\u5982\u679c\u4f60\u624b\u5934\u6709\u4e00\u53f0MacBook，\u7279\u522b\u662f\u642d\u8f7d\u4e86M2\u6216M3\u82af\u7247\u7684\u578b\u53f7，\u4f60\u53ef\u80fd\u4f1a\u53d1\u73b0\u5f88\u591aAI\u6a21\u578b\u90e8\u7f72\u6559\u7a0b\u90fd\u662f\u9488\u5bf9x86\u67b6\u6784\u7684，\u7528\u8d77\u6765\u603b\u662f\u4e0d\u592a\u987a\u624b\u3002\u4eca\u5929，\u6211\u4eec\u5c31\u6765\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\u3002<\/p>\n
Qwen3-Reranker-0.6B\u662f\u901a\u4e49\u5343\u95ee\u6700\u65b0\u63a8\u51fa\u7684\u91cd\u6392\u5e8f\u6a21\u578b，\u4e13\u95e8\u7528\u6765\u7ed9\u641c\u7d22\u7ed3\u679c\u201c\u6253\u5206\u6392\u961f\u201d\u3002\u60f3\u8c61\u4e00\u4e0b，\u4f60\u641c\u7d22\u201c\u5982\u4f55\u505a\u7ea2\u70e7\u8089\u201d，\u641c\u7d22\u5f15\u64ce\u8fd4\u56de\u4e86100\u4e2a\u7ed3\u679c，\u8fd9\u4e2a\u6a21\u578b\u80fd\u5e2e\u4f60\u5224\u65ad\u54ea\u4e2a\u83dc\u8c31\u6700\u9760\u8c31\u3001\u54ea\u4e2a\u89c6\u9891\u8bb2\u89e3\u6700\u8be6\u7ec6，\u7136\u540e\u6309\u76f8\u5173\u6027\u91cd\u65b0\u6392\u5e8f\u3002<\/p>\n
\u4f46\u95ee\u9898\u662f，\u5b98\u65b9\u6587\u6863\u901a\u5e38\u9ed8\u8ba4\u4f60\u5728x86\u670d\u52a1\u5668\u4e0a\u8fd0\u884c\u3002\u5982\u679c\u4f60\u7528\u7684\u662fMac M\u7cfb\u5217\u7535\u8111，\u6216\u8005\u6811\u8393\u6d3e\u8fd9\u7c7bARM\u67b6\u6784\u7684\u8bbe\u5907，\u76f4\u63a5\u7167\u642c\u6559\u7a0b\u53ef\u80fd\u4f1a\u9047\u5230\u5404\u79cd\u517c\u5bb9\u6027\u95ee\u9898\u3002\u8fd9\u7bc7\u6587\u7ae0\u5c31\u662f\u4e3a\u4f60\u51c6\u5907\u7684\u2014\u2014\u6211\u4f1a\u624b\u628a\u624b\u5e26\u4f60\u5b8c\u6210\u5728ARM\u67b6\u6784\u4e0a\u7684\u5b8c\u6574\u90e8\u7f72，\u8ba9\u4f60\u5728\u81ea\u5df1\u7684\u8bbe\u5907\u4e0a\u4e5f\u80fd\u987a\u7545\u4f7f\u7528\u8fd9\u4e2a\u5f3a\u5927\u7684\u91cd\u6392\u5e8f\u5de5\u5177\u3002<\/p>\n
\u5728\u5f00\u59cb\u4e4b\u524d，\u6211\u4eec\u5148\u786e\u8ba4\u4e00\u4e0b\u73af\u5883\u3002\u6253\u5f00\u4f60\u7684\u7ec8\u7aef，\u8f93\u5165\u4ee5\u4e0b\u547d\u4ee4：<\/p>\n
# \u67e5\u770b\u7cfb\u7edf\u67b6\u6784
\nuname -m<\/p>\n
# \u67e5\u770bPython\u7248\u672c
\npython3 –version<\/p>\n
# \u67e5\u770bpip\u7248\u672c
\npip3 –version<\/p>\n
\u4f60\u5e94\u8be5\u4f1a\u770b\u5230\u7c7b\u4f3c\u8fd9\u6837\u7684\u8f93\u51fa：<\/p>\n
\u5982\u679c\u4f60\u7684Python\u7248\u672c\u4f4e\u4e8e3.8，\u9700\u8981\u5148\u5347\u7ea7\u3002\u5728Mac\u4e0a，\u53ef\u4ee5\u7528Homebrew：<\/p>\n
# \u5b89\u88c5\u6216\u5347\u7ea7Python
\nbrew install python@3.10<\/p>\n
Qwen3-Reranker-0.6B\u6a21\u578b\u5927\u5c0f\u7ea61.2GB，\u6211\u4eec\u9700\u8981\u5148\u4e0b\u8f7d\u5230\u672c\u5730\u3002\u8fd9\u91cc\u63d0\u4f9b\u4e24\u79cd\u65b9\u5f0f：<\/p>\n
\u65b9\u5f0f\u4e00：\u4f7f\u7528Hugging Face CLI（\u63a8\u8350）<\/p>\n
# \u5b89\u88c5huggingface-hub
\npip3 install huggingface-hub<\/p>\n
# \u4e0b\u8f7d\u6a21\u578b
\npython3 -c "from huggingface_hub import snapshot_download; snapshot_download(repo_id='Qwen\/Qwen3-Reranker-0.6B', local_dir='.\/Qwen3-Reranker-0.6B')"<\/p>\n
\u65b9\u5f0f\u4e8c：\u624b\u52a8\u4e0b\u8f7d<\/p>\n
\u5982\u679c\u4f60\u7f51\u7edc\u73af\u5883\u7279\u6b8a，\u4e5f\u53ef\u4ee5\u4ece\u955c\u50cf\u7ad9\u4e0b\u8f7d：<\/p>\n
\u4e0b\u8f7d\u5b8c\u6210\u540e，\u4f60\u7684\u76ee\u5f55\u7ed3\u6784\u5e94\u8be5\u662f\u8fd9\u6837\u7684：<\/p>\n
Qwen3-Reranker-0.6B\/
\n\u251c\u2500\u2500 config.json
\n\u251c\u2500\u2500 model.safetensors
\n\u251c\u2500\u2500 tokenizer.json
\n\u251c\u2500\u2500 tokenizer_config.json
\n\u2514\u2500\u2500 …\u5176\u4ed6\u914d\u7f6e\u6587\u4ef6<\/p>\n
ARM\u67b6\u6784\u4e0a\u5b89\u88c5PyTorch\u9700\u8981\u7279\u522b\u6ce8\u610f\u3002PyTorch\u5b98\u65b9\u4e3aARM\u63d0\u4f9b\u4e86\u9884\u7f16\u8bd1\u7248\u672c，\u4f46\u9700\u8981\u6307\u5b9a\u6b63\u786e\u7684\u7248\u672c\u3002<\/p>\n
# \u9996\u5148\u5347\u7ea7pip
\npip3 install –upgrade pip<\/p>\n
# \u5b89\u88c5PyTorch（ARM\u517c\u5bb9\u7248\u672c）
\n# \u5bf9\u4e8eMac M\u7cfb\u5217，\u4f7f\u7528\u4ee5\u4e0b\u547d\u4ee4：
\npip3 install torch torchvision torchaudio<\/p>\n
# \u5bf9\u4e8e\u5176\u4ed6ARM\u8bbe\u5907（\u5982\u6811\u8393\u6d3e），\u53ef\u80fd\u9700\u8981\u4ece\u6e90\u7801\u7f16\u8bd1
\n# \u4f46\u901a\u5e38PyTorch\u5b98\u65b9\u4e5f\u63d0\u4f9bARM\u7248\u672c，\u53ef\u4ee5\u5148\u5c1d\u8bd5：
\npip3 install torch<\/p>\n
# \u5b89\u88c5\u5176\u4ed6\u4f9d\u8d56
\npip3 install transformers>=4.51.0
\npip3 install gradio>=4.0.0
\npip3 install accelerate safetensors sentencepiece<\/p>\n
\u91cd\u8981\u63d0\u793a：\u5982\u679c\u5728\u5b89\u88c5PyTorch\u65f6\u9047\u5230\u95ee\u9898，\u53ef\u4ee5\u8bbf\u95eePyTorch\u5b98\u7f51，\u9009\u62e9\u5bf9\u5e94\u7684ARM\u7248\u672c\u83b7\u53d6\u5b89\u88c5\u547d\u4ee4\u3002<\/p>\n
\u5728\u6a21\u578b\u76ee\u5f55\u4e0b\u521b\u5efa\u4e00\u4e2astart.sh\u6587\u4ef6：<\/p>\n
#!\/bin\/bash<\/p>\n
# Qwen3-Reranker-0.6B ARM\u67b6\u6784\u542f\u52a8\u811a\u672c
\n# \u9002\u7528\u4e8eMac M1\/M2\/M3\u53caARM\u670d\u52a1\u5668<\/p>\n
echo "\u6b63\u5728\u542f\u52a8 Qwen3-Reranker-0.6B \u670d\u52a1…"<\/p>\n
# \u68c0\u67e5Python\u7248\u672c
\npython_version=$(python3 -c 'import sys; print(f"{sys.version_info.major}.{sys.version_info.minor}")')
\necho "Python\u7248\u672c: $python_version"<\/p>\n
if [[ $(echo "$python_version < 3.8" | bc) -eq 1 ]]; then
\n echo "\u9519\u8bef: Python\u7248\u672c\u9700\u89813.8\u6216\u66f4\u9ad8"
\n exit 1
\nfi<\/p>\n
# \u68c0\u67e5\u5fc5\u8981\u4f9d\u8d56
\necho "\u68c0\u67e5\u4f9d\u8d56…"
\nfor package in torch transformers gradio accelerate; do
\n if ! python3 -c "import $package" 2>\/dev\/null; then
\n echo "\u7f3a\u5c11\u4f9d\u8d56: $package"
\n echo "\u6b63\u5728\u5b89\u88c5…"
\n pip3 install $package
\n fi
\ndone<\/p>\n
# \u8bbe\u7f6e\u73af\u5883\u53d8\u91cf（\u9488\u5bf9ARM\u4f18\u5316）
\nexport PYTORCH_ENABLE_MPS_FALLBACK=1 # \u5141\u8bb8MPS\u56de\u9000\u5230CPU
\nexport GRADIO_SERVER_PORT=7860 # \u8bbe\u7f6e\u7aef\u53e3<\/p>\n
# \u542f\u52a8\u670d\u52a1
\necho "\u542f\u52a8Web\u670d\u52a1…"
\necho "\u670d\u52a1\u5c06\u5728 http:\/\/localhost:7860 \u53ef\u7528"
\necho "\u6309 Ctrl+C \u505c\u6b62\u670d\u52a1"<\/p>\n
python3 app.py<\/p>\n
\u7ed9\u811a\u672c\u6dfb\u52a0\u6267\u884c\u6743\u9650：<\/p>\n
chmod +x start.sh<\/p>\n
\u521b\u5efaapp.py\u6587\u4ef6，\u8fd9\u662fWeb\u670d\u52a1\u7684\u6838\u5fc3：<\/p>\n
#!\/usr\/bin\/env python3
\n# -*- coding: utf-8 -*-
\n"""
\nQwen3-Reranker-0.6B Web\u670d\u52a1
\nARM\u67b6\u6784\u9002\u914d\u7248\u672c
\n"""<\/p>\n
import os
\nimport sys
\nimport torch
\nimport gradio as gr
\nfrom transformers import AutoModelForSequenceClassification, AutoTokenizer
\nimport logging<\/p>\n
# \u8bbe\u7f6e\u65e5\u5fd7
\nlogging.basicConfig(level=logging.INFO)
\nlogger = logging.getLogger(__name__)<\/p>\n
def check_arm_environment():
\n """\u68c0\u67e5ARM\u73af\u5883\u5e76\u4f18\u5316\u8bbe\u7f6e"""
\n logger.info("\u68c0\u67e5\u8fd0\u884c\u73af\u5883…")<\/p>\n
# \u68c0\u67e5\u67b6\u6784
\n import platform
\n arch = platform.machine()
\n logger.info(f"\u7cfb\u7edf\u67b6\u6784: {arch}")<\/p>\n
# ARM\u67b6\u6784\u7279\u5b9a\u4f18\u5316
\n if arch in ['arm64', 'aarch64']:
\n logger.info("\u68c0\u6d4b\u5230ARM\u67b6\u6784，\u8fdb\u884c\u4f18\u5316\u8bbe\u7f6e")<\/p>\n
# \u5bf9\u4e8eMac M\u7cfb\u5217，\u4f7f\u7528MPS\u52a0\u901f
\n if torch.backends.mps.is_available():
\n device = torch.device("mps")
\n logger.info("\u4f7f\u7528MPS\u52a0\u901f (Apple Silicon)")
\n else:
\n device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
\n logger.info(f"\u4f7f\u7528\u8bbe\u5907: {device}")
\n else:
\n device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
\n logger.info(f"\u4f7f\u7528\u8bbe\u5907: {device}")<\/p>\n
return device<\/p>\n
def load_model(model_path=".\/Qwen3-Reranker-0.6B"):
\n """\u52a0\u8f7d\u6a21\u578b\u548c\u5206\u8bcd\u5668"""
\n logger.info("\u5f00\u59cb\u52a0\u8f7d\u6a21\u578b…")<\/p>\n
try:
\n # \u68c0\u67e5\u6a21\u578b\u6587\u4ef6\u662f\u5426\u5b58\u5728
\n if not os.path.exists(model_path):
\n logger.error(f"\u6a21\u578b\u8def\u5f84\u4e0d\u5b58\u5728: {model_path}")
\n logger.info("\u8bf7\u786e\u4fdd\u6a21\u578b\u6587\u4ef6\u5df2\u4e0b\u8f7d\u5230\u6b63\u786e\u4f4d\u7f6e")
\n return None, None<\/p>\n
# \u52a0\u8f7d\u5206\u8bcd\u5668
\n logger.info("\u52a0\u8f7d\u5206\u8bcd\u5668…")
\n tokenizer = AutoTokenizer.from_pretrained(
\n model_path,
\n trust_remote_code=True
\n )<\/p>\n
# \u52a0\u8f7d\u6a21\u578b
\n logger.info("\u52a0\u8f7d\u6a21\u578b…")
\n model = AutoModelForSequenceClassification.from_pretrained(
\n model_path,
\n trust_remote_code=True,
\n torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32
\n )<\/p>\n
# \u79fb\u52a8\u5230\u5bf9\u5e94\u8bbe\u5907
\n device = check_arm_environment()
\n model = model.to(device)
\n model.eval()<\/p>\n
logger.info("\u6a21\u578b\u52a0\u8f7d\u5b8c\u6210!")
\n return model, tokenizer, device<\/p>\n
except Exception as e:
\n logger.error(f"\u52a0\u8f7d\u6a21\u578b\u5931\u8d25: {str(e)}")
\n return None, None, None<\/p>\n
def rerank_documents(query, documents, instruction=None, batch_size=8):
\n """
\n \u91cd\u6392\u5e8f\u6587\u6863<\/p>\n
\u53c2\u6570:
\n – query: \u67e5\u8be2\u6587\u672c
\n – documents: \u6587\u6863\u5217\u8868，\u6bcf\u884c\u4e00\u4e2a
\n – instruction: \u4efb\u52a1\u6307\u4ee4（\u53ef\u9009）
\n – batch_size: \u6279\u5904\u7406\u5927\u5c0f
\n """
\n global model, tokenizer, device<\/p>\n
if model is None or tokenizer is None:
\n return "\u9519\u8bef: \u6a21\u578b\u672a\u52a0\u8f7d\u6210\u529f"<\/p>\n
try:
\n # \u5904\u7406\u6587\u6863\u8f93\u5165
\n if isinstance(documents, str):
\n doc_list = [doc.strip() for doc in documents.split('\\\\n') if doc.strip()]
\n else:
\n doc_list = documents<\/p>\n
if not doc_list:
\n return "\u9519\u8bef: \u6ca1\u6709\u6709\u6548\u7684\u6587\u6863"<\/p>\n
# \u9650\u5236\u6587\u6863\u6570\u91cf（\u6027\u80fd\u8003\u8651）
\n if len(doc_list) > 100:
\n doc_list = doc_list[:100]
\n logger.warning(f"\u6587\u6863\u6570\u91cf\u8d85\u8fc7100，\u53ea\u5904\u7406\u524d100\u4e2a")<\/p>\n
# \u51c6\u5907\u8f93\u5165
\n pairs = [[query, doc] for doc in doc_list]<\/p>\n
# \u5206\u6279\u5904\u7406
\n scores = []
\n for i in range(0, len(pairs), batch_size):
\n batch_pairs = pairs[i:i+batch_size]<\/p>\n
# \u7f16\u7801
\n inputs = tokenizer(
\n batch_pairs,
\n padding=True,
\n truncation=True,
\n max_length=512,
\n return_tensors="pt"
\n ).to(device)<\/p>\n
# \u63a8\u7406
\n with torch.no_grad():
\n outputs = model(**inputs)
\n batch_scores = outputs.logits[:, 0].cpu().numpy()
\n scores.extend(batch_scores.tolist())<\/p>\n
# \u6392\u5e8f
\n sorted_indices = sorted(range(len(scores)), key=lambda i: scores[i], reverse=True)<\/p>\n
# \u683c\u5f0f\u5316\u7ed3\u679c
\n result = "\u91cd\u6392\u5e8f\u7ed3\u679c:\\\\n\\\\n"
\n for rank, idx in enumerate(sorted_indices[:10], 1): # \u53ea\u663e\u793a\u524d10\u4e2a
\n score = scores[idx]
\n doc = doc_list[idx]
\n result += f"{rank}. [\u5f97\u5206: {score:.4f}] {doc}\\\\n"<\/p>\n
if len(doc_list) > 10:
\n result += f"\\\\n… \u5171 {len(doc_list)} \u4e2a\u6587\u6863，\u663e\u793a\u524d10\u4e2a"<\/p>\n
return result<\/p>\n
except Exception as e:
\n logger.error(f"\u63a8\u7406\u9519\u8bef: {str(e)}")
\n return f"\u5904\u7406\u51fa\u9519: {str(e)}"<\/p>\n
# \u5168\u5c40\u53d8\u91cf
\nmodel, tokenizer, device = load_model()<\/p>\n
def create_interface():
\n """\u521b\u5efaGradio\u754c\u9762"""<\/p>\n
# \u754c\u9762\u8bf4\u660e
\n description = """
\n # Qwen3-Reranker-0.6B \u91cd\u6392\u5e8f\u670d\u52a1<\/p>\n
**ARM\u67b6\u6784\u9002\u914d\u7248\u672c** – \u4e13\u4e3aMac M\u7cfb\u5217\u53caARM\u670d\u52a1\u5668\u4f18\u5316<\/p>\n
\u8f93\u5165\u67e5\u8be2\u6587\u672c\u548c\u5019\u9009\u6587\u6863，\u6a21\u578b\u4f1a\u6839\u636e\u76f8\u5173\u6027\u5bf9\u6587\u6863\u8fdb\u884c\u91cd\u65b0\u6392\u5e8f\u3002
\n """<\/p>\n
# \u793a\u4f8b
\n examples = [
\n [
\n "\u4ec0\u4e48\u662f\u4eba\u5de5\u667a\u80fd？",
\n "\u4eba\u5de5\u667a\u80fd\u662f\u8ba1\u7b97\u673a\u79d1\u5b66\u7684\u4e00\u4e2a\u5206\u652f\u3002\\\\n\u673a\u5668\u5b66\u4e60\u662f\u4eba\u5de5\u667a\u80fd\u7684\u4e00\u79cd\u5b9e\u73b0\u65b9\u5f0f\u3002\\\\n\u6df1\u5ea6\u5b66\u4e60\u57fa\u4e8e\u795e\u7ecf\u7f51\u7edc\u3002\\\\nPython\u662f\u4e00\u79cd\u7f16\u7a0b\u8bed\u8a00\u3002",
\n "Given a technical query, retrieve relevant documents"
\n ],
\n [
\n "\u5982\u4f55\u5b66\u4e60\u7f16\u7a0b？",
\n "\u7f16\u7a0b\u9700\u8981\u903b\u8f91\u601d\u7ef4\u3002\\\\n\u53ef\u4ee5\u4ecePython\u5f00\u59cb\u5b66\u4e60\u3002\\\\n\u591a\u5199\u4ee3\u7801\u662f\u8fdb\u6b65\u7684\u5173\u952e\u3002\\\\n\u7b97\u6cd5\u548c\u6570\u636e\u7ed3\u6784\u5f88\u91cd\u8981\u3002",
\n ""
\n ]
\n ]<\/p>\n
# \u521b\u5efa\u754c\u9762
\n with gr.Blocks(title="Qwen3-Reranker-0.6B", theme=gr.themes.Soft()) as demo:
\n gr.Markdown(description)<\/p>\n
with gr.Row():
\n with gr.Column(scale=2):
\n query_input = gr.Textbox(
\n label="\u67e5\u8be2\u6587\u672c",
\n placeholder="\u8f93\u5165\u4f60\u8981\u641c\u7d22\u7684\u95ee\u9898…",
\n lines=2
\n )<\/p>\n
documents_input = gr.Textbox(
\n label="\u6587\u6863\u5217\u8868",
\n placeholder="\u6bcf\u884c\u8f93\u5165\u4e00\u4e2a\u5019\u9009\u6587\u6863…",
\n lines=8
\n )<\/p>\n
instruction_input = gr.Textbox(
\n label="\u4efb\u52a1\u6307\u4ee4 (\u53ef\u9009)",
\n placeholder="\u4f8b\u5982: Given a web search query, retrieve relevant passages",
\n lines=2
\n )<\/p>\n
batch_size_slider = gr.Slider(
\n minimum=1,
\n maximum=32,
\n value=8,
\n step=1,
\n label="\u6279\u5904\u7406\u5927\u5c0f"
\n )<\/p>\n
submit_btn = gr.Button("\u5f00\u59cb\u91cd\u6392\u5e8f", variant="primary")<\/p>\n
with gr.Column(scale=3):
\n output_text = gr.Textbox(
\n label="\u6392\u5e8f\u7ed3\u679c",
\n lines=15,
\n interactive=False
\n )<\/p>\n
# \u793a\u4f8b
\n gr.Examples(
\n examples=examples,
\n inputs=[query_input, documents_input, instruction_input],
\n label="\u70b9\u51fb\u4f7f\u7528\u793a\u4f8b"
\n )<\/p>\n
# \u4e8b\u4ef6\u5904\u7406
\n submit_btn.click(
\n fn=rerank_documents,
\n inputs=[query_input, documents_input, instruction_input, batch_size_slider],
\n outputs=output_text
\n )<\/p>\n
# \u56de\u8f66\u952e\u63d0\u4ea4
\n query_input.submit(
\n fn=rerank_documents,
\n inputs=[query_input, documents_input, instruction_input, batch_size_slider],
\n outputs=output_text
\n )<\/p>\n
# \u7cfb\u7edf\u4fe1\u606f
\n with gr.Accordion("\u7cfb\u7edf\u4fe1\u606f", open=False):
\n gr.Markdown(f"""
\n **\u8fd0\u884c\u73af\u5883:**
\n – \u67b6\u6784: {platform.machine() if 'platform' in locals() else '\u672a\u77e5'}
\n – \u8bbe\u5907: {str(device) if device else '\u672a\u77e5'}
\n – \u6a21\u578b: Qwen3-Reranker-0.6B
\n – \u4e0a\u4e0b\u6587\u957f\u5ea6: 32K tokens
\n – \u652f\u6301\u8bed\u8a00: 100+ \u79cd
\n """)<\/p>\n
return demo<\/p>\n
if __name__ == "__main__":
\n # \u68c0\u67e5\u6a21\u578b\u662f\u5426\u52a0\u8f7d\u6210\u529f
\n if model is None:
\n logger.error("\u6a21\u578b\u52a0\u8f7d\u5931\u8d25，\u8bf7\u68c0\u67e5:")
\n logger.error("1. \u6a21\u578b\u6587\u4ef6\u662f\u5426\u5b58\u5728")
\n logger.error("2. \u4f9d\u8d56\u662f\u5426\u5b89\u88c5\u5b8c\u6574")
\n logger.error("3. \u8bbe\u5907\u5185\u5b58\u662f\u5426\u5145\u8db3")
\n sys.exit(1)<\/p>\n
# \u542f\u52a8\u670d\u52a1
\n demo = create_interface()
\n demo.launch(
\n server_name="0.0.0.0",
\n server_port=7860,
\n share=False,
\n show_error=True
\n )<\/p>\n
\u521b\u5efa\u4e00\u4e2aconfig.json\u6587\u4ef6\u6765\u4fdd\u5b58\u4e00\u4e9b\u8bbe\u7f6e：<\/p>\n
{
\n "model_name": "Qwen3-Reranker-0.6B",
\n "model_path": ".\/Qwen3-Reranker-0.6B",
\n "max_documents": 100,
\n "default_batch_size": 8,
\n "max_length": 512,
\n "device": "auto",
\n "arm_optimization": true,
\n "use_mps_if_available": true,
\n "float_precision": "fp16",
\n "language_support": ["en", "zh", "es", "fr", "de", "ja", "ko", "ru"]
\n}<\/p>\n
ARM\u8bbe\u5907（\u7279\u522b\u662fMac M\u7cfb\u5217）\u6709\u7edf\u4e00\u5185\u5b58\u67b6\u6784，\u6211\u4eec\u9700\u8981\u9488\u5bf9\u8fd9\u4e2a\u7279\u70b9\u8fdb\u884c\u4f18\u5316：<\/p>\n
# \u5728app.py\u4e2d\u6dfb\u52a0\u5185\u5b58\u4f18\u5316\u51fd\u6570
\ndef optimize_for_arm():
\n """ARM\u67b6\u6784\u5185\u5b58\u4f18\u5316"""
\n import psutil
\n import torch<\/p>\n
# \u83b7\u53d6\u7cfb\u7edf\u5185\u5b58
\n memory_info = psutil.virtual_memory()
\n total_memory_gb = memory_info.total \/ (1024**3)<\/p>\n
logger.info(f"\u7cfb\u7edf\u603b\u5185\u5b58: {total_memory_gb:.1f} GB")<\/p>\n
# \u6839\u636e\u5185\u5b58\u5927\u5c0f\u8c03\u6574\u914d\u7f6e
\n if total_memory_gb < 8:
\n # \u5c0f\u4e8e8GB\u5185\u5b58，\u4f7f\u7528\u66f4\u4fdd\u5b88\u7684\u8bbe\u7f6e
\n config = {
\n "batch_size": 4,
\n "max_documents": 50,
\n "use_fp16": False, # \u4f7f\u7528FP32\u8282\u7701\u663e\u5b58
\n "enable_cache": False
\n }
\n logger.info("\u5185\u5b58\u8f83\u5c0f，\u4f7f\u7528\u4f18\u5316\u914d\u7f6e")
\n elif total_memory_gb < 16:
\n # 8-16GB\u5185\u5b58
\n config = {
\n "batch_size": 8,
\n "max_documents": 100,
\n "use_fp16": True,
\n "enable_cache": True
\n }
\n logger.info("\u4e2d\u7b49\u5185\u5b58\u914d\u7f6e")
\n else:
\n # 16GB\u4ee5\u4e0a\u5185\u5b58
\n config = {
\n "batch_size": 16,
\n "max_documents": 100,
\n "use_fp16": True,
\n "enable_cache": True
\n }
\n logger.info("\u5927\u5185\u5b58\u914d\u7f6e")<\/p>\n
return config<\/p>\n
\u5bf9\u4e8eMac M\u7cfb\u5217\u82af\u7247，PyTorch\u652f\u6301MPS（Metal Performance Shaders）\u540e\u7aef：<\/p>\n
def setup_mps_acceleration():
\n """\u914d\u7f6eMPS\u52a0\u901f"""
\n if not torch.backends.mps.is_available():
\n logger.warning("MPS\u4e0d\u53ef\u7528，\u5c06\u4f7f\u7528CPU")
\n return False<\/p>\n
# \u542f\u7528MPS
\n torch.backends.mps.enabled = True<\/p>\n
# MPS\u7279\u5b9a\u4f18\u5316
\n if torch.backends.mps.is_built():
\n # \u8bbe\u7f6e\u5185\u5b58\u9650\u5236（\u907f\u514dOOM）
\n torch.mps.set_per_process_memory_fraction(0.7) # \u4f7f\u752870%\u5185\u5b58<\/p>\n
# \u542f\u7528MPS fallback（\u91cd\u8981！）
\n os.environ['PYTORCH_ENABLE_MPS_FALLBACK'] = '1'<\/p>\n
logger.info("MPS\u52a0\u901f\u5df2\u542f\u7528")
\n return True<\/p>\n
return False<\/p>\n
\u521b\u5efa\u4e00\u4e2a\u6027\u80fd\u76d1\u63a7\u811a\u672cmonitor.py：<\/p>\n
#!\/usr\/bin\/env python3
\n# \u6027\u80fd\u76d1\u63a7\u811a\u672c<\/p>\n
import time
\nimport psutil
\nimport torch<\/p>\n
def monitor_performance():
\n """\u76d1\u63a7\u7cfb\u7edf\u6027\u80fd"""
\n process = psutil.Process()<\/p>\n
while True:
\n # CPU\u4f7f\u7528\u7387
\n cpu_percent = psutil.cpu_percent(interval=1)<\/p>\n
# \u5185\u5b58\u4f7f\u7528
\n memory_info = process.memory_info()
\n memory_mb = memory_info.rss \/ 1024 \/ 1024<\/p>\n
# GPU\/MPS\u5185\u5b58（\u5982\u679c\u53ef\u7528）
\n if torch.backends.mps.is_available():
\n gpu_memory = torch.mps.current_allocated_memory() \/ 1024 \/ 1024
\n gpu_info = f"MPS\u5185\u5b58: {gpu_memory:.1f} MB"
\n else:
\n gpu_info = "MPS\u4e0d\u53ef\u7528"<\/p>\n
print(f"\\\\rCPU: {cpu_percent}% | \u5185\u5b58: {memory_mb:.1f} MB | {gpu_info}", end="")
\n time.sleep(2)<\/p>\n
if __name__ == "__main__":
\n try:
\n monitor_performance()
\n except KeyboardInterrupt:
\n print("\\\\n\u76d1\u63a7\u7ed3\u675f")<\/p>\n
\u521b\u5efa\u4e00\u4e2a\u5b8c\u6574\u7684\u90e8\u7f72\u811a\u672cdeploy_arm.sh：<\/p>\n
#!\/bin\/bash<\/p>\n
echo "=== Qwen3-Reranker-0.6B ARM\u90e8\u7f72\u811a\u672c ==="
\necho "\u9002\u7528\u4e8eMac M\u7cfb\u5217\u53caARM\u670d\u52a1\u5668"
\necho ""<\/p>\n
# \u68c0\u67e5\u76ee\u5f55
\nif [ ! -d "Qwen3-Reranker-0.6B" ]; then
\n echo "\u521b\u5efa\u9879\u76ee\u76ee\u5f55…"
\n mkdir -p Qwen3-Reranker-0.6B
\nfi<\/p>\n
cd Qwen3-Reranker-0.6B<\/p>\n
echo "\u6b65\u9aa41: \u68c0\u67e5Python\u73af\u5883…"
\npython3 –version
\nif [ $? -ne 0 ]; then
\n echo "\u9519\u8bef: Python3\u672a\u5b89\u88c5"
\n exit 1
\nfi<\/p>\n
echo "\u6b65\u9aa42: \u5b89\u88c5\u4f9d\u8d56…"
\necho "\u5b89\u88c5PyTorch…"
\npip3 install torch torchvision torchaudio<\/p>\n
echo "\u5b89\u88c5\u5176\u4ed6\u4f9d\u8d56…"
\npip3 install transformers>=4.51.0 gradio>=4.0.0 accelerate safetensors sentencepiece psutil<\/p>\n
echo "\u6b65\u9aa43: \u4e0b\u8f7d\u6a21\u578b\u6587\u4ef6…"
\nif [ ! -f "config.json" ]; then
\n echo "\u6a21\u578b\u6587\u4ef6\u4e0d\u5b58\u5728，\u5f00\u59cb\u4e0b\u8f7d…"<\/p>\n
# \u68c0\u67e5\u662f\u5426\u5b89\u88c5huggingface-hub
\n if ! command -v huggingface-cli &> \/dev\/null; then
\n pip3 install huggingface-hub
\n fi<\/p>\n
# \u4e0b\u8f7d\u6a21\u578b
\n python3 -c "
\nimport sys
\nfrom huggingface_hub import snapshot_download
\nimport os<\/p>\n
try:
\n print('\u5f00\u59cb\u4e0b\u8f7d\u6a21\u578b…')
\n snapshot_download(
\n repo_id='Qwen\/Qwen3-Reranker-0.6B',
\n local_dir='.',
\n local_dir_use_symlinks=False
\n )
\n print('\u4e0b\u8f7d\u5b8c\u6210!')
\nexcept Exception as e:
\n print(f'\u4e0b\u8f7d\u5931\u8d25: {e}')
\n print('\u8bf7\u624b\u52a8\u4e0b\u8f7d\u6a21\u578b\u6587\u4ef6\u5230\u5f53\u524d\u76ee\u5f55')
\n sys.exit(1)
\n"
\nelse
\n echo "\u6a21\u578b\u6587\u4ef6\u5df2\u5b58\u5728，\u8df3\u8fc7\u4e0b\u8f7d"
\nfi<\/p>\n
echo "\u6b65\u9aa44: \u521b\u5efa\u914d\u7f6e\u6587\u4ef6…"
\ncat > app.py << 'EOF'
\n# \u8fd9\u91cc\u7c98\u8d34\u4e0a\u9762\u521b\u5efa\u7684app.py\u5185\u5bb9
\nEOF<\/p>\n
cat > start.sh << 'EOF'
\n#!\/bin\/bash
\n# \u8fd9\u91cc\u7c98\u8d34\u4e0a\u9762\u521b\u5efa\u7684start.sh\u5185\u5bb9
\nEOF<\/p>\n
chmod +x start.sh<\/p>\n
cat > config.json << 'EOF'
\n{
\n "model_name": "Qwen3-Reranker-0.6B",
\n "model_path": ".",
\n "max_documents": 100,
\n "default_batch_size": 8,
\n "arm_optimized": true
\n}
\nEOF<\/p>\n
echo "\u6b65\u9aa45: \u9a8c\u8bc1\u5b89\u88c5…"
\necho "\u8fd0\u884c\u5feb\u901f\u6d4b\u8bd5…"
\npython3 -c "
\nimport torch
\nprint(f'PyTorch\u7248\u672c: {torch.__version__}')
\nprint(f'\u8bbe\u5907\u53ef\u7528\u6027:')
\nprint(f' CPU: \u662f')
\nprint(f' CUDA: {torch.cuda.is_available()}')
\nprint(f' MPS: {torch.backends.mps.is_available()}')
\n"<\/p>\n
echo ""
\necho "=== \u90e8\u7f72\u5b8c\u6210! ==="
\necho ""
\necho "\u542f\u52a8\u670d\u52a1:"
\necho " .\/start.sh"
\necho ""
\necho "\u8bbf\u95ee\u5730\u5740:"
\necho " http:\/\/localhost:7860"
\necho ""
\necho "\u505c\u6b62\u670d\u52a1: Ctrl+C"<\/p>\n
\u8fd0\u884c\u90e8\u7f72\u811a\u672c：<\/p>\n
# \u7ed9\u811a\u672c\u6dfb\u52a0\u6267\u884c\u6743\u9650
\nchmod +x deploy_arm.sh<\/p>\n
# \u8fd0\u884c\u90e8\u7f72\u811a\u672c
\n.\/deploy_arm.sh<\/p>\n
\u90e8\u7f72\u5b8c\u6210\u540e，\u542f\u52a8\u670d\u52a1：<\/p>\n
cd Qwen3-Reranker-0.6B
\n.\/start.sh<\/p>\n
\u4f60\u5e94\u8be5\u770b\u5230\u7c7b\u4f3c\u8fd9\u6837\u7684\u8f93\u51fa：<\/p>\n
\u6b63\u5728\u542f\u52a8 Qwen3-Reranker-0.6B \u670d\u52a1…
\nPython\u7248\u672c: 3.10
\n\u68c0\u67e5\u4f9d\u8d56…
\n\u542f\u52a8Web\u670d\u52a1…
\n\u670d\u52a1\u5c06\u5728 http:\/\/localhost:7860 \u53ef\u7528
\nRunning on local URL: http:\/\/0.0.0.0:7860<\/p>\n
\u6253\u5f00\u6d4f\u89c8\u5668，\u8bbf\u95ee http:\/\/localhost:7860，\u4f60\u5e94\u8be5\u80fd\u770b\u5230Web\u754c\u9762\u3002<\/p>\n
\u5728Web\u754c\u9762\u4e2d，\u5c1d\u8bd5\u8f93\u5165：<\/p>\n
\u67e5\u8be2\u6587\u672c：<\/p>\n
\u5982\u4f55\u5b66\u4e60Python\u7f16\u7a0b？<\/p>\n
\u6587\u6863\u5217\u8868：<\/p>\n
Python\u662f\u4e00\u79cd\u9ad8\u7ea7\u7f16\u7a0b\u8bed\u8a00\u3002
\n\u673a\u5668\u5b66\u4e60\u9700\u8981\u6570\u5b66\u57fa\u7840\u3002
\nPython\u9002\u5408\u521d\u5b66\u8005，\u8bed\u6cd5\u7b80\u6d01\u3002
\nJava\u662f\u53e6\u4e00\u79cd\u7f16\u7a0b\u8bed\u8a00\u3002
\n\u591a\u5199\u4ee3\u7801\u662f\u5b66\u4e60\u7f16\u7a0b\u7684\u6700\u597d\u65b9\u6cd5\u3002<\/p>\n
\u70b9\u51fb"\u5f00\u59cb\u91cd\u6392\u5e8f"，\u4f60\u4f1a\u770b\u5230\u6587\u6863\u6309\u7167\u76f8\u5173\u6027\u91cd\u65b0\u6392\u5e8f\u3002<\/p>\n
\u95ee\u9898： \u542f\u52a8\u65f6\u63d0\u793a\u6a21\u578b\u52a0\u8f7d\u5931\u8d25<\/p>\n
\u89e3\u51b3\u6b65\u9aa4：<\/p>\n
ls -lh Qwen3-Reranker-0.6B\/
\n# \u5e94\u8be5\u770b\u5230\u81f3\u5c111.2GB\u7684model.safetensors\u6587\u4ef6<\/p>\n
pip3 show transformers
\n# \u786e\u4fdd\u7248\u672c >= 4.51.0<\/p>\n
rm -rf ~\/.cache\/huggingface\/hub
\n# \u91cd\u65b0\u4e0b\u8f7d\u6a21\u578b<\/p>\n
\u95ee\u9898： \u8fd0\u884c\u65f6\u62a5\u5185\u5b58\u9519\u8bef<\/p>\n
\u89e3\u51b3\u65b9\u6848：<\/p>\n
\u51cf\u5c0f\u6279\u5904\u7406\u5927\u5c0f：<\/p>\n
\u4f7f\u7528CPU\u6a21\u5f0f（\u5982\u679cMPS\u5185\u5b58\u4e0d\u8db3）：<\/p>\n<\/li>\n
# \u5728app.py\u4e2d\u5f3a\u5236\u4f7f\u7528CPU
\ndevice = torch.device("cpu")<\/p>\n
# \u8fd0\u884c\u76d1\u63a7\u811a\u672c
\npython3 monitor.py<\/p>\n
\u5982\u679c\u89c9\u5f97\u901f\u5ea6\u4e0d\u591f\u5feb，\u53ef\u4ee5\u5c1d\u8bd5：<\/p>\n
# \u52a0\u8f7d\u65f6\u4f7f\u75288\u4f4d\u91cf\u5316
\nmodel = AutoModelForSequenceClassification.from_pretrained(
\n model_path,
\n load_in_8bit=True, # 8\u4f4d\u91cf\u5316
\n device_map="auto"
\n)<\/p>\n
# \u8bbe\u7f6ePyTorch\u7ebf\u7a0b\u6570
\nexport OMP_NUM_THREADS=4
\nexport MKL_NUM_THREADS=4<\/p>\n
\u95ee\u9898： 7860\u7aef\u53e3\u88ab\u5360\u7528<\/p>\n
\u89e3\u51b3\u65b9\u6848：<\/p>\n
# \u67e5\u770b\u7aef\u53e3\u5360\u7528
\nlsof -i:7860<\/p>\n
# \u5982\u679c\u88ab\u5360\u7528，\u53ef\u4ee5\u4fee\u6539\u7aef\u53e3
\n# \u5728start.sh\u4e2d\u4fee\u6539：
\nexport GRADIO_SERVER_PORT=7861<\/p>\n
\u6216\u8005\u76f4\u63a5\u4fee\u6539app.py\u4e2d\u7684\u542f\u52a8\u53c2\u6570：<\/p>\n
demo.launch(server_port=7861) # \u4f7f\u7528\u5176\u4ed6\u7aef\u53e3<\/p>\n
\u5047\u8bbe\u4f60\u6709\u4e00\u4e2a\u7b80\u5355\u7684\u6587\u6863\u68c0\u7d22\u7cfb\u7edf，\u53ef\u4ee5\u8fd9\u6837\u96c6\u6210\u91cd\u6392\u5e8f：<\/p>\n
import requests
\nimport json<\/p>\n
class DocumentSearch:
\n def __init__(self, reranker_url="http:\/\/localhost:7860"):
\n self.reranker_url = reranker_url
\n self.documents = [] # \u4f60\u7684\u6587\u6863\u5e93<\/p>\n
def search(self, query, top_k=10):
\n """\u641c\u7d22\u5e76\u91cd\u6392\u5e8f"""
\n # 1. \u521d\u6b65\u68c0\u7d22（\u7b80\u5355\u5173\u952e\u8bcd\u5339\u914d）
\n candidates = self.initial_search(query)<\/p>\n
# 2. \u4f7f\u7528\u91cd\u6392\u5e8f\u6a21\u578b\u4f18\u5316\u7ed3\u679c
\n if candidates:
\n sorted_docs = self.rerank(query, candidates)
\n return sorted_docs[:top_k]
\n return []<\/p>\n
def initial_search(self, query):
\n """\u521d\u6b65\u68c0\u7d22（\u8fd9\u91cc\u7b80\u5316\u5b9e\u73b0）"""
\n # \u5b9e\u9645\u4e2d\u53ef\u80fd\u4f7f\u7528Elasticsearch\u7b49
\n results = []
\n for doc in self.documents:
\n if query.lower() in doc.lower():
\n results.append(doc)
\n return results<\/p>\n
def rerank(self, query, documents):
\n """\u8c03\u7528\u91cd\u6392\u5e8f\u670d\u52a1"""
\n try:
\n # \u51c6\u5907\u8bf7\u6c42\u6570\u636e
\n data = {
\n "data": [
\n query,
\n "\\\\n".join(documents),
\n "", # \u53ef\u9009\u6307\u4ee4
\n 8 # \u6279\u5904\u7406\u5927\u5c0f
\n ]
\n }<\/p>\n
# \u53d1\u9001\u8bf7\u6c42
\n response = requests.post(
\n f"{self.reranker_url}\/api\/predict",
\n json=data,
\n timeout=10
\n )<\/p>\n
if response.status_code == 200:
\n result = response.json()
\n # \u89e3\u6790\u7ed3\u679c…
\n return self.parse_rerank_result(result, documents)
\n else:
\n print(f"\u8bf7\u6c42\u5931\u8d25: {response.status_code}")
\n return documents<\/p>\n
except Exception as e:
\n print(f"\u91cd\u6392\u5e8f\u51fa\u9519: {e}")
\n return documents<\/p>\n
def parse_rerank_result(self, result, original_docs):
\n """\u89e3\u6790\u91cd\u6392\u5e8f\u7ed3\u679c"""
\n # \u6839\u636eAPI\u8fd4\u56de\u683c\u5f0f\u89e3\u6790
\n # \u8fd9\u91cc\u9700\u8981\u6839\u636e\u5b9e\u9645API\u8c03\u6574
\n return original_docs # \u7b80\u5316\u8fd4\u56de<\/p>\n
\u5982\u679c\u4f60\u9700\u8981\u5904\u7406\u5927\u91cf\u67e5\u8be2，\u53ef\u4ee5\u521b\u5efa\u6279\u91cf\u5904\u7406\u811a\u672c：<\/p>\n
import pandas as pd
\nfrom tqdm import tqdm<\/p>\n
def batch_rerank(input_csv, output_csv):
\n """\u6279\u91cf\u91cd\u6392\u5e8f"""
\n # \u8bfb\u53d6\u6570\u636e
\n df = pd.read_csv(input_csv)<\/p>\n
results = []<\/p>\n
for _, row in tqdm(df.iterrows(), total=len(df)):
\n query = row['query']
\n documents = row['documents'].split('|') # \u5047\u8bbe\u6587\u6863\u7528|\u5206\u9694<\/p>\n
# \u8c03\u7528\u91cd\u6392\u5e8f
\n sorted_docs = rerank_documents(query, documents)<\/p>\n
# \u4fdd\u5b58\u7ed3\u679c
\n results.append({
\n 'query': query,
\n 'original_docs': '|'.join(documents),
\n 'sorted_docs': '|'.join(sorted_docs),
\n 'top_result': sorted_docs[0] if sorted_docs else ''
\n })<\/p>\n
# \u4fdd\u5b58\u7ed3\u679c
\n result_df = pd.DataFrame(results)
\n result_df.to_csv(output_csv, index=False)
\n print(f"\u5904\u7406\u5b8c\u6210，\u7ed3\u679c\u4fdd\u5b58\u5230 {output_csv}")<\/p>\n
\u901a\u8fc7\u8fd9\u7bc7\u6559\u7a0b，\u4f60\u5e94\u8be5\u5df2\u7ecf\u6210\u529f\u5728ARM\u67b6\u6784\u8bbe\u5907\u4e0a\u90e8\u7f72\u4e86Qwen3-Reranker-0.6B\u6a21\u578b\u3002\u6211\u4eec\u6765\u56de\u987e\u4e00\u4e0b\u5173\u952e\u70b9：<\/p>\n
\u90e8\u7f72\u8981\u70b9\u603b\u7ed3：<\/p>\n
ARM\u7279\u6709\u4f18\u52bf：<\/p>\n
\u6027\u80fd\u5efa\u8bae：<\/p>\n
\u4e0b\u4e00\u6b65\u5efa\u8bae：<\/p>\n
\u8fd9\u4e2a\u90e8\u7f72\u4e0d\u4ec5\u8ba9\u4f60\u80fd\u5728\u672c\u5730\u4f7f\u7528\u5f3a\u5927\u7684\u91cd\u6392\u5e8f\u529f\u80fd，\u4e5f\u4e3a\u7406\u89e3\u5982\u4f55\u5728ARM\u67b6\u6784\u4e0a\u90e8\u7f72AI\u6a21\u578b\u63d0\u4f9b\u4e86\u5b9e\u8df5\u7ecf\u9a8c\u3002\u65e0\u8bba\u662f\u4e2a\u4eba\u9879\u76ee\u8fd8\u662f\u539f\u578b\u5f00\u53d1，\u672c\u5730\u90e8\u7f72\u90fd\u80fd\u63d0\u4f9b\u5feb\u901f\u8fed\u4ee3\u548c\u9690\u79c1\u4fdd\u62a4\u7684\u4f18\u52bf\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":"
Qwen3-Reranker-0.6B\u90e8\u7f72\u6559\u7a0b：ARM\u67b6\u6784\u670d\u52a1\u5668（\u5982Mac M2\/M3）\u9002\u914d\u6307\u5357
\n1. \u5f15\u8a00：\u4e3a\u4ec0\u4e48\u8981\u5728ARM\u67b6\u6784\u4e0a\u90e8\u7f72\u91cd\u6392\u5e8f\u6a21\u578b？
\n\u5982\u679c\u4f60\u624b\u5934\u6709\u4e00\u53f0MacBook，\u7279\u522b\u662f\u642d\u8f7d\u4e86M2\u6216M3\u82af\u7247\u7684\u578b\u53f7，\u4f60\u53ef\u80fd\u4f1a\u53d1\u73b0\u5f88\u591aAI\u6a21\u578b\u90e8\u7f72\u6559\u7a0b\u90fd\u662f\u9488\u5bf9x86\u67b6\u6784\u7684，\u7528\u8d77\u6765\u603b\u662f\u4e0d\u592a\u987a\u624b\u3002\u4eca\u5929，\u6211\u4eec\u5c31\u6765\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\u3002
\nQwen3-Reranker<\/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":[7591,6714,9255],"topic":[],"class_list":["post-82829","post","type-post","status-publish","format-standard","hentry","category-server","tag-ai","tag-arm","tag-9255"],"yoast_head":"\n