\u5982\u679c\u4f60\u6b63\u5728\u5bfb\u627e\u4e00\u4e2a\u65e2\u5f3a\u5927\u53c8\u80fd\u5728\u666e\u901a\u670d\u52a1\u5668\u4e0a\u6d41\u7545\u8fd0\u884c\u7684\u5927\u8bed\u8a00\u6a21\u578b，\u90a3\u4e48Gemma-3-12B-IT\u5f88\u53ef\u80fd\u5c31\u662f\u4f60\u7684\u7406\u60f3\u9009\u62e9\u3002<\/p>\n
\u8ba9\u6211\u5148\u8bf4\u8bf4\u6211\u7684\u4eb2\u8eab\u7ecf\u5386\u3002\u524d\u6bb5\u65f6\u95f4，\u6211\u9700\u8981\u4e3a\u4e00\u4e2a\u5c0f\u578b\u5f00\u53d1\u56e2\u961f\u90e8\u7f72\u4e00\u4e2aAI\u52a9\u624b，\u9884\u7b97\u6709\u9650，\u670d\u52a1\u5668\u914d\u7f6e\u4e5f\u4e0d\u7b97\u9ad8（32GB\u5185\u5b58，\u6ca1\u6709\u9ad8\u7aefGPU）\u3002\u8bd5\u8fc7\u51e0\u4e2a\u5927\u6a21\u578b，\u8981\u4e48\u8dd1\u4e0d\u8d77\u6765，\u8981\u4e48\u54cd\u5e94\u6162\u5f97\u50cf\u8717\u725b\u3002\u76f4\u5230\u6211\u9047\u5230\u4e86Gemma-3-12B-IT，\u60c5\u51b5\u624d\u5f7b\u5e95\u6539\u53d8\u3002<\/p>\n
\u8fd9\u4e2a\u6a21\u578b\u6709\u4ec0\u4e48\u7279\u522b\u4e4b\u5904？\u7b80\u5355\u6765\u8bf4，\u5b83\u662fGoogle\u6700\u65b0\u4e00\u4ee3\u7684\u8f7b\u91cf\u7ea7\u5f00\u6e90\u6a21\u578b，\u62e5\u6709120\u4ebf\u53c2\u6570\u2014\u2014\u8fd9\u4e2a\u89c4\u6a21\u6b63\u597d\u5361\u5728\u201c\u80fd\u529b\u8db3\u591f\u5f3a\u201d\u548c\u201c\u8d44\u6e90\u6d88\u8017\u53ef\u63a5\u53d7\u201d\u7684\u751c\u871c\u70b9\u4e0a\u3002\u76f8\u6bd4\u52a8\u8f84\u51e0\u767e\u4ebf\u53c2\u6570\u7684\u5927\u6a21\u578b，\u5b83\u80fd\u5728\u666e\u901a\u670d\u52a1\u5668\u4e0a\u987a\u7545\u8fd0\u884c；\u76f8\u6bd4\u66f4\u5c0f\u7684\u6a21\u578b，\u5b83\u7684\u7406\u89e3\u80fd\u529b\u548c\u751f\u6210\u8d28\u91cf\u53c8\u660e\u663e\u9ad8\u51fa\u4e00\u622a\u3002<\/p>\n
\u66f4\u91cd\u8981\u7684\u662f，\u8fd9\u662f\u201c\u6307\u4ee4\u5fae\u8c03\u201d\u7248\u672c（IT\u4ee3\u8868Instruction Tuned）\u3002\u8fd9\u610f\u5473\u7740\u5b83\u4e13\u95e8\u9488\u5bf9\u4eba\u7c7b\u5bf9\u8bdd\u8fdb\u884c\u4e86\u4f18\u5316，\u4f60\u95ee\u5b83\u95ee\u9898，\u5b83\u4e0d\u4f1a\u7ed9\u4f60\u4e00\u5806\u6280\u672f\u672f\u8bed，\u800c\u662f\u50cf\u670b\u53cb\u804a\u5929\u4e00\u6837\u7ed9\u51fa\u5b9e\u7528\u7684\u56de\u7b54\u3002\u5199\u4ee3\u7801\u3001\u89e3\u91ca\u6982\u5ff5\u3001\u521b\u4f5c\u5185\u5bb9，\u6837\u6837\u90fd\u80fd\u5e72\u3002<\/p>\n
\u4f46\u95ee\u9898\u6765\u4e86：\u600e\u4e48\u5728\u8d44\u6e90\u6709\u9650\u7684\u670d\u52a1\u5668\u4e0a，\u8ba9\u8fd9\u4e2a\u6a21\u578b\u8dd1\u5f97\u53c8\u5feb\u53c8\u7a33？\u8fd9\u5c31\u662f\u6211\u4eca\u5929\u8981\u5206\u4eab\u7684\u6838\u5fc3\u5185\u5bb9\u3002\u6211\u4f1a\u5e26\u4f60\u4e00\u6b65\u6b65\u5b8c\u6210\u90e8\u7f72，\u5e76\u5206\u4eab\u6211\u6478\u7d22\u51fa\u6765\u7684\u4f18\u5316\u6280\u5de7，\u8ba9\u4f60\u7528\u6700\u5c11\u7684\u8d44\u6e90，\u83b7\u5f97\u6700\u597d\u7684\u4f53\u9a8c\u3002<\/p>\n
\u5728\u5f00\u59cb\u5b89\u88c5\u4e4b\u524d，\u505a\u597d\u51c6\u5907\u5de5\u4f5c\u80fd\u8ba9\u4f60\u5c11\u8d70\u5f88\u591a\u5f2f\u8def\u3002\u5f88\u591a\u4eba\u4e00\u4e0a\u6765\u5c31\u4e0b\u8f7d\u6a21\u578b\u3001\u8fd0\u884c\u547d\u4ee4，\u7ed3\u679c\u9047\u5230\u5404\u79cd\u95ee\u9898，\u6298\u817e\u534a\u5929\u53c8\u5f97\u4ece\u5934\u518d\u6765\u3002<\/p>\n
\u9996\u5148，\u770b\u770b\u4f60\u7684\u670d\u52a1\u5668\u591f\u4e0d\u591f\u683c\u3002Gemma-3-12B-IT\u5bf9\u786c\u4ef6\u7684\u8981\u6c42\u6bd4\u8f83\u53cb\u597d，\u4f46\u4e5f\u4e0d\u662f\u4ec0\u4e48\u673a\u5668\u90fd\u80fd\u8dd1\u3002<\/p>\n
\u6700\u4f4e\u914d\u7f6e\u8981\u6c42：<\/p>\n
\u5982\u679c\u6709GPU\u4f1a\u66f4\u597d：<\/p>\n
\u6211\u7528\u7684\u6d4b\u8bd5\u73af\u5883\u662f\u4e00\u53f0\u4e91\u670d\u52a1\u5668：4\u6838CPU，32GB\u5185\u5b58，\u6ca1\u6709\u72ec\u7acbGPU\u3002\u8fd9\u4e2a\u914d\u7f6e\u8dd1\u8d77\u6765\u5b8c\u5168\u6ca1\u95ee\u9898，\u54cd\u5e94\u901f\u5ea6\u5728\u53ef\u63a5\u53d7\u8303\u56f4\u5185（\u4e00\u822c\u95ee\u98983-5\u79d2\u5185\u56de\u590d）\u3002\u5982\u679c\u4f60\u6709GPU，\u901f\u5ea6\u4f1a\u5feb\u5f88\u591a，\u4f46\u6ca1\u6709GPU\u4e5f\u80fd\u7528\u3002<\/p>\n
\u786e\u4fdd\u4f60\u7684\u7cfb\u7edf\u662f\u5e72\u51c0\u7684，\u6216\u8005\u81f3\u5c11\u6ca1\u6709\u51b2\u7a81\u7684Python\u73af\u5883\u3002\u6211\u63a8\u8350\u4f7f\u7528Ubuntu 20.04\u621622.04，\u8fd9\u662f\u6700\u7a33\u5b9a\u7684\u9009\u62e9\u3002<\/p>\n
# \u66f4\u65b0\u7cfb\u7edf\u5305
\nsudo apt update
\nsudo apt upgrade -y<\/p>\n
# \u5b89\u88c5\u57fa\u7840\u4f9d\u8d56
\nsudo apt install -y python3-pip python3-venv git curl wget<\/p>\n
# \u68c0\u67e5Python\u7248\u672c（\u9700\u89813.11\u6216\u4ee5\u4e0a）
\npython3 –version<\/p>\n
\u5982\u679cPython\u7248\u672c\u4f4e\u4e8e3.11，\u4f60\u9700\u8981\u5148\u5347\u7ea7\u3002\u4e0d\u8fc7\u5927\u591a\u6570\u73b0\u4ee3Ubuntu\u7cfb\u7edf\u90fd\u81ea\u5e263.10\u62163.11，\u4e00\u822c\u6ca1\u95ee\u9898\u3002<\/p>\n
\u8fd9\u662f\u975e\u5e38\u91cd\u8981\u7684\u4e00\u6b65！\u4e0d\u8981\u5728\u7cfb\u7edfPython\u91cc\u76f4\u63a5\u5b89\u88c5，\u4e00\u5b9a\u8981\u521b\u5efa\u865a\u62df\u73af\u5883\u3002\u8fd9\u6837\u65e2\u80fd\u907f\u514d\u4f9d\u8d56\u51b2\u7a81，\u4e5f\u65b9\u4fbf\u540e\u7eed\u7ba1\u7406\u3002<\/p>\n
# \u521b\u5efa\u9879\u76ee\u76ee\u5f55
\nmkdir -p ~\/gemma-3-webui
\ncd ~\/gemma-3-webui<\/p>\n
# \u521b\u5efaPython\u865a\u62df\u73af\u5883
\npython3 -m venv venv<\/p>\n
# \u6fc0\u6d3b\u865a\u62df\u73af\u5883
\nsource venv\/bin\/activate<\/p>\n
# \u4f60\u4f1a\u770b\u5230\u547d\u4ee4\u884c\u524d\u9762\u51fa\u73b0 (venv)，\u8868\u793a\u6fc0\u6d3b\u6210\u529f<\/p>\n
\u6fc0\u6d3b\u865a\u62df\u73af\u5883\u540e，\u6240\u6709Python\u5305\u7684\u5b89\u88c5\u90fd\u4f1a\u5c40\u9650\u5728\u8fd9\u4e2a\u73af\u5883\u91cc，\u4e0d\u4f1a\u5f71\u54cd\u7cfb\u7edf\u5176\u4ed6\u90e8\u5206\u3002\u9000\u51fa\u65f6\u7528 deactivate \u547d\u4ee4\u3002<\/p>\n
\u73b0\u5728\u8fdb\u5165\u6b63\u9898，\u6211\u4eec\u6765\u5b9e\u9645\u90e8\u7f72\u8fd9\u4e2a\u6a21\u578b\u7684Web\u754c\u9762\u3002\u6211\u63d0\u4f9b\u7684\u8fd9\u4e2a\u65b9\u6cd5\u7ecf\u8fc7\u591a\u6b21\u6d4b\u8bd5，\u662f\u6700\u7a33\u5b9a\u3001\u6700\u7701\u4e8b\u7684\u65b9\u6848\u3002<\/p>\n
\u6211\u4e3a\u4f60\u51c6\u5907\u4e86\u4e00\u4e2a\u5b8c\u6574\u7684\u90e8\u7f72\u811a\u672c，\u5305\u542b\u4e86\u6240\u6709\u5fc5\u8981\u7684\u6b65\u9aa4\u3002\u628a\u8fd9\u4e2a\u811a\u672c\u4fdd\u5b58\u5230\u4f60\u7684\u670d\u52a1\u5668\u4e0a：<\/p>\n
# \u521b\u5efa\u90e8\u7f72\u811a\u672c
\ncat > deploy_gemma.sh << 'EOF'
\n#!\/bin\/bash<\/p>\n
echo "=== Gemma-3-12B-IT WebUI \u90e8\u7f72\u811a\u672c ==="
\necho "\u5f00\u59cb\u65f6\u95f4: $(date)"<\/p>\n
# \u68c0\u67e5\u662f\u5426\u5728\u865a\u62df\u73af\u5883\u4e2d
\nif [ -z "$VIRTUAL_ENV" ]; then
\n echo "\u9519\u8bef：\u8bf7\u5148\u6fc0\u6d3bPython\u865a\u62df\u73af\u5883"
\n echo "\u6267\u884c: source venv\/bin\/activate"
\n exit 1
\nfi<\/p>\n
# \u6b65\u9aa41：\u5b89\u88c5PyTorch（\u6839\u636e\u662f\u5426\u6709GPU\u9009\u62e9）
\necho "\u6b65\u9aa41：\u5b89\u88c5PyTorch…"
\nif command -v nvidia-smi &> \/dev\/null; then
\n echo "\u68c0\u6d4b\u5230GPU，\u5b89\u88c5CUDA\u7248\u672c\u7684PyTorch"
\n pip3 install torch torchvision torchaudio –index-url https:\/\/download.pytorch.org\/whl\/cu118
\nelse
\n echo "\u672a\u68c0\u6d4b\u5230GPU，\u5b89\u88c5CPU\u7248\u672c\u7684PyTorch"
\n pip3 install torch torchvision torchaudio –index-url https:\/\/download.pytorch.org\/whl\/cpu
\nfi<\/p>\n
# \u6b65\u9aa42：\u5b89\u88c5WebUI\u4f9d\u8d56
\necho "\u6b65\u9aa42：\u5b89\u88c5WebUI\u4f9d\u8d56…"
\npip3 install gradio==4.19.0
\npip3 install transformers==4.36.0
\npip3 install accelerate==0.25.0
\npip3 install sentencepiece==0.1.99
\npip3 install protobuf==3.20.3<\/p>\n
# \u6b65\u9aa43：\u4e0b\u8f7d\u6a21\u578b（\u5982\u679c\u672c\u5730\u6ca1\u6709）
\necho "\u6b65\u9aa43：\u68c0\u67e5\u6a21\u578b\u6587\u4ef6…"
\nMODEL_PATH="\/root\/ai-models\/LLM-Research\/gemma-3-12b-it"
\nif [ ! -d "$MODEL_PATH" ]; then
\n echo "\u6a21\u578b\u76ee\u5f55\u4e0d\u5b58\u5728，\u8bf7\u786e\u4fdd\u6a21\u578b\u5df2\u6b63\u786e\u653e\u7f6e"
\n echo "\u6a21\u578b\u5e94\u8be5\u4f4d\u4e8e: $MODEL_PATH"
\n echo "\u5982\u679c\u9700\u8981\u4e0b\u8f7d，\u8bf7\u53c2\u8003\u5b98\u65b9\u6587\u6863"
\n exit 1
\nelse
\n echo "\u6a21\u578b\u76ee\u5f55\u5b58\u5728: $MODEL_PATH"
\nfi<\/p>\n
# \u6b65\u9aa44：\u521b\u5efaWebUI\u5e94\u7528\u6587\u4ef6
\necho "\u6b65\u9aa44：\u521b\u5efaWebUI\u5e94\u7528…"
\ncat > app.py << 'PYEOF'
\nimport gradio as gr
\nfrom transformers import AutoTokenizer, AutoModelForCausalLM
\nimport torch
\nimport time<\/p>\n
# \u6a21\u578b\u8def\u5f84
\nMODEL_PATH = "\/root\/ai-models\/LLM-Research\/gemma-3-12b-it"<\/p>\n
# \u52a0\u8f7d\u6a21\u578b\u548c\u5206\u8bcd\u5668
\nprint("\u6b63\u5728\u52a0\u8f7d\u6a21\u578b，\u8fd9\u53ef\u80fd\u9700\u8981\u51e0\u5206\u949f…")
\nstart_time = time.time()<\/p>\n
tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH)
\nmodel = AutoModelForCausalLM.from_pretrained(
\n MODEL_PATH,
\n torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
\n device_map="auto",
\n low_cpu_mem_usage=True
\n)<\/p>\n
print(f"\u6a21\u578b\u52a0\u8f7d\u5b8c\u6210，\u8017\u65f6: {time.time() – start_time:.2f}\u79d2")<\/p>\n
def generate_response(message, history, temperature=0.7, max_tokens=512):
\n """\u751f\u6210\u56de\u590d"""
\n try:
\n # \u6784\u5efa\u5bf9\u8bdd\u5386\u53f2
\n prompt = ""
\n if history:
\n for user_msg, assistant_msg in history:
\n prompt += f"\u7528\u6237: {user_msg}\\\\n\u52a9\u624b: {assistant_msg}\\\\n"
\n prompt += f"\u7528\u6237: {message}\\\\n\u52a9\u624b: "<\/p>\n
# \u7f16\u7801\u8f93\u5165
\n inputs = tokenizer(prompt, return_tensors="pt", truncation=True, max_length=2048)<\/p>\n
# \u79fb\u52a8\u5230\u6b63\u786e\u7684\u8bbe\u5907
\n if torch.cuda.is_available():
\n inputs = {k: v.cuda() for k, v in inputs.items()}<\/p>\n
# \u751f\u6210\u56de\u590d
\n with torch.no_grad():
\n outputs = model.generate(
\n **inputs,
\n max_new_tokens=max_tokens,
\n temperature=temperature,
\n top_p=0.9,
\n do_sample=True,
\n pad_token_id=tokenizer.eos_token_id
\n )<\/p>\n
# \u89e3\u7801\u56de\u590d
\n response = tokenizer.decode(outputs[0][inputs['input_ids'].shape[1]:], skip_special_tokens=True)<\/p>\n
return response.strip()<\/p>\n
except Exception as e:
\n return f"\u751f\u6210\u56de\u590d\u65f6\u51fa\u9519: {str(e)}"<\/p>\n
# \u521b\u5efaGradio\u754c\u9762
\nwith gr.Blocks(title="Gemma-3-12B-IT \u804a\u5929\u52a9\u624b", theme=gr.themes.Soft()) as demo:
\n gr.Markdown("# 🤖 Gemma-3-12B-IT \u804a\u5929\u52a9\u624b")
\n gr.Markdown("Google\u6700\u65b0120\u4ebf\u53c2\u6570\u6307\u4ee4\u5fae\u8c03\u6a21\u578b，\u652f\u6301\u4ee3\u7801\u751f\u6210\u3001\u77e5\u8bc6\u95ee\u7b54\u3001\u6587\u672c\u521b\u4f5c\u7b49")<\/p>\n
# \u804a\u5929\u754c\u9762
\n chatbot = gr.Chatbot(height=500, label="\u5bf9\u8bdd\u5386\u53f2")<\/p>\n
with gr.Row():
\n msg = gr.Textbox(
\n label="\u8f93\u5165\u4f60\u7684\u95ee\u9898",
\n placeholder="\u4f8b\u5982：\u5199\u4e00\u4e2aPython\u5feb\u901f\u6392\u5e8f\u51fd\u6570…",
\n scale=4
\n )
\n submit_btn = gr.Button("\u53d1\u9001", variant="primary", scale=1)<\/p>\n
with gr.Row():
\n clear_btn = gr.Button("\u6e05\u7a7a\u5bf9\u8bdd", variant="secondary")<\/p>\n
with gr.Accordion("\u9ad8\u7ea7\u53c2\u6570", open=False):
\n temperature = gr.Slider(
\n minimum=0.1,
\n maximum=1.5,
\n value=0.7,
\n step=0.1,
\n label="Temperature（\u968f\u673a\u6027）",
\n info="\u503c\u8d8a\u9ad8\u56de\u7b54\u8d8a\u6709\u521b\u610f，\u503c\u8d8a\u4f4e\u56de\u7b54\u8d8a\u4e25\u8c28"
\n )
\n max_tokens = gr.Slider(
\n minimum=64,
\n maximum=2048,
\n value=512,
\n step=64,
\n label="\u6700\u5927\u751f\u6210\u957f\u5ea6",
\n info="\u63a7\u5236\u56de\u590d\u7684\u6700\u5927\u957f\u5ea6"
\n )<\/p>\n
# \u793a\u4f8b\u95ee\u9898
\n gr.Examples(
\n examples=[
\n ["\u7528Python\u5199\u4e00\u4e2a\u5feb\u901f\u6392\u5e8f\u7b97\u6cd5"],
\n ["\u89e3\u91ca\u4e00\u4e0b\u4ec0\u4e48\u662f\u673a\u5668\u5b66\u4e60"],
\n ["\u5e2e\u6211\u5199\u4e00\u5c01\u5de5\u4f5c\u6c47\u62a5\u90ae\u4ef6"],
\n ["JavaScript\u548cPython\u6709\u4ec0\u4e48\u533a\u522b？"],
\n ["\u5199\u4e00\u4e2a\u5173\u4e8e\u4eba\u5de5\u667a\u80fd\u7684\u77ed\u6545\u4e8b"]
\n ],
\n inputs=msg,
\n label="\u8bd5\u8bd5\u8fd9\u4e9b\u95ee\u9898"
\n )<\/p>\n
# \u4e8b\u4ef6\u5904\u7406
\n def respond(message, chat_history, temp, tokens):
\n bot_message = 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_btn.click(respond, [msg, chatbot, temperature, max_tokens], [msg, chatbot])
\n clear_btn.click(lambda: None, None, chatbot, queue=False)<\/p>\n
if __name__ == "__main__":
\n demo.launch(
\n server_name="0.0.0.0",
\n server_port=7860,
\n share=False,
\n show_error=True
\n )
\nPYEOF<\/p>\n
echo "\u6b65\u9aa45：\u521b\u5efa\u7ba1\u7406\u811a\u672c…"
\ncat > manage.sh << 'MGEOF'
\n#!\/bin\/bash<\/p>\n
SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
\nLOG_DIR="$SCRIPT_DIR\/logs"
\nPID_FILE="$SCRIPT_DIR\/app.pid"<\/p>\n
# \u521b\u5efa\u65e5\u5fd7\u76ee\u5f55
\nmkdir -p "$LOG_DIR"<\/p>\n
start() {
\n echo "\u542f\u52a8 Gemma-3 WebUI \u670d\u52a1…"
\n cd "$SCRIPT_DIR"
\n source venv\/bin\/activate
\n nohup python app.py > "$LOG_DIR\/app.log" 2>&1 &
\n echo $! > "$PID_FILE"
\n echo "\u670d\u52a1\u5df2\u542f\u52a8，PID: $(cat $PID_FILE)"
\n echo "Web\u754c\u9762: http:\/\/$(hostname -I | awk '{print $1}'):7860"
\n}<\/p>\n
stop() {
\n if [ -f "$PID_FILE" ]; then
\n PID=$(cat "$PID_FILE")
\n echo "\u505c\u6b62\u670d\u52a1 (PID: $PID)…"
\n kill $PID 2>\/dev\/null
\n rm -f "$PID_FILE"
\n echo "\u670d\u52a1\u5df2\u505c\u6b62"
\n else
\n echo "\u670d\u52a1\u672a\u8fd0\u884c"
\n fi
\n}<\/p>\n
restart() {
\n stop
\n sleep 2
\n start
\n}<\/p>\n
status() {
\n if [ -f "$PID_FILE" ]; then
\n PID=$(cat "$PID_FILE")
\n if ps -p $PID > \/dev\/null; then
\n echo "\u670d\u52a1\u6b63\u5728\u8fd0\u884c (PID: $PID)"
\n echo "\u8bbf\u95ee\u5730\u5740: http:\/\/$(hostname -I | awk '{print $1}'):7860"
\n else
\n echo "PID\u6587\u4ef6\u5b58\u5728\u4f46\u8fdb\u7a0b\u672a\u8fd0\u884c"
\n rm -f "$PID_FILE"
\n fi
\n else
\n echo "\u670d\u52a1\u672a\u8fd0\u884c"
\n fi
\n}<\/p>\n
logs() {
\n if [ -f "$LOG_DIR\/app.log" ]; then
\n tail -f "$LOG_DIR\/app.log"
\n else
\n echo "\u65e5\u5fd7\u6587\u4ef6\u4e0d\u5b58\u5728"
\n fi
\n}<\/p>\n
case "$1" in
\n start)
\n start
\n ;;
\n stop)
\n stop
\n ;;
\n restart)
\n restart
\n ;;
\n status)
\n status
\n ;;
\n logs)
\n logs
\n ;;
\n *)
\n echo "\u4f7f\u7528\u65b9\u6cd5: $0 {start|stop|restart|status|logs}"
\n exit 1
\n ;;
\nesac
\nMGEOF<\/p>\n
chmod +x manage.sh<\/p>\n
echo "\u6b65\u9aa46：\u521b\u5efaSupervisord\u914d\u7f6e（\u53ef\u9009）…"
\ncat > gemma-webui.conf << 'SUPEOF'
\n[program:gemma-webui]
\ndirectory=\/root\/gemma-3-webui
\ncommand=\/root\/gemma-3-webui\/venv\/bin\/python app.py
\nautostart=true
\nautorestart=true
\nstartretries=3
\nuser=root
\nredirect_stderr=true
\nstdout_logfile=\/root\/gemma-3-webui\/logs\/supervisor.log
\nstdout_logfile_maxbytes=10MB
\nstdout_logfile_backups=5
\nenvironment=PYTHONUNBUFFERED="1"
\nSUPEOF<\/p>\n
echo "=== \u90e8\u7f72\u5b8c\u6210 ==="
\necho ""
\necho "\u4e0b\u4e00\u6b65\u64cd\u4f5c："
\necho "1. \u542f\u52a8\u670d\u52a1: .\/manage.sh start"
\necho "2. \u67e5\u770b\u72b6\u6001: .\/manage.sh status"
\necho "3. \u8bbf\u95eeWeb\u754c\u9762: http:\/\/\u4f60\u7684\u670d\u52a1\u5668IP:7860"
\necho ""
\necho "\u5982\u679c\u9700\u8981\u5f00\u673a\u81ea\u542f，\u53ef\u4ee5\u5c06\u4ee5\u4e0b\u547d\u4ee4\u6dfb\u52a0\u5230 \/etc\/rc.local:"
\necho "cd \/root\/gemma-3-webui && .\/manage.sh start"
\nEOF<\/p>\n
# \u7ed9\u811a\u672c\u6267\u884c\u6743\u9650
\nchmod +x deploy_gemma.sh<\/p>\n
\u8fd9\u4e2a\u811a\u672c\u505a\u4e86\u51e0\u4ef6\u91cd\u8981\u7684\u4e8b\u60c5：<\/p>\n
\u73b0\u5728\u8fd0\u884c\u90e8\u7f72\u811a\u672c：<\/p>\n
# \u786e\u4fdd\u5728\u865a\u62df\u73af\u5883\u4e2d
\nsource venv\/bin\/activate<\/p>\n
# \u6267\u884c\u90e8\u7f72\u811a\u672c
\n.\/deploy_gemma.sh<\/p>\n
\u811a\u672c\u8fd0\u884c\u8fc7\u7a0b\u4e2d，\u4f60\u4f1a\u770b\u5230\u6bcf\u4e00\u6b65\u7684\u8fdb\u5ea6\u3002\u6700\u8017\u65f6\u7684\u90e8\u5206\u662f\u5b89\u88c5PyTorch\u548c\u52a0\u8f7d\u6a21\u578b，\u53ef\u80fd\u9700\u8981\u51e0\u5206\u949f\u5230\u5341\u51e0\u5206\u949f，\u53d6\u51b3\u4e8e\u4f60\u7684\u7f51\u7edc\u548c\u670d\u52a1\u5668\u6027\u80fd\u3002<\/p>\n
\u90e8\u7f72\u5b8c\u6210\u540e，\u542f\u52a8\u670d\u52a1\u5f88\u7b80\u5355：<\/p>\n
# \u542f\u52a8\u670d\u52a1
\n.\/manage.sh start<\/p>\n
# \u67e5\u770b\u72b6\u6001
\n.\/manage.sh status<\/p>\n
\u5982\u679c\u4e00\u5207\u6b63\u5e38，\u4f60\u4f1a\u770b\u5230\u7c7b\u4f3c\u8fd9\u6837\u7684\u8f93\u51fa：<\/p>\n
\u670d\u52a1\u6b63\u5728\u8fd0\u884c (PID: 12345)
\n\u8bbf\u95ee\u5730\u5740: http:\/\/192.168.1.100:7860<\/p>\n
\u73b0\u5728\u6253\u5f00\u6d4f\u89c8\u5668，\u8bbf\u95ee\u663e\u793a\u7684\u5730\u5740，\u5c31\u80fd\u770b\u5230Gemma-3\u7684\u804a\u5929\u754c\u9762\u4e86！<\/p>\n
\u90e8\u7f72\u53ea\u662f\u7b2c\u4e00\u6b65，\u771f\u6b63\u7684\u6311\u6218\u662f\u8ba9\u6a21\u578b\u5728\u6709\u9650\u8d44\u6e90\u4e0b\u8868\u73b0\u6700\u4f73\u3002\u4e0b\u9762\u662f\u6211\u603b\u7ed3\u7684\u51e0\u4e2a\u5173\u952e\u4f18\u5316\u6280\u5de7\u3002<\/p>\n
120\u4ebf\u53c2\u6570\u7684\u6a21\u578b\u542c\u8d77\u6765\u5f88\u5927，\u4f46\u901a\u8fc7\u4e00\u4e9b\u6280\u5de7，\u6211\u4eec\u53ef\u4ee5\u572832GB\u5185\u5b58\u7684\u670d\u52a1\u5668\u4e0a\u6d41\u7545\u8fd0\u884c\u3002<\/p>\n
\u6280\u5de71：\u4f7f\u7528\u534a\u7cbe\u5ea6\u6d6e\u70b9\u6570<\/p>\n
\u8fd9\u662f\u6700\u91cd\u8981\u7684\u4f18\u5316！\u9ed8\u8ba4\u60c5\u51b5\u4e0b，PyTorch\u4f7f\u7528float32（\u5355\u7cbe\u5ea6），\u6bcf\u4e2a\u53c2\u6570\u53604\u5b57\u8282\u3002120\u4ebf\u53c2\u6570\u5c31\u9700\u898148GB\u5185\u5b58，\u8fd9\u663e\u7136\u4e0d\u884c\u3002<\/p>\n
# \u4f18\u5316\u524d：\u4f7f\u7528float32
\nmodel = AutoModelForCausalLM.from_pretrained(MODEL_PATH)<\/p>\n
# \u4f18\u5316\u540e：\u4f7f\u7528float16（\u534a\u7cbe\u5ea6）
\nmodel = AutoModelForCausalLM.from_pretrained(
\n MODEL_PATH,
\n torch_dtype=torch.float16, # \u5173\u952e\u53c2\u6570！
\n device_map="auto"
\n)<\/p>\n
\u4f7f\u7528float16\u540e，\u6bcf\u4e2a\u53c2\u6570\u53ea\u53602\u5b57\u8282，\u5185\u5b58\u9700\u6c42\u76f4\u63a5\u51cf\u534a，\u53d8\u621024GB\u3002\u5982\u679c\u4f60\u7684GPU\u652f\u6301bfloat16，\u6548\u679c\u4f1a\u66f4\u597d\u3002<\/p>\n
\u6280\u5de72：\u542f\u7528\u4f4eCPU\u5185\u5b58\u4f7f\u7528\u6a21\u5f0f<\/p>\n
model = AutoModelForCausalLM.from_pretrained(
\n MODEL_PATH,
\n torch_dtype=torch.float16,
\n device_map="auto",
\n low_cpu_mem_usage=True # \u8fd9\u4e2a\u53c2\u6570\u5f88\u91cd\u8981！
\n)<\/p>\n
\u8fd9\u4e2a\u53c2\u6570\u544a\u8bc9transformers\u5e93\u5728\u52a0\u8f7d\u6a21\u578b\u65f6\u5c3d\u91cf\u51cf\u5c11CPU\u5185\u5b58\u7684\u4f7f\u7528，\u7279\u522b\u662f\u5728\u5185\u5b58\u6709\u9650\u7684\u7cfb\u7edf\u4e0a\u3002<\/p>\n
\u6280\u5de73：\u5206\u7247\u52a0\u8f7d（\u5982\u679c\u5185\u5b58\u771f\u7684\u5f88\u7d27\u5f20）<\/p>\n
\u5982\u679c\u4f60\u7684\u5185\u5b58\u5c0f\u4e8e24GB，\u53ef\u4ee5\u5c1d\u8bd5\u5206\u7247\u52a0\u8f7d：<\/p>\n
from transformers import AutoConfig<\/p>\n
# \u5148\u52a0\u8f7d\u914d\u7f6e
\nconfig = AutoConfig.from_pretrained(MODEL_PATH)<\/p>\n
# \u7136\u540e\u5206\u7247\u52a0\u8f7d\u6a21\u578b
\nmodel = AutoModelForCausalLM.from_pretrained(
\n MODEL_PATH,
\n torch_dtype=torch.float16,
\n device_map="auto",
\n low_cpu_mem_usage=True,
\n offload_folder="offload", # \u4e34\u65f6\u5378\u8f7d\u6587\u4ef6\u5939
\n offload_state_dict=True # \u5378\u8f7d\u72b6\u6001\u5b57\u5178
\n)<\/p>\n
\u6ca1\u6709\u4eba\u559c\u6b22\u7b49\u5f85，\u7279\u522b\u662f\u5728\u5bf9\u8bdd\u65f6\u3002\u4e0b\u9762\u8fd9\u4e9b\u6280\u5de7\u80fd\u663e\u8457\u63d0\u5347\u54cd\u5e94\u901f\u5ea6\u3002<\/p>\n
\u6280\u5de71：\u8c03\u6574\u751f\u6210\u53c2\u6570<\/p>\n
# \u751f\u6210\u56de\u590d\u65f6\u7684\u4f18\u5316\u53c2\u6570
\noutputs = model.generate(
\n **inputs,
\n max_new_tokens=512, # \u9650\u5236\u751f\u6210\u957f\u5ea6
\n temperature=0.7, # \u5e73\u8861\u521b\u610f\u548c\u51c6\u786e
\n top_p=0.9, # \u6838\u91c7\u6837，\u63d0\u9ad8\u8d28\u91cf
\n do_sample=True, # \u542f\u7528\u91c7\u6837
\n pad_token_id=tokenizer.eos_token_id,<\/p>\n
# \u901f\u5ea6\u4f18\u5316\u53c2\u6570
\n no_repeat_ngram_size=3, # \u907f\u514d\u91cd\u590d
\n repetition_penalty=1.1, # \u91cd\u590d\u60e9\u7f5a
\n length_penalty=1.0, # \u957f\u5ea6\u60e9\u7f5a
\n)<\/p>\n
\u6280\u5de72：\u4f7f\u7528KV\u7f13\u5b58（\u5982\u679c\u652f\u6301）<\/p>\n
KV\u7f13\u5b58\u80fd\u663e\u8457\u63d0\u5347\u8fde\u7eed\u5bf9\u8bdd\u7684\u901f\u5ea6：<\/p>\n
# \u7b2c\u4e00\u6b21\u751f\u6210
\noutputs = model.generate(**inputs, use_cache=True)
\npast_key_values = outputs.past_key_values<\/p>\n
# \u540e\u7eed\u751f\u6210\u65f6\u590d\u7528\u7f13\u5b58
\nnext_outputs = model.generate(
\n next_inputs,
\n past_key_values=past_key_values,
\n use_cache=True
\n)<\/p>\n
\u6280\u5de73：\u6279\u5904\u7406\u8bf7\u6c42<\/p>\n
\u5982\u679c\u4f60\u9700\u8981\u5904\u7406\u591a\u4e2a\u8bf7\u6c42，\u5c3d\u91cf\u6279\u5904\u7406：<\/p>\n
# \u5355\u4e2a\u8bf7\u6c42
\nresponse1 = generate_response("\u95ee\u98981")
\nresponse2 = generate_response("\u95ee\u98982")<\/p>\n
# \u6279\u5904\u7406（\u5982\u679c\u903b\u8f91\u5141\u8bb8）
\nresponses = batch_generate(["\u95ee\u98981", "\u95ee\u98982"])<\/p>\n
\u901f\u5ea6\u5f88\u91cd\u8981，\u4f46\u8d28\u91cf\u66f4\u91cd\u8981\u3002\u8fd9\u4e9b\u6280\u5de7\u80fd\u8ba9Gemma-3\u7684\u56de\u7b54\u66f4\u7b26\u5408\u4f60\u7684\u671f\u671b\u3002<\/p>\n
\u6280\u5de71：\u6e29\u5ea6\u53c2\u6570\u8c03\u4f18<\/p>\n
\u6e29\u5ea6（Temperature）\u63a7\u5236\u968f\u673a\u6027，\u4e0d\u540c\u4efb\u52a1\u9700\u8981\u4e0d\u540c\u7684\u8bbe\u7f6e：<\/p>\n
# \u4e0d\u540c\u4efb\u52a1\u7684\u6700\u4f73\u6e29\u5ea6\u8bbe\u7f6e
\nTASK_SETTINGS = {
\n "\u4ee3\u7801\u751f\u6210": 0.2, # \u4f4e\u6e29\u5ea6，\u786e\u4fdd\u4ee3\u7801\u51c6\u786e
\n "\u77e5\u8bc6\u95ee\u7b54": 0.7, # \u4e2d\u7b49\u6e29\u5ea6，\u5e73\u8861\u51c6\u786e\u548c\u6d41\u7545
\n "\u521b\u610f\u5199\u4f5c": 1.0, # \u9ad8\u6e29\u5ea6，\u66f4\u6709\u521b\u610f
\n "\u7ffb\u8bd1\u4efb\u52a1": 0.5, # \u4f4e\u6e29\u5ea6，\u4fdd\u6301\u539f\u610f
\n "\u603b\u7ed3\u5f52\u7eb3": 0.6, # \u4e2d\u7b49\u504f\u4f4e，\u786e\u4fdd\u5173\u952e\u4fe1\u606f
\n}<\/p>\n
def get_optimal_temperature(task_type):
\n return TASK_SETTINGS.get(task_type, 0.7)<\/p>\n
\u6280\u5de72：\u66f4\u597d\u7684\u63d0\u793a\u8bcd\u5de5\u7a0b<\/p>\n
Gemma-3\u5bf9\u63d0\u793a\u8bcd\u5f88\u654f\u611f，\u597d\u7684\u63d0\u793a\u8bcd\u80fd\u5927\u5e45\u63d0\u5347\u56de\u7b54\u8d28\u91cf：<\/p>\n
# \u4e0d\u597d\u7684\u63d0\u793a\u8bcd
\nprompt = "\u5199\u4ee3\u7801"<\/p>\n
# \u597d\u7684\u63d0\u793a\u8bcd
\ngood_prompt = """\u8bf7\u7528Python\u5199\u4e00\u4e2a\u5feb\u901f\u6392\u5e8f\u51fd\u6570，\u8981\u6c42：
\n1. \u5305\u542b\u8be6\u7ec6\u7684\u6ce8\u91ca
\n2. \u5904\u7406\u7a7a\u5217\u8868\u548c\u5355\u5143\u7d20\u5217\u8868\u7684\u60c5\u51b5
\n3. \u6dfb\u52a0\u7c7b\u578b\u63d0\u793a
\n4. \u5305\u542b\u4f7f\u7528\u793a\u4f8b<\/p>\n
\u8bf7\u786e\u4fdd\u4ee3\u7801\u7b26\u5408PEP8\u89c4\u8303\u3002"""<\/p>\n
\u6280\u5de73：\u7cfb\u7edf\u6307\u4ee4\u8bbe\u7f6e<\/p>\n
\u867d\u7136Gemma-3\u4e0d\u662f\u804a\u5929\u6a21\u578b，\u4f46\u6211\u4eec\u53ef\u4ee5\u901a\u8fc7\u7cfb\u7edf\u6307\u4ee4\u6765\u5f15\u5bfc\u5b83：<\/p>\n
system_instruction = """\u4f60\u662f\u4e00\u4e2a\u4e13\u4e1a\u7684\u7f16\u7a0b\u52a9\u624b，\u64c5\u957fPython\u3001JavaScript\u548cJava\u3002
\n\u4f60\u7684\u56de\u7b54\u5e94\u8be5：
\n1. \u51c6\u786e\u4e14\u5b9e\u7528
\n2. \u5305\u542b\u4ee3\u7801\u793a\u4f8b
\n3. \u89e3\u91ca\u5173\u952e\u6982\u5ff5
\n4. \u7ed9\u51fa\u6700\u4f73\u5b9e\u8df5\u5efa\u8bae<\/p>\n
\u5982\u679c\u7528\u6237\u7684\u95ee\u9898\u4e0d\u6e05\u695a，\u8bf7\u793c\u8c8c\u5730\u8bf7\u6c42\u6f84\u6e05\u3002"""<\/p>\n
prompt = f"{system_instruction}\\\\n\\\\n\u7528\u6237: {user_message}\\\\n\u52a9\u624b: "<\/p>\n
\u5bf9\u4e8e\u751f\u4ea7\u73af\u5883，\u7a33\u5b9a\u6027\u81f3\u5173\u91cd\u8981\u3002<\/p>\n
\u6280\u5de71：\u6dfb\u52a0\u8d85\u65f6\u548c\u91cd\u8bd5\u673a\u5236<\/p>\n
import signal
\nfrom functools import wraps<\/p>\n
def timeout(seconds=30):
\n """\u8d85\u65f6\u88c5\u9970\u5668"""
\n def decorator(func):
\n @wraps(func)
\n def wrapper(*args, **kwargs):
\n def handler(signum, frame):
\n raise TimeoutError("\u751f\u6210\u8d85\u65f6")<\/p>\n
# \u8bbe\u7f6e\u8d85\u65f6
\n signal.signal(signal.SIGALRM, handler)
\n signal.alarm(seconds)<\/p>\n
try:
\n result = func(*args, **kwargs)
\n finally:
\n signal.alarm(0) # \u53d6\u6d88\u8d85\u65f6<\/p>\n
return result
\n return wrapper
\n return decorator<\/p>\n
@timeout(30)
\ndef safe_generate_response(message, history):
\n return generate_response(message, history)<\/p>\n
\u6280\u5de72：\u5185\u5b58\u76d1\u63a7\u548c\u6e05\u7406<\/p>\n
import gc
\nimport torch<\/p>\n
def clean_memory():
\n """\u6e05\u7406GPU\u548cCPU\u5185\u5b58"""
\n if torch.cuda.is_available():
\n torch.cuda.empty_cache()
\n torch.cuda.synchronize()<\/p>\n
gc.collect()<\/p>\n
# \u5728\u957f\u65f6\u95f4\u8fd0\u884c\u540e\u8c03\u7528
\nclean_memory()<\/p>\n
\u6280\u5de73：\u5b9e\u73b0\u5065\u5eb7\u68c0\u67e5\u63a5\u53e3<\/p>\n
# \u5728app.py\u4e2d\u6dfb\u52a0\u5065\u5eb7\u68c0\u67e5
\n@app.route('\/health')
\ndef health_check():
\n try:
\n # \u7b80\u5355\u6d4b\u8bd5\u6a21\u578b\u662f\u5426\u53ef\u7528
\n test_input = tokenizer("test", return_tensors="pt")
\n return jsonify({"status": "healthy", "model": "gemma-3-12b-it"})
\n except Exception as e:
\n return jsonify({"status": "unhealthy", "error": str(e)}), 500<\/p>\n
\u7406\u8bba\u8bf4\u518d\u591a，\u4e0d\u5982\u770b\u770b\u5b9e\u9645\u6548\u679c\u3002\u4e0b\u9762\u662f\u6211\u5728\u5b9e\u9645\u4f7f\u7528\u4e2d\u7684\u51e0\u4e2a\u6848\u4f8b\u3002<\/p>\n
\u6211\u7684\u9700\u6c42：\u9700\u8981\u4e00\u4e2aPython\u51fd\u6570，\u4eceAPI\u83b7\u53d6\u6570\u636e，\u5904\u7406\u5f02\u5e38，\u5e76\u8fd4\u56de\u7ed3\u6784\u5316\u7ed3\u679c\u3002<\/p>\n
\u6211\u7ed9Gemma-3\u7684\u63d0\u793a：<\/p>\n
\u5199\u4e00\u4e2aPython\u51fd\u6570，\u4ece\u6307\u5b9a\u7684API\u7aef\u70b9\u83b7\u53d6JSON\u6570\u636e\u3002
\n\u8981\u6c42：
\n1. \u6dfb\u52a0\u8bf7\u6c42\u8d85\u65f6\u548c\u91cd\u8bd5\u673a\u5236
\n2. \u5904\u7406HTTP\u9519\u8bef\u548cJSON\u89e3\u6790\u9519\u8bef
\n3. \u6dfb\u52a0\u7c7b\u578b\u63d0\u793a\u548c\u6587\u6863\u5b57\u7b26\u4e32
\n4. \u8fd4\u56de\u4e00\u4e2a\u5305\u542b\u72b6\u6001\u548c\u6570\u636e\u7684\u7ed3\u679c\u5b57\u5178<\/p>\n
Gemma-3\u751f\u6210\u7684\u4ee3\u7801：<\/p>\n
import requests
\nimport json
\nfrom typing import Dict, Any, Optional
\nfrom time import sleep<\/p>\n
def fetch_api_data(
\n url: str,
\n timeout: int = 10,
\n max_retries: int = 3,
\n retry_delay: int = 2
\n) -> Dict[str, Any]:
\n """
\n \u4eceAPI\u7aef\u70b9\u83b7\u53d6JSON\u6570\u636e，\u5305\u542b\u9519\u8bef\u5904\u7406\u548c\u91cd\u8bd5\u673a\u5236\u3002<\/p>\n
\u53c2\u6570:
\n url: API\u7aef\u70b9URL
\n timeout: \u8bf7\u6c42\u8d85\u65f6\u65f6\u95f4（\u79d2）
\n max_retries: \u6700\u5927\u91cd\u8bd5\u6b21\u6570
\n retry_delay: \u91cd\u8bd5\u5ef6\u8fdf（\u79d2）<\/p>\n
\u8fd4\u56de:
\n \u5305\u542b\u72b6\u6001\u548c\u6570\u636e\u7684\u7ed3\u679c\u5b57\u5178
\n {
\n "success": bool,
\n "data": Optional[Dict],
\n "error": Optional[str],
\n "status_code": Optional[int]
\n }
\n """
\n headers = {
\n "User-Agent": "Mozilla\/5.0",
\n "Accept": "application\/json"
\n }<\/p>\n
for attempt in range(max_retries):
\n try:
\n response = requests.get(
\n url,
\n headers=headers,
\n timeout=timeout
\n )
\n response.raise_for_status() # \u68c0\u67e5HTTP\u9519\u8bef<\/p>\n
# \u89e3\u6790JSON
\n data = response.json()<\/p>\n
return {
\n "success": True,
\n "data": data,
\n "error": None,
\n "status_code": response.status_code
\n }<\/p>\n
except requests.exceptions.Timeout:
\n error_msg = f"\u8bf7\u6c42\u8d85\u65f6 (\u5c1d\u8bd5 {attempt + 1}\/{max_retries})"
\n if attempt < max_retries – 1:
\n sleep(retry_delay)
\n continue<\/p>\n
except requests.exceptions.HTTPError as e:
\n return {
\n "success": False,
\n "data": None,
\n "error": f"HTTP\u9519\u8bef: {str(e)}",
\n "status_code": e.response.status_code if e.response else None
\n }<\/p>\n
except json.JSONDecodeError:
\n return {
\n "success": False,
\n "data": None,
\n "error": "JSON\u89e3\u6790\u5931\u8d25",
\n "status_code": response.status_code if 'response' in locals() else None
\n }<\/p>\n
except Exception as e:
\n return {
\n "success": False,
\n "data": None,
\n "error": f"\u672a\u77e5\u9519\u8bef: {str(e)}",
\n "status_code": None
\n }<\/p>\n
return {
\n "success": False,
\n "data": None,
\n "error": "\u8fbe\u5230\u6700\u5927\u91cd\u8bd5\u6b21\u6570",
\n "status_code": None
\n }<\/p>\n
# \u4f7f\u7528\u793a\u4f8b
\nif __name__ == "__main__":
\n result = fetch_api_data("https:\/\/api.example.com\/data")
\n if result["success"]:
\n print("\u83b7\u53d6\u6570\u636e\u6210\u529f:", result["data"])
\n else:
\n print("\u83b7\u53d6\u6570\u636e\u5931\u8d25:", result["error"])<\/p>\n
\u6211\u7684\u8bc4\u4ef7：\u4ee3\u7801\u8d28\u91cf\u5f88\u9ad8！\u4e0d\u4ec5\u5b9e\u73b0\u4e86\u6240\u6709\u9700\u6c42，\u8fd8\u6dfb\u52a0\u4e86\u6211\u6ca1\u60f3\u5230\u7684\u7ec6\u8282，\u6bd4\u5982User-Agent\u5934\u548c\u5b8c\u6574\u7684\u9519\u8bef\u5904\u7406\u3002\u6211\u76f4\u63a5\u7528\u4e86\u8fd9\u4e2a\u51fd\u6570，\u53ea\u505a\u4e86\u5f88\u5c0f\u7684\u8c03\u6574\u3002<\/p>\n
\u6211\u7684\u9700\u6c42：\u5411\u65b0\u624b\u89e3\u91ca\u4ec0\u4e48\u662f"\u5f02\u6b65\u7f16\u7a0b"\u3002<\/p>\n
\u6211\u7ed9Gemma-3\u7684\u63d0\u793a：<\/p>\n
\u7528\u7b80\u5355\u7684\u6bd4\u55bb\u5411\u7f16\u7a0b\u65b0\u624b\u89e3\u91ca\u4ec0\u4e48\u662f\u5f02\u6b65\u7f16\u7a0b\u3002
\n\u8981\u6c42：
\n1. \u7528\u751f\u6d3b\u4e2d\u7684\u4f8b\u5b50\u505a\u6bd4\u55bb
\n2. \u5bf9\u6bd4\u540c\u6b65\u7f16\u7a0b
\n3. \u8bf4\u660e\u4f18\u7f3a\u70b9
\n4. \u7ed9\u51fa\u4e00\u4e2a\u7b80\u5355\u7684\u4ee3\u7801\u793a\u4f8b<\/p>\n
Gemma-3\u7684\u56de\u7b54：<\/p>\n
\u60f3\u8c61\u4e00\u4e0b\u4f60\u8981\u505a\u4e00\u987f\u996d，\u9700\u8981\u5b8c\u6210\u4e09\u4e2a\u4efb\u52a1：\u716e\u996d（20\u5206\u949f）\u3001\u7092\u83dc（10\u5206\u949f）\u3001\u716e\u6c64（15\u5206\u949f）\u3002<\/p>\n
\u540c\u6b65\u7f16\u7a0b（\u4f20\u7edf\u65b9\u5f0f）： \u5c31\u50cf\u4e00\u4e2a\u4eba\u6309\u987a\u5e8f\u505a：<\/p>\n
\u5f02\u6b65\u7f16\u7a0b（\u73b0\u4ee3\u65b9\u5f0f）： \u5c31\u50cf\u806a\u660e\u7684\u53a8\u5e08：<\/p>\n
\u4ee3\u7801\u793a\u4f8b\u5bf9\u6bd4：<\/p>\n
# \u540c\u6b65\u7248\u672c（\u4f4e\u6548）
\nimport time<\/p>\n
def cook_rice():
\n time.sleep(20) # \u716e\u996d20\u5206\u949f
\n print("\u996d\u597d\u4e86")<\/p>\n
def cook_dish():
\n time.sleep(10) # \u7092\u83dc10\u5206\u949f
\n print("\u83dc\u597d\u4e86")<\/p>\n
def cook_soup():
\n time.sleep(15) # \u716e\u6c6415\u5206\u949f
\n print("\u6c64\u597d\u4e86")<\/p>\n
# \u6309\u987a\u5e8f\u6267\u884c，\u603b\u8017\u65f645\u5206\u949f
\ncook_rice()
\ncook_dish()
\ncook_soup()<\/p>\n
# \u5f02\u6b65\u7248\u672c（\u9ad8\u6548）
\nimport asyncio<\/p>\n
async def cook_rice():
\n await asyncio.sleep(20)
\n print("\u996d\u597d\u4e86")<\/p>\n
async def cook_dish():
\n await asyncio.sleep(10)
\n print("\u83dc\u597d\u4e86")<\/p>\n
async def cook_soup():
\n await asyncio.sleep(15)
\n print("\u6c64\u597d\u4e86")<\/p>\n
# \u540c\u65f6\u5f00\u59cb，\u603b\u8017\u65f6\u7ea620\u5206\u949f
\nasync def main():
\n await asyncio.gather(
\n cook_rice(),
\n cook_dish(),
\n cook_soup()
\n )<\/p>\n
asyncio.run(main())<\/p>\n
\u4f18\u70b9：<\/p>\n
\u7f3a\u70b9：<\/p>\n
\u6211\u7684\u8bc4\u4ef7：\u89e3\u91ca\u5f97\u975e\u5e38\u6e05\u695a！\u7528\u505a\u996d\u7684\u6bd4\u55bb\u8ba9\u62bd\u8c61\u6982\u5ff5\u53d8\u5f97\u5177\u4f53，\u4ee3\u7801\u793a\u4f8b\u4e5f\u5f88\u8d34\u5207\u3002\u6211\u76f4\u63a5\u628a\u8fd9\u6bb5\u89e3\u91ca\u7528\u5728\u4e86\u56e2\u961f\u57f9\u8bad\u4e2d\u3002<\/p>\n
\u6211\u7684\u9700\u6c42：\u5199\u4e00\u7bc7\u5173\u4e8e"Python\u865a\u62df\u73af\u5883\u6700\u4f73\u5b9e\u8df5"\u7684\u6280\u672f\u535a\u5ba2\u5927\u7eb2\u3002<\/p>\n
\u6211\u7ed9Gemma-3\u7684\u63d0\u793a：<\/p>\n
\u5e2e\u6211\u5236\u5b9a\u4e00\u7bc7\u6280\u672f\u535a\u5ba2\u7684\u5927\u7eb2，\u4e3b\u9898\u662f"Python\u865a\u62df\u73af\u5883\u6700\u4f73\u5b9e\u8df5"\u3002
\n\u8981\u6c42：
\n1. \u9762\u5411\u4e2d\u7ea7Python\u5f00\u53d1\u8005
\n2. \u5305\u542b\u5b9e\u9645\u95ee\u9898\u548c\u89e3\u51b3\u65b9\u6848
\n3. \u6709\u5177\u4f53\u7684\u547d\u4ee4\u793a\u4f8b
\n4. \u7ed3\u6784\u6e05\u6670，\u9002\u5408\u5199\u62102000\u5b57\u6587\u7ae0<\/p>\n
Gemma-3\u751f\u6210\u7684\u5927\u7eb2：<\/p>\n
# Python\u865a\u62df\u73af\u5883\u5b8c\u5168\u6307\u5357：\u4ece\u5165\u95e8\u5230\u6700\u4f73\u5b9e\u8df5<\/p>\n
## 1. \u4e3a\u4ec0\u4e48\u9700\u8981\u865a\u62df\u73af\u5883？（\u75db\u70b9\u5206\u6790）
\n– \u4f9d\u8d56\u51b2\u7a81\u7684\u771f\u5b9e\u6848\u4f8b
\n– \u9879\u76ee\u9694\u79bb\u7684\u91cd\u8981\u6027
\n– \u56e2\u961f\u534f\u4f5c\u7684\u4e00\u81f4\u6027\u8981\u6c42<\/p>\n
## 2. \u865a\u62df\u73af\u5883\u5de5\u5177\u5bf9\u6bd4
\n### 2.1 venv（Python\u5185\u7f6e）
\n– \u4f18\u70b9：\u65e0\u9700\u5b89\u88c5\u3001\u7b80\u5355\u53ef\u9760
\n– \u7f3a\u70b9：\u529f\u80fd\u76f8\u5bf9\u57fa\u7840
\n– \u9002\u7528\u573a\u666f：\u5927\u591a\u6570\u9879\u76ee<\/p>\n
### 2.2 virtualenv
\n– \u4f18\u70b9：\u66f4\u5feb\u7684\u521b\u5efa\u901f\u5ea6\u3001\u66f4\u591a\u529f\u80fd
\n– \u7f3a\u70b9：\u9700\u8981\u989d\u5916\u5b89\u88c5
\n– \u9002\u7528\u573a\u666f：\u9700\u8981\u9ad8\u7ea7\u529f\u80fd\u7684\u9879\u76ee<\/p>\n
### 2.3 conda
\n– \u4f18\u70b9：\u8de8\u8bed\u8a00\u3001\u4e8c\u8fdb\u5236\u5305\u7ba1\u7406
\n– \u7f3a\u70b9：\u4f53\u79ef\u8f83\u5927\u3001\u901f\u5ea6\u8f83\u6162
\n– \u9002\u7528\u573a\u666f：\u6570\u636e\u79d1\u5b66\u3001\u673a\u5668\u5b66\u4e60\u9879\u76ee<\/p>\n
### 2.4 pipenv & poetry
\n– \u4f18\u70b9：\u4f9d\u8d56\u9501\u5b9a\u3001\u66f4\u597d\u7684\u5de5\u4f5c\u6d41
\n– \u7f3a\u70b9：\u5b66\u4e60\u66f2\u7ebf\u8f83\u9661
\n– \u9002\u7528\u573a\u666f：\u751f\u4ea7\u73af\u5883\u3001\u56e2\u961f\u9879\u76ee<\/p>\n
## 3. \u624b\u628a\u624b\u5b9e\u6218\u6559\u7a0b
\n### 3.1 \u4f7f\u7528venv\u521b\u5efa\u548c\u7ba1\u7406\u73af\u5883
\n```bash
\n# \u521b\u5efa\u73af\u5883
\npython -m venv myenv<\/p>\n
# \u6fc0\u6d3b\u73af\u5883（Linux\/Mac）
\nsource myenv\/bin\/activate<\/p>\n
# \u6fc0\u6d3b\u73af\u5883（Windows）
\nmyenv\\\\Scripts\\\\activate<\/p>\n
# \u5b89\u88c5\u5305
\npip install requests pandas<\/p>\n
# \u5bfc\u51fa\u4f9d\u8d56
\npip freeze > requirements.txt<\/p>\n
# \u9000\u51fa\u73af\u5883
\ndeactivate<\/p>\n
# \u4ecerequirements.txt\u5b89\u88c5
\npip install -r requirements.txt<\/p>\n
# \u7cbe\u786e\u590d\u5236\u73af\u5883
\npip list –format=freeze > requirements.txt<\/p>\n
# \u4f7f\u7528pip-tools\u7ba1\u7406\u4f9d\u8d56
\npip-compile requirements.in
\npip-sync requirements.txt<\/p>\n
**\u6211\u7684\u8bc4\u4ef7**：\u5927\u7eb2\u975e\u5e38\u4e13\u4e1a\u548c\u5168\u9762！\u4e0d\u4ec5\u6db5\u76d6\u4e86\u6240\u6709\u91cd\u8981\u4e3b\u9898，\u8fd8\u7ed9\u51fa\u4e86\u5177\u4f53\u7684\u547d\u4ee4\u793a\u4f8b\u3002\u6211\u6309\u7167\u8fd9\u4e2a\u5927\u7eb2\u5199\u6587\u7ae0，\u8282\u7701\u4e86\u5927\u91cf\u6784\u601d\u65f6\u95f4\u3002<\/p>\n
## 6. \u6027\u80fd\u76d1\u63a7\u548c\u95ee\u9898\u6392\u67e5<\/p>\n
\u90e8\u7f72\u4f18\u5316\u540e，\u6211\u4eec\u9700\u8981\u77e5\u9053\u6a21\u578b\u8fd0\u884c\u5f97\u600e\u4e48\u6837\u3002\u8fd9\u91cc\u5206\u4eab\u4e00\u4e9b\u76d1\u63a7\u548c\u6392\u67e5\u7684\u6280\u5de7\u3002<\/p>\n
### 6.1 \u76d1\u63a7\u5173\u952e\u6307\u6807<\/p>\n
\u521b\u5efa\u4e00\u4e2a\u7b80\u5355\u7684\u76d1\u63a7\u811a\u672c：<\/p>\n
```python
\n# monitor.py
\nimport psutil
\nimport torch
\nimport time
\nfrom datetime import datetime<\/p>\n
def monitor_system():
\n """\u76d1\u63a7\u7cfb\u7edf\u8d44\u6e90"""
\n cpu_percent = psutil.cpu_percent(interval=1)
\n memory = psutil.virtual_memory()<\/p>\n
info = {
\n "timestamp": datetime.now().strftime("%Y-%m-%d %H:%M:%S"),
\n "cpu_percent": cpu_percent,
\n "memory_percent": memory.percent,
\n "memory_used_gb": memory.used \/ (1024**3),
\n "memory_available_gb": memory.available \/ (1024**3),
\n }<\/p>\n
if torch.cuda.is_available():
\n gpu_memory = torch.cuda.memory_allocated() \/ (1024**3)
\n gpu_memory_max = torch.cuda.max_memory_allocated() \/ (1024**3)
\n info.update({
\n "gpu_memory_used_gb": gpu_memory,
\n "gpu_memory_max_gb": gpu_memory_max,
\n })<\/p>\n
return info<\/p>\n
def log_performance(prompt, response, generation_time):
\n """\u8bb0\u5f55\u751f\u6210\u6027\u80fd"""
\n with open("performance.log", "a") as f:
\n log_entry = {
\n "time": datetime.now().isoformat(),
\n "prompt_length": len(prompt),
\n "response_length": len(response),
\n "generation_time": generation_time,
\n "tokens_per_second": len(response.split()) \/ generation_time if generation_time > 0 else 0
\n }
\n f.write(str(log_entry) + "\\\\n")<\/p>\n
# \u5728\u751f\u6210\u51fd\u6570\u4e2d\u6dfb\u52a0\u76d1\u63a7
\ndef generate_with_monitoring(message, history):
\n start_time = time.time()<\/p>\n
# \u751f\u6210\u524d\u7684\u7cfb\u7edf\u72b6\u6001
\n before = monitor_system()<\/p>\n
# \u751f\u6210\u56de\u590d
\n response = generate_response(message, history)<\/p>\n
# \u751f\u6210\u540e\u7684\u7cfb\u7edf\u72b6\u6001
\n after = monitor_system()
\n generation_time = time.time() – start_time<\/p>\n
# \u8bb0\u5f55\u6027\u80fd
\n log_performance(message, response, generation_time)<\/p>\n
# \u6253\u5370\u76d1\u63a7\u4fe1\u606f
\n print(f"\u751f\u6210\u65f6\u95f4: {generation_time:.2f}\u79d2")
\n print(f"\u5185\u5b58\u53d8\u5316: {after['memory_used_gb'] – before['memory_used_gb']:.2f} GB")<\/p>\n
return response<\/p>\n
\u95ee\u98981：\u54cd\u5e94\u901f\u5ea6\u53d8\u6162<\/p>\n
\u53ef\u80fd\u539f\u56e0\u548c\u89e3\u51b3\u65b9\u6848：<\/p>\n
# \u68c0\u67e5\u7cfb\u7edf\u8d1f\u8f7d
\ntop -c<\/p>\n
# \u68c0\u67e5\u5185\u5b58\u4f7f\u7528
\nfree -h<\/p>\n
# \u68c0\u67e5GPU\u4f7f\u7528（\u5982\u679c\u6709）
\nnvidia-smi<\/p>\n
# \u89e3\u51b3\u65b9\u6848：
\n# 1. \u91cd\u542f\u670d\u52a1\u91ca\u653e\u5185\u5b58
\n.\/manage.sh restart<\/p>\n
# 2. \u51cf\u5c11\u5e76\u53d1\u8bf7\u6c42
\n# 3. \u8c03\u6574\u751f\u6210\u53c2\u6570（\u51cf\u5c11max_tokens）
\n# 4. \u6e05\u7406\u6a21\u578b\u7f13\u5b58<\/p>\n
\u95ee\u98982：\u5185\u5b58\u4e0d\u8db3\u9519\u8bef<\/p>\n
# \u9519\u8bef\u4fe1\u606f：CUDA out of memory \u6216 RuntimeError<\/p>\n
# \u89e3\u51b3\u65b9\u6848：
\n# 1. \u4f7f\u7528\u66f4\u5c0f\u7684\u6279\u6b21
\nmodel.generate(…, batch_size=1)<\/p>\n
# 2. \u4f7f\u7528\u68af\u5ea6\u68c0\u67e5\u70b9（\u5982\u679c\u8bad\u7ec3）
\nmodel.gradient_checkpointing_enable()<\/p>\n
# 3. \u4f7f\u7528CPU\u5378\u8f7d
\nmodel = AutoModelForCausalLM.from_pretrained(
\n MODEL_PATH,
\n device_map="auto",
\n offload_folder="offload",
\n offload_state_dict=True
\n)<\/p>\n
# 4. \u4f7f\u7528\u91cf\u5316（\u5982\u679c\u652f\u6301）
\nfrom transformers import BitsAndBytesConfig
\nquantization_config = BitsAndBytesConfig(load_in_8bit=True)
\nmodel = AutoModelForCausalLM.from_pretrained(
\n MODEL_PATH,
\n quantization_config=quantization_config
\n)<\/p>\n
\u95ee\u98983：\u751f\u6210\u8d28\u91cf\u4e0b\u964d<\/p>\n
# \u8c03\u6574\u8fd9\u4e9b\u53c2\u6570\u6539\u5584\u8d28\u91cf：<\/p>\n
# 1. \u964d\u4f4e\u6e29\u5ea6，\u63d0\u9ad8\u786e\u5b9a\u6027
\ntemperature=0.3<\/p>\n
# 2. \u4f7f\u7528top-p\u91c7\u6837\u800c\u4e0d\u662ftop-k
\ntop_p=0.9
\ntop_k=0 # \u7981\u7528top-k<\/p>\n
# 3. \u6dfb\u52a0\u91cd\u590d\u60e9\u7f5a
\nrepetition_penalty=1.2<\/p>\n
# 4. \u4f7f\u7528\u66f4\u597d\u7684\u63d0\u793a\u8bcd
\nprompt = """\u8bf7\u4ed4\u7ec6\u601d\u8003\u540e\u56de\u7b54\u4ee5\u4e0b\u95ee\u9898\u3002
\n\u786e\u4fdd\u56de\u7b54\u51c6\u786e\u3001\u5b8c\u6574\u3001\u6709\u7528\u3002<\/p>\n
\u95ee\u9898：{user_question}<\/p>\n
\u56de\u7b54："""<\/p>\n
\u5b9a\u671f\u68c0\u67e5\u8fd9\u4e9b\u9879\u76ee，\u786e\u4fdd\u6700\u4f73\u6027\u80fd：<\/p>\n
# \u6027\u80fd\u68c0\u67e5\u6e05\u5355
\n\u6bcf\u65e5\u68c0\u67e5:
\n – [ ] \u5185\u5b58\u4f7f\u7528\u662f\u5426\u6b63\u5e38 (<80%)
\n – [ ] \u54cd\u5e94\u65f6\u95f4\u662f\u5426\u7a33\u5b9a (<10\u79d2)
\n – [ ] \u9519\u8bef\u7387\u662f\u5426\u6b63\u5e38 (<1%)
\n – [ ] \u670d\u52a1\u662f\u5426\u53ef\u8bbf\u95ee<\/p>\n
\u6bcf\u5468\u68c0\u67e5:
\n – [ ] \u6e05\u7406\u65e5\u5fd7\u6587\u4ef6
\n – [ ] \u66f4\u65b0Python\u5305
\n – [ ] \u5907\u4efd\u6a21\u578b\u548c\u914d\u7f6e
\n – [ ] \u68c0\u67e5\u78c1\u76d8\u7a7a\u95f4<\/p>\n
\u6bcf\u6708\u68c0\u67e5:
\n – [ ] \u8bc4\u4f30\u751f\u6210\u8d28\u91cf
\n – [ ] \u4f18\u5316\u63d0\u793a\u8bcd\u6a21\u677f
\n – [ ] \u5ba1\u67e5\u6027\u80fd\u65e5\u5fd7
\n – [ ] \u8003\u8651\u6a21\u578b\u66f4\u65b0<\/p>\n
\u7ecf\u8fc7\u51e0\u4e2a\u6708\u7684\u5b9e\u9645\u4f7f\u7528\u548c\u4f18\u5316，\u6211\u5bf9Gemma-3-12B-IT\u5728\u4f4e\u8d44\u6e90\u670d\u52a1\u5668\u4e0a\u7684\u8868\u73b0\u6709\u4e86\u6df1\u523b\u7684\u7406\u89e3\u3002\u4e0b\u9762\u662f\u6211\u7684\u603b\u7ed3\u548c\u4e00\u4e9b\u5b9e\u7528\u5efa\u8bae\u3002<\/p>\n
Gemma-3-12B-IT\u7684\u4f18\u52bf：<\/p>\n
\u90e8\u7f72\u4f18\u5316\u7684\u5173\u952e\u70b9：<\/p>\n
\u5982\u679c\u4f60\u662f\u5c0f\u56e2\u961f\u6216\u4e2a\u4eba\u5f00\u53d1\u8005：<\/p>\n
\u5982\u679c\u4f60\u662f\u4f01\u4e1a\u7528\u6237：<\/p>\n
\u5982\u679c\u4f60\u8d44\u6e90\u975e\u5e38\u6709\u9650（\u5185\u5b58<24GB）：<\/p>\n
\u6280\u672f\u603b\u662f\u5728\u8fdb\u6b65，\u8fd9\u91cc\u6709\u4e00\u4e9b\u53ef\u4ee5\u7ee7\u7eed\u63a2\u7d22\u7684\u65b9\u5411：<\/p>\n
\u90e8\u7f72\u548c\u4f18\u5316\u5927\u8bed\u8a00\u6a21\u578b\u662f\u4e00\u4e2a\u6301\u7eed\u7684\u8fc7\u7a0b，\u4e0d\u662f\u4e00\u52b3\u6c38\u9038\u7684\u3002\u6211\u7684\u5efa\u8bae\u662f：<\/p>\n
Gemma-3-12B-IT\u662f\u4e00\u4e2a\u5f3a\u5927\u7684\u5de5\u5177，\u4f46\u5de5\u5177\u7684\u4ef7\u503c\u5728\u4e8e\u5982\u4f55\u4f7f\u7528\u3002\u901a\u8fc7\u5408\u7406\u7684\u90e8\u7f72\u548c\u4f18\u5316，\u4f60\u5b8c\u5168\u53ef\u4ee5\u5728\u6709\u9650\u7684\u8d44\u6e90\u4e0b，\u83b7\u5f97\u63a5\u8fd1\u9ad8\u7aef\u914d\u7f6e\u7684\u4f7f\u7528\u4f53\u9a8c\u3002<\/p>\n
\u8bb0\u4f4f，\u6700\u597d\u7684\u4f18\u5316\u4e0d\u662f\u8ffd\u6c42\u6781\u81f4\u7684\u6027\u80fd，\u800c\u662f\u5728\u8d44\u6e90\u9650\u5236\u548c\u7528\u6237\u4f53\u9a8c\u4e4b\u95f4\u627e\u5230\u6700\u4f73\u7684\u5e73\u8861\u70b9\u3002\u5e0c\u671b\u8fd9\u7bc7\u6559\u7a0b\u80fd\u5e2e\u52a9\u4f60\u987a\u5229\u90e8\u7f72Gemma-3，\u5e76\u5728\u5b9e\u9645\u9879\u76ee\u4e2d\u53d1\u6325\u5b83\u7684\u4ef7\u503c\u3002<\/p>\n
\u83b7\u53d6\u66f4\u591aAI\u955c\u50cf<\/p>\n
\u60f3\u63a2\u7d22\u66f4\u591aAI\u955c\u50cf\u548c\u5e94\u7528\u573a\u666f？\u8bbf\u95ee CSDN\u661f\u56fe\u955c\u50cf\u5e7f\u573a，\u63d0\u4f9b\u4e30\u5bcc\u7684\u9884\u7f6e\u955c\u50cf，\u8986\u76d6\u5927\u6a21\u578b\u63a8\u7406\u3001\u56fe\u50cf\u751f\u6210\u3001\u89c6\u9891\u751f\u6210\u3001\u6a21\u578b\u5fae\u8c03\u7b49\u591a\u4e2a\u9886\u57df，\u652f\u6301\u4e00\u952e\u90e8\u7f72\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"
Gemma-3-12B-IT\u5165\u95e8\u6559\u7a0b：120\u4ebf\u53c2\u6570\u6a21\u578b\u5728\u4f4e\u8d44\u6e90\u670d\u52a1\u5668\u4e0a\u7684\u90e8\u7f72\u4f18\u5316\u6280\u5de7
\n1. \u5f00\u7bc7：\u4e3a\u4ec0\u4e48\u9009\u62e9Gemma-3-12B-IT？
\n\u5982\u679c\u4f60\u6b63\u5728\u5bfb\u627e\u4e00\u4e2a\u65e2\u5f3a\u5927\u53c8\u80fd\u5728\u666e\u901a\u670d\u52a1\u5668\u4e0a\u6d41\u7545\u8fd0\u884c\u7684\u5927\u8bed\u8a00\u6a21\u578b，\u90a3\u4e48Gemma-3-12B-IT\u5f88\u53ef\u80fd\u5c31\u662f\u4f60\u7684\u7406\u60f3\u9009\u62e9\u3002
\n\u8ba9\u6211\u5148\u8bf4\u8bf4\u6211\u7684\u4eb2\u8eab\u7ecf\u5386\u3002\u524d\u6bb5\u65f6\u95f4，\u6211\u9700\u8981\u4e3a\u4e00\u4e2a\u5c0f\u578b\u5f00\u53d1\u56e2\u961f\u90e8\u7f72\u4e00\u4e2aAI\u52a9\u624b，\u9884\u7b97\u6709\u9650，\u670d\u52a1\u5668\u914d\u7f6e\u4e5f\u4e0d\u7b97\u9ad8ÿ<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[1871,4817,956,2094],"topic":[],"class_list":["post-83420","post","type-post","status-publish","format-standard","hentry","category-server","tag-ai","tag-4817","tag-956","tag-2094"],"yoast_head":"\n