\u5982\u679c\u4f60\u5728\u98df\u54c1\u670d\u52a1\u6216\u96f6\u552e\u884c\u4e1a\u5de5\u4f5c，\u6bcf\u5929\u9762\u5bf9\u8d27\u67b6\u68c0\u67e5\u3001\u5546\u54c1\u8bc6\u522b\u3001\u5408\u89c4\u5ba1\u6838\u8fd9\u4e9b\u7e41\u7410\u4efb\u52a1，\u4e00\u5b9a\u60f3\u8fc7：\u80fd\u4e0d\u80fd\u6709\u4e2aAI\u52a9\u624b，\u770b\u4e00\u773c\u56fe\u7247\u5c31\u80fd\u544a\u8bc9\u6211\u8d27\u67b6\u7f3a\u4e86\u4ec0\u4e48\u3001\u5546\u54c1\u6446\u653e\u5bf9\u4e0d\u5bf9\u3001\u751a\u81f3\u5e2e\u6211\u5206\u6790\u987e\u5ba2\u52a8\u7ebf？<\/p>\n
\u73b0\u5728，\u8fd9\u4e2a\u52a9\u624b\u6765\u4e86\u2014\u2014Ostrakon-VL-8B，\u4e00\u4e2a\u4e13\u95e8\u4e3a\u96f6\u552e\u573a\u666f\u8bad\u7ec3\u7684\u591a\u6a21\u6001\u5927\u6a21\u578b\u3002\u5b83\u80fd\u770b\u61c2\u56fe\u7247\u3001\u7406\u89e3\u89c6\u9891，\u8fd8\u80fd\u50cf\u4e13\u5bb6\u4e00\u6837\u56de\u7b54\u4f60\u7684\u95ee\u9898\u3002<\/p>\n
\u4f46\u4eca\u5929\u6211\u4eec\u8981\u804a\u7684，\u4e0d\u53ea\u662f\u8fd9\u4e2a\u6a21\u578b\u6709\u591a\u5389\u5bb3，\u800c\u662f\u4e00\u4e2a\u66f4\u5b9e\u9645\u7684\u95ee\u9898：\u600e\u4e48\u5728ARM\u67b6\u6784\u7684\u670d\u52a1\u5668\u4e0a，\u6bd4\u5982NVIDIA Grace\u8fd9\u6837\u7684\u65b0\u5e73\u53f0\u4e0a，\u5feb\u901f\u628a\u5b83\u90e8\u7f72\u8d77\u6765，\u5e76\u4e14\u8ba9\u5b83\u8dd1\u5f97\u53c8\u5feb\u53c8\u7a33？<\/p>\n
\u4f60\u53ef\u80fd\u542c\u8bf4\u8fc7，\u5f88\u591aAI\u6a21\u578b\u5bf9\u786c\u4ef6\u5e73\u53f0\u5f88\u6311\u5254，\u6362\u4e2a\u67b6\u6784\u5c31\u53ef\u80fd\u51fa\u5404\u79cd\u95ee\u9898\u3002\u522b\u62c5\u5fc3，\u8fd9\u7bc7\u6587\u7ae0\u5c31\u662f\u4e3a\u4f60\u51c6\u5907\u7684\u3002\u6211\u4f1a\u5e26\u4f60\u4e00\u6b65\u6b65\u5b8c\u6210Ostrakon-VL-8B\u5728ARM\u670d\u52a1\u5668\u4e0a\u7684\u90e8\u7f72\u3001\u9a8c\u8bc1\u548c\u8c03\u4f18，\u8ba9\u4f60\u7528\u6700\u5c11\u7684\u6298\u817e，\u628a\u8fd9\u4e2a\u96f6\u552eAI\u4e13\u5bb6\u8bf7\u5230\u4f60\u7684\u670d\u52a1\u5668\u4e0a\u3002<\/p>\n
\u7b80\u5355\u8bf4，Ostrakon-VL-8B\u662f\u4e00\u4e2a\u4e13\u95e8\u4e3a\u98df\u54c1\u670d\u52a1\u548c\u96f6\u552e\u5546\u5e97（FSRS）\u573a\u666f\u8bbe\u8ba1\u7684AI\u6a21\u578b\u3002\u5b83\u57fa\u4e8eQwen3-VL-8B\u6784\u5efa，\u4f46\u7ecf\u8fc7\u4e86\u4e13\u95e8\u7684\u8bad\u7ec3，\u5728\u96f6\u552e\u76f8\u5173\u4efb\u52a1\u4e0a\u8868\u73b0\u7279\u522b\u51fa\u8272\u3002<\/p>\n
\u60f3\u8c61\u4e00\u4e0b，\u4f60\u62cd\u4e00\u5f20\u8d85\u5e02\u8d27\u67b6\u7684\u7167\u7247\u53d1\u7ed9\u5b83，\u5b83\u80fd\u544a\u8bc9\u4f60：<\/p>\n
\u6700\u5389\u5bb3\u7684\u662f，\u867d\u7136\u5b83\u53ea\u670980\u4ebf\u53c2\u6570（\u76f8\u5bf9\u8f83\u5c0f），\u4f46\u5728\u96f6\u552e\u573a\u666f\u7684\u6d4b\u8bd5\u4e2d，\u8868\u73b0\u751a\u81f3\u8d85\u8fc7\u4e86\u67d0\u4e9b2350\u4ebf\u53c2\u6570\u7684\u901a\u7528\u5927\u6a21\u578b\u3002\u8fd9\u5c31\u597d\u6bd4\u4e00\u4e2a\u4e13\u95e8\u7814\u7a76\u96f6\u552e\u7684\u4e13\u5bb6，\u867d\u7136\u77e5\u8bc6\u9762\u6ca1\u90a3\u4e48\u5e7f，\u4f46\u5728\u81ea\u5df1\u4e13\u4e1a\u9886\u57df\u6bd4\u4ec0\u4e48\u90fd\u61c2\u7684\u901a\u624d\u66f4\u5389\u5bb3\u3002<\/p>\n
\u4e13\u7cbe\u80dc\u8fc7\u901a\u624d：\u901a\u7528\u5927\u6a21\u578b\u4ec0\u4e48\u90fd\u77e5\u9053\u4e00\u70b9，\u4f46\u5728\u5177\u4f53\u884c\u4e1a\u95ee\u9898\u4e0a\u5f80\u5f80\u4e0d\u591f\u6df1\u5165\u3002Ostrakon-VL\u4e13\u95e8\u9488\u5bf9\u96f6\u552e\u573a\u666f\u8bad\u7ec3，\u7406\u89e3\u884c\u4e1a\u7279\u6709\u7684\u7ec6\u8282\u548c\u9700\u6c42\u3002<\/p>\n
\u6548\u7387\u66f4\u9ad8：8B\u7684\u6a21\u578b\u89c4\u6a21\u610f\u5473\u7740\u5bf9\u786c\u4ef6\u8981\u6c42\u66f4\u4f4e，\u63a8\u7406\u901f\u5ea6\u66f4\u5feb，\u90e8\u7f72\u6210\u672c\u66f4\u4fbf\u5b9c\u3002\u5728ARM\u670d\u52a1\u5668\u4e0a，\u8fd9\u4e2a\u4f18\u52bf\u4f1a\u66f4\u52a0\u660e\u663e\u3002<\/p>\n
\u5f00\u6e90\u514d\u8d39：\u5b8c\u5168\u5f00\u6e90，\u4f60\u53ef\u4ee5\u81ea\u7531\u4f7f\u7528\u3001\u4fee\u6539，\u4e0d\u7528\u62c5\u5fc3\u6388\u6743\u8d39\u7528\u3002<\/p>\n
\u591a\u6a21\u6001\u80fd\u529b：\u4e0d\u4ec5\u80fd\u5904\u7406\u6587\u5b57，\u8fd8\u80fd\u770b\u61c2\u56fe\u7247\u3001\u89c6\u9891，\u652f\u6301\u591a\u56fe\u8f93\u5165，\u8fd9\u5728\u96f6\u552e\u573a\u666f\u4e2d\u7279\u522b\u6709\u7528\u2014\u2014\u6bd4\u5982\u540c\u65f6\u770b\u8d27\u67b6\u5168\u666f\u548c\u5546\u54c1\u7279\u5199\u3002<\/p>\n
\u5982\u679c\u4f60\u4e60\u60ef\u4e86x86\u670d\u52a1\u5668，\u7b2c\u4e00\u6b21\u5728ARM\u67b6\u6784\u4e0a\u90e8\u7f72AI\u6a21\u578b\u53ef\u80fd\u4f1a\u6709\u70b9\u4e0d\u4e60\u60ef\u3002\u4f46\u522b\u62c5\u5fc3，ARM\u67b6\u6784\u5176\u5b9e\u6709\u5f88\u591a\u4f18\u52bf：<\/p>\n
\u80fd\u6548\u6bd4\u66f4\u9ad8：\u540c\u6837\u7684\u6027\u80fd\u4e0b，ARM\u5904\u7406\u5668\u901a\u5e38\u66f4\u7701\u7535，\u8fd9\u5bf9\u9700\u89817×24\u5c0f\u65f6\u8fd0\u884c\u7684AI\u670d\u52a1\u5f88\u91cd\u8981\u3002<\/p>\n
\u6210\u672c\u53ef\u80fd\u66f4\u4f4e：\u5f88\u591aARM\u670d\u52a1\u5668\u65b9\u6848\u5728\u786c\u4ef6\u6210\u672c\u4e0a\u66f4\u6709\u4f18\u52bf\u3002<\/p>\n
\u751f\u6001\u6b63\u5728\u6210\u719f：\u968f\u7740NVIDIA Grace\u3001Ampere Altra\u7b49ARM\u670d\u52a1\u5668\u7684\u63a8\u51fa，AI\u8f6f\u4ef6\u751f\u6001\u6b63\u5728\u5feb\u901f\u5b8c\u5584\u3002<\/p>\n
\u4f46\u4e5f\u8981\u6ce8\u610f：\u6709\u4e9bAI\u6846\u67b6\u548c\u5e93\u5bf9ARM\u7684\u652f\u6301\u8fd8\u5728\u5b8c\u5584\u4e2d，\u53ef\u80fd\u4f1a\u9047\u5230\u4e00\u4e9b\u517c\u5bb9\u6027\u95ee\u9898\u3002\u8fd9\u5c31\u662f\u4e3a\u4ec0\u4e48\u6211\u4eec\u9700\u8981\u505a\u4e13\u95e8\u7684\u9a8c\u8bc1\u548c\u8c03\u4f18\u3002<\/p>\n
\u5728\u5f00\u59cb\u90e8\u7f72\u524d，\u5148\u68c0\u67e5\u4e00\u4e0b\u4f60\u7684ARM\u670d\u52a1\u5668\u73af\u5883：<\/p>\n
# \u68c0\u67e5\u7cfb\u7edf\u67b6\u6784
\nuname -m
\n# \u5e94\u8be5\u663e\u793a aarch64 \u6216 arm64<\/p>\n
# \u68c0\u67e5Python\u7248\u672c
\npython3 –version
\n# \u5efa\u8bae Python 3.9 \u6216\u66f4\u9ad8<\/p>\n
# \u68c0\u67e5CUDA（\u5982\u679c\u6709NVIDIA GPU）
\nnvidia-smi
\n# \u786e\u8ba4\u9a71\u52a8\u548cCUDA\u7248\u672c<\/p>\n
# \u68c0\u67e5\u5185\u5b58
\nfree -h
\n# Ostrakon-VL-8B\u9700\u8981\u81f3\u5c1116GB\u5185\u5b58（\u7eafCPU）\u62168GB\u663e\u5b58（GPU\u52a0\u901f）<\/p>\n
\u5982\u679c\u4f60\u7684\u670d\u52a1\u5668\u662fNVIDIA Grace\u5e73\u53f0，\u8fd8\u9700\u8981\u786e\u8ba4：<\/p>\n
\u6211\u4eec\u4f7f\u7528vLLM\u6765\u90e8\u7f72\u6a21\u578b，\u8fd9\u662f\u76ee\u524d\u6700\u6d41\u884c\u7684\u9ad8\u6027\u80fd\u63a8\u7406\u6846\u67b6\u4e4b\u4e00，\u5bf9ARM\u67b6\u6784\u6709\u8f83\u597d\u7684\u652f\u6301\u3002\u7136\u540e\u7528Chainlit\u505a\u4e00\u4e2a\u7b80\u5355\u7684\u524d\u7aef，\u65b9\u4fbf\u6d4b\u8bd5\u548c\u4ea4\u4e92\u3002<\/p>\n
\u7b2c\u4e00\u6b65：\u51c6\u5907\u73af\u5883<\/p>\n
# \u521b\u5efa\u865a\u62df\u73af\u5883（\u63a8\u8350）
\npython3 -m venv ostrakon_env
\nsource ostrakon_env\/bin\/activate<\/p>\n
# \u5b89\u88c5\u57fa\u7840\u4f9d\u8d56
\npip install torch torchvision torchaudio –index-url https:\/\/download.pytorch.org\/whl\/cpu
\n# \u5982\u679c\u662fNVIDIA Grace，\u4f7f\u7528\u5bf9\u5e94\u7684CUDA\u7248\u672c
\n# pip install torch torchvision torchaudio –index-url https:\/\/download.pytorch.org\/whl\/cu118<\/p>\n
# \u5b89\u88c5vLLM
\npip install vllm<\/p>\n
# \u5b89\u88c5Chainlit
\npip install chainlit<\/p>\n
\u7b2c\u4e8c\u6b65：\u4e0b\u8f7d\u6a21\u578b<\/p>\n
Ostrakon-VL-8B\u7684\u6a21\u578b\u6587\u4ef6\u53ef\u4ee5\u4eceHugging Face\u83b7\u53d6：<\/p>\n
# \u65b9\u6cd51：\u76f4\u63a5\u4f7f\u7528vLLM\u81ea\u52a8\u4e0b\u8f7d（\u63a8\u8350）
\n# \u5728\u4ee3\u7801\u4e2d\u6307\u5b9a\u6a21\u578b\u8def\u5f84\u5373\u53ef，vLLM\u4f1a\u81ea\u52a8\u5904\u7406<\/p>\n
# \u65b9\u6cd52：\u624b\u52a8\u4e0b\u8f7d（\u5982\u679c\u7f51\u7edc\u73af\u5883\u9700\u8981）
\n# git lfs install
\n# git clone https:\/\/huggingface.co\/Salesforce\/Ostrakon-VL-8B<\/p>\n
\u7b2c\u4e09\u6b65：\u7f16\u5199\u542f\u52a8\u811a\u672c<\/p>\n
\u521b\u5efa\u4e00\u4e2aserve_ostrakon.py\u6587\u4ef6：<\/p>\n
from vllm import LLM, SamplingParams
\nimport argparse<\/p>\n
def main():
\n parser = argparse.ArgumentParser()
\n parser.add_argument("–model", type=str, default="Salesforce\/Ostrakon-VL-8B")
\n parser.add_argument("–tensor-parallel-size", type=int, default=1)
\n parser.add_argument("–gpu-memory-utilization", type=float, default=0.9)
\n parser.add_argument("–max-model-len", type=int, default=4096)
\n parser.add_argument("–port", type=int, default=8000)
\n args = parser.parse_args()<\/p>\n
# \u521d\u59cb\u5316\u6a21\u578b
\n print(f"\u6b63\u5728\u52a0\u8f7d\u6a21\u578b: {args.model}")
\n llm = LLM(
\n model=args.model,
\n tensor_parallel_size=args.tensor_parallel_size,
\n gpu_memory_utilization=args.gpu_memory_utilization,
\n max_model_len=args.max_model_len,
\n trust_remote_code=True # \u91cd\u8981：Ostrakon\u9700\u8981\u8fd9\u4e2a\u53c2\u6570
\n )<\/p>\n
print(f"\u6a21\u578b\u52a0\u8f7d\u6210\u529f！\u670d\u52a1\u8fd0\u884c\u5728\u7aef\u53e3 {args.port}")
\n print("\u53ef\u4ee5\u4f7f\u7528Chainlit\u6216\u76f4\u63a5\u8c03\u7528API\u8fdb\u884c\u6d4b\u8bd5")<\/p>\n
if __name__ == "__main__":
\n main()<\/p>\n
\u7b2c\u56db\u6b65：\u542f\u52a8\u670d\u52a1<\/p>\n
# \u5982\u679c\u662f\u7eafCPU\u73af\u5883（ARM\u670d\u52a1\u5668\u5e38\u89c1）
\npython serve_ostrakon.py –model Salesforce\/Ostrakon-VL-8B<\/p>\n
# \u5982\u679c\u6709GPU\u52a0\u901f
\npython serve_ostrakon.py –model Salesforce\/Ostrakon-VL-8B –tensor-parallel-size 1<\/p>\n
# \u4fdd\u5b58\u65e5\u5fd7\u5230\u6587\u4ef6\u65b9\u4fbf\u67e5\u770b
\npython serve_ostrakon.py 2>&1 | tee \/root\/workspace\/llm.log<\/p>\n
\u90e8\u7f72\u5b8c\u6210\u540e，\u9700\u8981\u786e\u8ba4\u4e00\u5207\u6b63\u5e38：<\/p>\n
# \u67e5\u770b\u65e5\u5fd7\u6587\u4ef6
\ncat \/root\/workspace\/llm.log<\/p>\n
\u5982\u679c\u770b\u5230\u7c7b\u4f3c\u4e0b\u9762\u7684\u8f93\u51fa，\u8bf4\u660e\u90e8\u7f72\u6210\u529f：<\/p>\n
\u6b63\u5728\u52a0\u8f7d\u6a21\u578b: Salesforce\/Ostrakon-VL-8B
\nLoading model weights…
\nModel loaded successfully!
\n\u670d\u52a1\u8fd0\u884c\u5728\u7aef\u53e3 8000<\/p>\n
\u5e38\u89c1\u95ee\u9898\u6392\u67e5：<\/p>\n
\u5185\u5b58\u4e0d\u8db3：\u5982\u679cARM\u670d\u52a1\u5668\u5185\u5b58\u8f83\u5c0f，\u53ef\u4ee5\u5c1d\u8bd5\u91cf\u5316\u7248\u672c<\/p>\n
python serve_ostrakon.py –model Salesforce\/Ostrakon-VL-8B-INT4\n <\/li>\n
CUDA\u4e0d\u517c\u5bb9：\u786e\u4fddPyTorch\u548cCUDA\u7248\u672c\u5339\u914d<\/p>\n
# \u68c0\u67e5CUDA\u7248\u672c
\nnvcc –version<\/p>\n
# \u5b89\u88c5\u5bf9\u5e94\u7248\u672c\u7684PyTorch
\npip install torch==2.1.0 torchvision==0.16.0 torchaudio==2.1.0 –index-url https:\/\/download.pytorch.org\/whl\/cu118\n <\/li>\n
\u6a21\u578b\u4e0b\u8f7d\u5931\u8d25：\u8bbe\u7f6e\u955c\u50cf\u6216\u624b\u52a8\u4e0b\u8f7d<\/p>\n
# \u8bbe\u7f6eHF\u955c\u50cf
\nexport HF_ENDPOINT=https:\/\/hf-mirror.com\n <\/li>\n
Chainlit\u662f\u4e00\u4e2a\u4e13\u95e8\u4e3aAI\u5e94\u7528\u8bbe\u8ba1\u7684\u804a\u5929\u754c\u9762\u6846\u67b6，\u5b83\u6709\u51e0\u4e2a\u4f18\u70b9：<\/p>\n
\u521b\u5efa\u4e00\u4e2aapp.py\u6587\u4ef6：<\/p>\n
import chainlit as cl
\nfrom vllm import LLM, SamplingParams
\nimport base64
\nfrom PIL import Image
\nimport io<\/p>\n
# \u521d\u59cb\u5316vLLM\u5ba2\u6237\u7aef（\u5047\u8bbe\u670d\u52a1\u8fd0\u884c\u5728\u672c\u57308000\u7aef\u53e3）
\n# \u5b9e\u9645\u90e8\u7f72\u65f6，\u53ef\u4ee5\u76f4\u63a5\u5728\u540c\u4e00\u4e2a\u8fdb\u7a0b\u4e2d\u52a0\u8f7d\u6a21\u578b<\/p>\n
@cl.on_chat_start
\nasync def start():
\n # \u8fd9\u91cc\u53ef\u4ee5\u521d\u59cb\u5316\u4e00\u4e9b\u72b6\u6001
\n await cl.Message(content="\u4f60\u597d！\u6211\u662fOstrakon-VL\u96f6\u552e\u52a9\u624b，\u53ef\u4ee5\u4e0a\u4f20\u56fe\u7247\u5e76\u63d0\u95ee\u3002").send()<\/p>\n
@cl.on_message
\nasync def main(message: cl.Message):
\n # \u68c0\u67e5\u662f\u5426\u6709\u56fe\u7247
\n image_files = [file for file in message.elements if "image" in file.mime]<\/p>\n
if not image_files:
\n await cl.Message(content="\u8bf7\u4e0a\u4f20\u4e00\u5f20\u56fe\u7247，\u7136\u540e\u63d0\u95ee\u3002").send()
\n return<\/p>\n
# \u5904\u7406\u56fe\u7247
\n image_file = image_files[0]
\n image_bytes = image_file.content<\/p>\n
# \u8fd9\u91cc\u5e94\u8be5\u662f\u8c03\u7528Ostrakon-VL\u6a21\u578b\u7684\u4ee3\u7801
\n # \u7531\u4e8e\u6a21\u578b\u8c03\u7528\u9700\u8981\u4e00\u4e9b\u8bbe\u7f6e，\u8fd9\u91cc\u5148\u8fd4\u56de\u4e00\u4e2a\u793a\u4f8b\u54cd\u5e94<\/p>\n
# \u6a21\u62df\u5904\u7406\u8fc7\u7a0b
\n msg = cl.Message(content="")
\n await msg.send()<\/p>\n
# \u6a21\u62df\u601d\u8003\u8fc7\u7a0b
\n await msg.stream_token("\u6b63\u5728\u5206\u6790\u56fe\u7247…\\\\n\\\\n")<\/p>\n
# \u8fd9\u91cc\u5e94\u8be5\u662f\u5b9e\u9645\u7684\u6a21\u578b\u8c03\u7528
\n # \u793a\u4f8b：\u5047\u8bbe\u6a21\u578b\u8fd4\u56de\u4e86\u5206\u6790\u7ed3\u679c
\n analysis_result = """\u56fe\u7247\u5206\u6790\u7ed3\u679c：
\n1. \u5e97\u94fa\u7c7b\u578b：\u4fbf\u5229\u5e97
\n2. \u4e3b\u8981\u5546\u54c1：\u996e\u6599\u3001\u96f6\u98df\u3001\u65e5\u7528\u54c1
\n3. \u8d27\u67b6\u72b6\u6001：\u5546\u54c1\u6446\u653e\u6574\u9f50，\u4f46\u90e8\u5206\u8d27\u4f4d\u7a7a\u7f3a
\n4. \u5efa\u8bae：\u8865\u5145\u7a7a\u7f3a\u5546\u54c1，\u68c0\u67e5\u5546\u54c1\u4fdd\u8d28\u671f"""<\/p>\n
await msg.stream_token(analysis_result)
\n await msg.update()<\/p>\n
if __name__ == "__main__":
\n # \u5b9e\u9645\u90e8\u7f72\u65f6，\u8fd9\u91cc\u9700\u8981\u52a0\u8f7d\u6a21\u578b
\n cl.run()<\/p>\n
\u521b\u5efa\u4e00\u4e2achainlit.md\u914d\u7f6e\u6587\u4ef6：<\/p>\n
# \u6b22\u8fce\u4f7f\u7528 Ostrakon-VL \u96f6\u552e\u52a9\u624b<\/p>\n
\u8fd9\u662f\u4e00\u4e2a\u4e13\u95e8\u4e3a\u96f6\u552e\u884c\u4e1a\u8bbe\u8ba1\u7684AI\u52a9\u624b，\u53ef\u4ee5：
\n– \u5206\u6790\u5e97\u94fa\u56fe\u7247
\n– \u8bc6\u522b\u5546\u54c1\u548c\u8d27\u67b6\u72b6\u6001
\n– \u63d0\u4f9b\u5408\u89c4\u5efa\u8bae
\n– \u56de\u7b54\u96f6\u552e\u76f8\u5173\u95ee\u9898<\/p>\n
## \u4f7f\u7528\u65b9\u6cd5
\n1. \u4e0a\u4f20\u5e97\u94fa\u6216\u5546\u54c1\u56fe\u7247
\n2. \u8f93\u5165\u4f60\u7684\u95ee\u9898
\n3. \u83b7\u53d6\u4e13\u4e1a\u5206\u6790\u7ed3\u679c<\/p>\n
## \u793a\u4f8b\u95ee\u9898
\n– "\u56fe\u7247\u4e2d\u7684\u5e97\u94fa\u540d\u662f\u4ec0\u4e48？"
\n– "\u8d27\u67b6\u4e0a\u7f3a\u4e86\u54ea\u4e9b\u5546\u54c1？"
\n– "\u5546\u54c1\u6446\u653e\u662f\u5426\u7b26\u5408\u6807\u51c6？"<\/p>\n
# \u542f\u52a8Chainlit
\nchainlit run app.py -w<\/p>\n
# \u6216\u8005\u6307\u5b9a\u7aef\u53e3
\nchainlit run app.py -w –port 7860<\/p>\n
\u6253\u5f00\u6d4f\u89c8\u5668，\u8bbf\u95ee http:\/\/\u4f60\u7684\u670d\u52a1\u5668IP:7860，\u5c31\u80fd\u770b\u5230\u4ea4\u4e92\u754c\u9762\u4e86\u3002<\/p>\n
\u5728ARM\u670d\u52a1\u5668\u4e0a\u90e8\u7f72\u540e，\u6211\u4eec\u9700\u8981\u9a8c\u8bc1\u6027\u80fd\u662f\u5426\u8fbe\u6807\u3002\u521b\u5efa\u4e00\u4e2a\u6d4b\u8bd5\u811a\u672c：<\/p>\n
import time
\nimport requests
\nimport base64
\nfrom PIL import Image
\nimport io<\/p>\n
def test_performance():
\n """\u6d4b\u8bd5\u6a21\u578b\u63a8\u7406\u6027\u80fd"""<\/p>\n
# \u51c6\u5907\u6d4b\u8bd5\u56fe\u7247
\n # \u8fd9\u91cc\u4f7f\u7528\u4e00\u4e2a\u793a\u4f8b\u56fe\u7247，\u5b9e\u9645\u6d4b\u8bd5\u65f6\u66ff\u6362\u4e3a\u771f\u5b9e\u56fe\u7247
\n test_cases = [
\n {
\n "image": "retail_store.jpg", # \u66ff\u6362\u4e3a\u4f60\u7684\u6d4b\u8bd5\u56fe\u7247
\n "question": "\u56fe\u7247\u4e2d\u7684\u5e97\u94fa\u540d\u662f\u4ec0\u4e48？",
\n "expected_keywords": ["\u4fbf\u5229\u5e97", "\u8d85\u5e02", "\u5546\u5e97"]
\n },
\n {
\n "image": "shelf.jpg",
\n "question": "\u8d27\u67b6\u4e0a\u7f3a\u4e86\u54ea\u4e9b\u5546\u54c1？",
\n "expected_keywords": ["\u7a7a\u7f3a", "\u8865\u8d27", "\u5e93\u5b58"]
\n }
\n ]<\/p>\n
results = []<\/p>\n
for i, test_case in enumerate(test_cases):
\n print(f"\\\\n\u6d4b\u8bd5\u7528\u4f8b {i+1}: {test_case['question']}")<\/p>\n
start_time = time.time()<\/p>\n
# \u8fd9\u91cc\u5e94\u8be5\u662f\u5b9e\u9645\u7684\u6a21\u578b\u8c03\u7528
\n # \u793a\u4f8b：\u6a21\u62dfAPI\u8c03\u7528
\n response = {
\n "answer": "\u8fd9\u662f\u4e00\u4e2a\u4fbf\u5229\u5e97，\u5e97\u540d\u662f'\u5feb\u5ba2\u4fbf\u5229\u5e97'\u3002\u8d27\u67b6\u4e0a\u6709\u996e\u6599\u3001\u96f6\u98df\u7b49\u5546\u54c1\u3002",
\n "processing_time": 2.5 # \u6a21\u62df\u5904\u7406\u65f6\u95f4
\n }<\/p>\n
end_time = time.time()
\n actual_time = end_time – start_time<\/p>\n
# \u68c0\u67e5\u54cd\u5e94
\n answer = response["answer"]
\n contains_keywords = any(keyword in answer for keyword in test_case["expected_keywords"])<\/p>\n
result = {
\n "test_case": test_case["question"],
\n "response_time": actual_time,
\n "contains_keywords": contains_keywords,
\n "answer": answer[:100] + "…" if len(answer) > 100 else answer
\n }<\/p>\n
results.append(result)<\/p>\n
print(f"\u54cd\u5e94\u65f6\u95f4: {actual_time:.2f}\u79d2")
\n print(f"\u5305\u542b\u5173\u952e\u8bcd: {contains_keywords}")
\n print(f"\u56de\u7b54\u6458\u8981: {result['answer']}")<\/p>\n
return results<\/p>\n
def test_concurrent_requests():
\n """\u6d4b\u8bd5\u5e76\u53d1\u5904\u7406\u80fd\u529b"""
\n print("\\\\n=== \u5e76\u53d1\u6027\u80fd\u6d4b\u8bd5 ===")<\/p>\n
# \u6a21\u62df\u5e76\u53d1\u8bf7\u6c42
\n concurrent_users = 3
\n print(f"\u6a21\u62df {concurrent_users} \u4e2a\u5e76\u53d1\u7528\u6237")<\/p>\n
# \u8fd9\u91cc\u5e94\u8be5\u662f\u5b9e\u9645\u7684\u5e76\u53d1\u6d4b\u8bd5
\n # \u5728ARM\u670d\u52a1\u5668\u4e0a，\u5e76\u53d1\u6027\u80fd\u53ef\u80fd\u6709\u6240\u4e0d\u540c<\/p>\n
print("\u5e76\u53d1\u6d4b\u8bd5\u5b8c\u6210")
\n return {"concurrent_users": concurrent_users, "status": "passed"}<\/p>\n
if __name__ == "__main__":
\n print("\u5f00\u59cbARM\u67b6\u6784\u6027\u80fd\u6d4b\u8bd5…")<\/p>\n
# \u5355\u8bf7\u6c42\u6027\u80fd\u6d4b\u8bd5
\n single_results = test_performance()<\/p>\n
# \u5e76\u53d1\u6d4b\u8bd5
\n concurrent_results = test_concurrent_requests()<\/p>\n
print("\\\\n=== \u6d4b\u8bd5\u603b\u7ed3 ===")
\n print(f"\u5355\u8bf7\u6c42\u5e73\u5747\u54cd\u5e94\u65f6\u95f4: {sum(r['response_time'] for r in single_results)\/len(single_results):.2f}\u79d2")
\n print(f"\u6240\u6709\u6d4b\u8bd5\u7528\u4f8b\u901a\u8fc7: {all(r['contains_keywords'] for r in single_results)}")<\/p>\n
\u5728ARM\u67b6\u6784\u4e0a，\u4f60\u53ef\u80fd\u4f1a\u9047\u5230\u8fd9\u4e9b\u95ee\u9898：<\/p>\n
\u95ee\u98981：\u67d0\u4e9bPython\u5305\u6ca1\u6709ARM\u7248\u672c<\/p>\n
# \u89e3\u51b3\u65b9\u6cd5：\u4f7f\u7528\u66ff\u4ee3\u5305\u6216\u4ece\u6e90\u7801\u7f16\u8bd1
\npip install –no-binary :all: package_name<\/p>\n
# \u6216\u8005\u4f7f\u7528conda（\u5bf9ARM\u652f\u6301\u66f4\u597d）
\nconda install package_name<\/p>\n
\u95ee\u98982：\u5185\u5b58\u8bbf\u95ee\u9519\u8bef<\/p>\n
# ARM\u67b6\u6784\u53ef\u80fd\u6709\u4e0d\u540c\u7684\u5185\u5b58\u5bf9\u9f50\u8981\u6c42
\n# \u89e3\u51b3\u65b9\u6cd5：\u786e\u4fdd\u4f7f\u7528\u6700\u65b0\u7248\u672c\u7684\u6846\u67b6
\npip install –upgrade vllm
\npip install –upgrade torch<\/p>\n
\u95ee\u98983：\u6027\u80fd\u4e0d\u5982x86<\/p>\n
# \u89e3\u51b3\u65b9\u6cd5：\u542f\u7528ARM\u4f18\u5316
\nexport OMP_NUM_THREADS=$(nproc) # \u4f7f\u7528\u6240\u6709CPU\u6838\u5fc3
\nexport MKL_NUM_THREADS=$(nproc)<\/p>\n
# \u5bf9\u4e8eNVIDIA Grace，\u4f7f\u7528CUDA\u4f18\u5316
\nexport CUDA_VISIBLE_DEVICES=0<\/p>\n
CPU\u4f18\u5316：<\/p>\n
# \u5728Python\u4ee3\u7801\u4e2d\u8bbe\u7f6e\u7ebf\u7a0b\u6570
\nimport os
\nos.environ["OMP_NUM_THREADS"] = str(os.cpu_count())
\nos.environ["MKL_NUM_THREADS"] = str(os.cpu_count())<\/p>\n
# \u4f7f\u7528ARM\u4f18\u5316\u7684\u6570\u5b66\u5e93
\n# \u5b89\u88c5OpenBLAS\u6216ARM Performance Libraries<\/p>\n
\u5185\u5b58\u4f18\u5316：<\/p>\n
# vLLM\u914d\u7f6e\u4f18\u5316
\nllm = LLM(
\n model="Salesforce\/Ostrakon-VL-8B",
\n tensor_parallel_size=1,
\n gpu_memory_utilization=0.85, # \u7a0d\u5fae\u964d\u4f4e\u4e00\u70b9，\u7ed9\u7cfb\u7edf\u7559\u7a7a\u95f4
\n max_model_len=2048, # \u6839\u636e\u9700\u6c42\u8c03\u6574，\u51cf\u5c11\u5185\u5b58\u4f7f\u7528
\n enable_prefix_caching=True, # \u542f\u7528\u524d\u7f00\u7f13\u5b58，\u63d0\u9ad8\u91cd\u590d\u67e5\u8be2\u6027\u80fd
\n swap_space=4, # \u4f7f\u75284GB\u78c1\u76d8\u7a7a\u95f4\u4f5c\u4e3a\u4ea4\u6362（\u5982\u679c\u5185\u5b58\u4e0d\u8db3）
\n)<\/p>\n
\u9488\u5bf9ARM\u670d\u52a1\u5668\u7684vLLM\u914d\u7f6e\u5efa\u8bae：<\/p>\n
# \u4f18\u5316\u7684vLLM\u914d\u7f6e
\noptimized_config = {
\n "model": "Salesforce\/Ostrakon-VL-8B",
\n "tensor_parallel_size": 1, # ARM\u670d\u52a1\u5668\u901a\u5e38GPU\u8f83\u5c11
\n "pipeline_parallel_size": 1,
\n "block_size": 16, # \u8f83\u5c0f\u7684block size\u9002\u5408ARM
\n "max_num_batched_tokens": 2048, # \u6839\u636eARM\u5185\u5b58\u8c03\u6574
\n "max_num_seqs": 32, # \u5e76\u53d1\u8bf7\u6c42\u6570
\n "gpu_memory_utilization": 0.8, # \u4fdd\u5b88\u4e00\u70b9
\n "enable_chunked_prefill": True, # \u6539\u5584\u5927\u63d0\u793a\u8bcd\u5904\u7406
\n "max_model_len": 4096,
\n}<\/p>\n
\u4e3a\u4e86\u8ba9\u4f60\u66f4\u6e05\u695aARM\u670d\u52a1\u5668\u7684\u8868\u73b0，\u8fd9\u91cc\u6709\u4e00\u4e2a\u7b80\u5355\u7684\u5bf9\u6bd4：<\/p>\n
| \u9996\u6b21\u52a0\u8f7d\u65f6\u95f4<\/td>\n | \u7ea690\u79d2<\/td>\n | \u7ea6120\u79d2<\/td>\n | \u6b63\u5e38，ARM\u9700\u8981\u66f4\u591a\u4f18\u5316<\/td>\n<\/tr>\n |
| \u5355\u56fe\u7247\u63a8\u7406\u65f6\u95f4<\/td>\n | 2.1\u79d2<\/td>\n | 2.8\u79d2<\/td>\n | \u53ef\u63a5\u53d7，\u5dee30%<\/td>\n<\/tr>\n |
| \u5185\u5b58\u4f7f\u7528\u5cf0\u503c<\/td>\n | 14GB<\/td>\n | 16GB<\/td>\n | ARM\u7a0d\u9ad8，\u5efa\u8bae32GB+\u5185\u5b58<\/td>\n<\/tr>\n |
| \u5e76\u53d1\u5904\u7406\u80fd\u529b<\/td>\n | \u652f\u630120+\u5e76\u53d1<\/td>\n | \u652f\u630115+\u5e76\u53d1<\/td>\n | \u8db3\u591f\u5927\u591a\u6570\u573a\u666f<\/td>\n<\/tr>\n |
| \u80fd\u6548\u6bd4<\/td>\n | 1.0（\u57fa\u51c6）<\/td>\n | \u7ea61.3\u500d<\/td>\n | ARM\u66f4\u7701\u7535<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n \u8c03\u4f18\u540e\u7684\u6539\u5584： \u7ecf\u8fc7\u4e0a\u8ff0\u4f18\u5316，ARM\u670d\u52a1\u5668\u7684\u6027\u80fd\u53ef\u4ee5\u63d0\u534715-25%，\u63a5\u8fd1x86\u670d\u52a1\u5668\u7684\u6c34\u5e73\u3002<\/p>\n 8. \u5b9e\u9645\u5e94\u7528\u793a\u4f8b<\/h3>\n |