# \u751f\u6210\u56de\u590d
\n with torch.no_grad():
\n outputs = self.model.generate(
\n inputs.input_ids,
\n max_new_tokens=max_tokens,
\n temperature=temperature,
\n do_sample=True,
\n top_p=0.9,
\n pad_token_id=self.tokenizer.eos_token_id
\n )<\/p>\n
# \u89e3\u7801\u56de\u590d
\n response = self.tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)<\/p>\n
# \u66f4\u65b0\u5386\u53f2
\n self.conversation_history.append((message, response))<\/p>\n
return response<\/p>\n
def _build_prompt(self, message, history):
\n # \u7b80\u5355\u7684\u63d0\u793a\u6784\u5efa
\n prompt = "<start_of_turn>user\\\\n"
\n prompt += message + "<end_of_turn>\\\\n"
\n prompt += "<start_of_turn>model\\\\n"
\n return prompt<\/p>\n
# \u521d\u59cb\u5316\u804a\u5929\u673a\u5668\u4eba
\nmodel_path = "\/root\/models\/gemma-3-12b-it" # \u4fee\u6539\u4e3a\u4f60\u7684\u6a21\u578b\u8def\u5f84
\nchatbot = GemmaChatbot(model_path)<\/p>\n
# \u521b\u5efaGradio\u754c\u9762
\ndef chat_interface(message, history, temperature, max_tokens):
\n response = chatbot.generate_response(message, history, temperature, max_tokens)
\n return response<\/p>\n
# \u6784\u5efaWeb\u754c\u9762
\nwith gr.Blocks(title="Gemma-3-12B-IT Chatbot") as demo:
\n gr.Markdown("# 🤖 Gemma-3-12B-IT \u804a\u5929\u52a9\u624b")
\n gr.Markdown("\u57fa\u4e8eGoogle Gemma-3-12B-IT\u6a21\u578b\u7684\u5bf9\u8bdd\u7cfb\u7edf")<\/p>\n
chatbot = gr.Chatbot(height=400)
\n msg = gr.Textbox(label="\u8f93\u5165\u4f60\u7684\u95ee\u9898", placeholder="\u5728\u8fd9\u91cc\u8f93\u5165…")<\/p>\n
with gr.Row():
\n temperature = gr.Slider(0.1, 1.5, value=0.7, label="Temperature", info="\u63a7\u5236\u56de\u590d\u7684\u968f\u673a\u6027")
\n max_tokens = gr.Slider(64, 2048, value=512, step=64, label="\u6700\u5927\u751f\u6210\u957f\u5ea6")<\/p>\n
with gr.Row():
\n submit = gr.Button("\u53d1\u9001")
\n clear = gr.Button("\u6e05\u7a7a\u5bf9\u8bdd")<\/p>\n
def respond(message, chat_history, temp, tokens):
\n bot_message = chatbot.generate_response(message, chat_history, temp, tokens)
\n chat_history.append((message, bot_message))
\n return "", chat_history<\/p>\n
msg.submit(respond, [msg, chatbot, temperature, max_tokens], [msg, chatbot])
\n submit.click(respond, [msg, chatbot, temperature, max_tokens], [msg, chatbot])<\/p>\n
clear.click(lambda: None, None, chatbot, queue=False)<\/p>\n
# \u542f\u52a8\u670d\u52a1
\nif __name__ == "__main__":
\n demo.launch(
\n server_name="0.0.0.0",
\n server_port=7860,
\n share=False
\n )<\/p>\n
\u521b\u5efa\u542f\u52a8\u811a\u672c<\/p>\n
# start.sh
\n#!\/bin\/bash<\/p>\n
# \u6fc0\u6d3b\u865a\u62df\u73af\u5883
\nsource ~\/gemma-3-deployment\/venv\/bin\/activate<\/p>\n
# \u542f\u52a8Web\u670d\u52a1
\ncd ~\/gemma-3-deployment
\npython app.py<\/p>\n
\u7ed9\u811a\u672c\u6267\u884c\u6743\u9650：<\/p>\n
chmod +x start.sh<\/p>\n
\u572832GB\u5185\u5b58\u7684\u670d\u52a1\u5668\u4e0a\u8fd0\u884c23GB\u7684\u6a21\u578b，\u5185\u5b58\u7ba1\u7406\u5f88\u91cd\u8981\u3002\u4ee5\u4e0b\u662f\u6211\u7528\u7684\u4e00\u4e9b\u4f18\u5316\u6280\u5de7：<\/p>\n
\u76d1\u63a7\u5185\u5b58\u4f7f\u7528<\/p>\n
# \u5b9e\u65f6\u76d1\u63a7\u5185\u5b58\u4f7f\u7528
\nwatch -n 1 "free -h"<\/p>\n
# \u6216\u8005\u4f7f\u7528htop
\nhtop<\/p>\n
\u4f18\u5316Python\u5185\u5b58\u4f7f\u7528<\/p>\n
\u5728app.py\u4e2d\u6dfb\u52a0\u5185\u5b58\u4f18\u5316\u914d\u7f6e：<\/p>\n
import os
\nimport gc<\/p>\n
# \u8bbe\u7f6ePyTorch\u5185\u5b58\u5206\u914d\u7b56\u7565
\nos.environ['PYTORCH_CUDA_ALLOC_CONF'] = 'max_split_size_mb:128'<\/p>\n
# \u5b9a\u671f\u6e05\u7406\u5185\u5b58
\ndef cleanup_memory():
\n gc.collect()
\n if torch.cuda.is_available():
\n torch.cuda.empty_cache()<\/p>\n
\u4f7f\u7528\u4ea4\u6362\u7a7a\u95f4（\u5982\u679c\u5185\u5b58\u7d27\u5f20）<\/p>\n
\u5982\u679c32GB\u5185\u5b58\u4e0d\u591f\u7528，\u53ef\u4ee5\u6dfb\u52a0\u4ea4\u6362\u7a7a\u95f4：<\/p>\n
# \u521b\u5efa8GB\u7684\u4ea4\u6362\u6587\u4ef6
\nsudo fallocate -l 8G \/swapfile
\nsudo chmod 600 \/swapfile
\nsudo mkswap \/swapfile
\nsudo swapon \/swapfile<\/p>\n
# \u6c38\u4e45\u751f\u6548
\necho '\/swapfile none swap sw 0 0' | sudo tee -a \/etc\/fstab<\/p>\n
\u90e8\u7f72\u5b8c\u6210\u540e，\u6211\u8fdb\u884c\u4e86\u4e00\u7cfb\u5217\u6d4b\u8bd5，\u770b\u770b\u8fd9\u4e2a\u6a21\u578b\u5728\u5b9e\u9645\u4f7f\u7528\u4e2d\u7684\u8868\u73b0\u5982\u4f55\u3002<\/p>\n
\u6a21\u578b\u52a0\u8f7d\u65f6\u95f4<\/p>\n
\u572832GB\u5185\u5b58\u7684\u670d\u52a1\u5668\u4e0a，\u7eafCPU\u73af\u5883\u4e0b\u7684\u52a0\u8f7d\u65f6\u95f4：<\/p>\n
\u5185\u5b58\u5360\u7528\u60c5\u51b5<\/p>\n
\u6a21\u578b\u8fd0\u884c\u65f6\u7684\u5185\u5b58\u4f7f\u7528：<\/p>\n
CPU\u4f7f\u7528\u7387<\/p>\n
\u6211\u6d4b\u8bd5\u4e86\u4e0d\u540c\u7c7b\u578b\u95ee\u9898\u7684\u54cd\u5e94\u901f\u5ea6：<\/p>\n
# \u6d4b\u8bd5\u4ee3\u7801\u793a\u4f8b
\ntest_cases = [
\n ("\u7b80\u5355\u95ee\u5019", "\u4f60\u597d，\u8bf7\u4ecb\u7ecd\u4e00\u4e0b\u4f60\u81ea\u5df1"),
\n ("\u4ee3\u7801\u751f\u6210", "\u5199\u4e00\u4e2aPython\u51fd\u6570\u6765\u8ba1\u7b97\u6590\u6ce2\u90a3\u5951\u6570\u5217"),
\n ("\u77e5\u8bc6\u95ee\u7b54", "\u89e3\u91ca\u4e00\u4e0b\u91cf\u5b50\u8ba1\u7b97\u7684\u57fa\u672c\u539f\u7406"),
\n ("\u903b\u8f91\u63a8\u7406", "\u5982\u679c\u6240\u6709\u7684\u732b\u90fd\u6015\u6c34，\u800c\u6c64\u59c6\u662f\u4e00\u53ea\u732b，\u90a3\u4e48\u6c64\u59c6\u6015\u6c34\u5417？"),
\n]<\/p>\n
\u54cd\u5e94\u65f6\u95f4\u7edf\u8ba1（\u7eafCPU\u73af\u5883）：<\/p>\n
| \u7b80\u5355\u95ee\u5019<\/td>\n | 3-5\u79d2<\/td>\n | 50-100<\/td>\n<\/tr>\n |
| \u4ee3\u7801\u751f\u6210<\/td>\n | 8-15\u79d2<\/td>\n | 150-300<\/td>\n<\/tr>\n |
| \u77e5\u8bc6\u95ee\u7b54<\/td>\n | 6-10\u79d2<\/td>\n | 100-200<\/td>\n<\/tr>\n |
| \u903b\u8f91\u63a8\u7406<\/td>\n | 5-8\u79d2<\/td>\n | 80-150<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n \u6bcf\u79d2\u751f\u6210token\u6570：\u5927\u7ea63-5 tokens\/\u79d2<\/p>\n \u8fd9\u4e2a\u901f\u5ea6\u5bf9\u4e8e\u5bf9\u8bdd\u573a\u666f\u6765\u8bf4\u5b8c\u5168\u53ef\u4ee5\u63a5\u53d7\u3002\u5982\u679c\u662fGPU\u73af\u5883（\u6bd4\u5982RTX 4090），\u901f\u5ea6\u53ef\u4ee5\u63d0\u5347\u523050-100 tokens\/\u79d2\u3002<\/p>\n 5.3 \u8d28\u91cf\u6d4b\u8bd5：\u4e0d\u540c\u573a\u666f\u7684\u8868\u73b0<\/h4>\n\u7f16\u7a0b\u80fd\u529b\u6d4b\u8bd5<\/p>\n \u6211\u8ba9\u6a21\u578b\u5199\u51e0\u4e2a\u5e38\u89c1\u7684\u7b97\u6cd5\u5b9e\u73b0：<\/p>\n \u7528\u6237：\u5199\u4e00\u4e2aPython\u7684\u5feb\u901f\u6392\u5e8f\u5b9e\u73b0<\/p>\n \u52a9\u624b： # \u6d4b\u8bd5 \u4ee3\u7801\u8d28\u91cf\u8bc4\u4ef7：<\/p>\n
|