\u5982\u679c\u4f60\u6b63\u5728\u5bfb\u627e\u4e00\u4e2a\u80fd\u771f\u6b63\u770b\u61c2\u96f6\u552e\u548c\u9910\u996e\u573a\u666f\u56fe\u7247\u7684AI\u52a9\u624b，\u90a3\u4e48Ostrakon-VL-8B\u7edd\u5bf9\u503c\u5f97\u4f60\u82b1\u65f6\u95f4\u4e86\u89e3\u4e00\u4e0b\u3002<\/p>\n
\u60f3\u8c61\u4e00\u4e0b\u8fd9\u6837\u7684\u573a\u666f：\u4f60\u6709\u4e00\u5bb6\u8fde\u9501\u8d85\u5e02，\u6bcf\u5929\u8981\u5904\u7406\u6210\u5343\u4e0a\u4e07\u7684\u8d27\u67b6\u7167\u7247，\u9700\u8981\u68c0\u67e5\u5546\u54c1\u6446\u653e\u662f\u5426\u5408\u89c4\u3001\u4ef7\u683c\u6807\u7b7e\u662f\u5426\u6b63\u786e\u3001\u5e93\u5b58\u662f\u5426\u5145\u8db3\u3002\u4f20\u7edf\u65b9\u6cd5\u8981\u4e48\u9760\u4eba\u5de5\u4e00\u5f20\u5f20\u770b，\u6548\u7387\u4f4e\u4e0b；\u8981\u4e48\u7528\u901a\u7528AI\u6a21\u578b，\u4f46\u6548\u679c\u603b\u662f\u4e0d\u5c3d\u5982\u4eba\u610f\u2014\u2014\u5b83\u4eec\u53ef\u80fd\u8ba4\u8bc6\u201c\u82f9\u679c\u201d，\u4f46\u5206\u4e0d\u6e05\u201c\u7ea2\u5bcc\u58eb\u201d\u548c\u201c\u560e\u5566\u201d，\u66f4\u770b\u4e0d\u61c2\u4fc3\u9500\u6d77\u62a5\u4e0a\u7684\u590d\u6742\u4fe1\u606f\u3002<\/p>\n
Ostrakon-VL-8B\u5c31\u662f\u4e3a\u89e3\u51b3\u8fd9\u7c7b\u95ee\u9898\u800c\u751f\u7684\u3002\u5b83\u4e0d\u662f\u4e00\u4e2a\u666e\u901a\u7684\u56fe\u6587\u5bf9\u8bdd\u6a21\u578b，\u800c\u662f\u4e13\u95e8\u9488\u5bf9\u98df\u54c1\u670d\u52a1\u548c\u96f6\u552e\u5546\u5e97\u573a\u666f\u8bad\u7ec3\u7684\u201c\u9886\u57df\u4e13\u5bb6\u201d\u3002\u6700\u8ba9\u4eba\u60ca\u559c\u7684\u662f，\u8fd9\u4e2a\u53ea\u670980\u4ebf\u53c2\u6570\u7684\u201c\u5c0f\u201d\u6a21\u578b，\u5728\u67d0\u4e9b\u4e13\u4e1a\u4efb\u52a1\u4e0a\u7684\u8868\u73b0，\u7adf\u7136\u80fd\u8d85\u8fc7\u90a3\u4e9b\u53c2\u6570\u5927\u51e0\u5341\u500d\u7684\u901a\u7528\u5927\u6a21\u578b\u3002<\/p>\n
\u4eca\u5929\u6211\u8981\u5e26\u4f60\u505a\u7684，\u5c31\u662f\u5728ARM\u67b6\u6784\u7684\u670d\u52a1\u5668\u4e0a（\u6bd4\u5982NVIDIA Grace\u5e73\u53f0），\u7528vLLM\u6846\u67b6\u5feb\u901f\u90e8\u7f72\u8fd9\u4e2a\u6a21\u578b，\u5e76\u7528\u4e00\u4e2a\u7b80\u5355\u7684\u524d\u7aef\u754c\u9762\u6765\u8c03\u7528\u5b83\u3002\u6574\u4e2a\u8fc7\u7a0b\u5c31\u50cf\u642d\u79ef\u6728\u4e00\u6837\u7b80\u5355，\u5373\u4f7f\u4f60\u4e0d\u662f\u6df1\u5ea6\u5b66\u4e60\u4e13\u5bb6，\u4e5f\u80fd\u8ddf\u7740\u4e00\u6b65\u6b65\u5b8c\u6210\u3002<\/p>\n
\u5728\u5f00\u59cb\u4e4b\u524d，\u6211\u4eec\u5148\u5feb\u901f\u4e86\u89e3\u4e00\u4e0b\u8981\u7528\u5230\u7684\u51e0\u4e2a\u5173\u952e\u7ec4\u4ef6：<\/p>\n
Ostrakon-VL-8B：\u8fd9\u662f\u4eca\u5929\u7684\u4e3b\u89d2\u3002\u5b83\u57fa\u4e8eQwen3-VL-8B\u6784\u5efa，\u4e13\u95e8\u9488\u5bf9\u96f6\u552e\u548c\u9910\u996e\u573a\u666f\u505a\u4e86\u6df1\u5ea6\u4f18\u5316\u3002\u7b80\u5355\u8bf4，\u5b83\u7279\u522b\u64c5\u957f：<\/p>\n
vLLM：\u8fd9\u662f\u4e00\u4e2a\u9ad8\u6027\u80fd\u7684\u63a8\u7406\u6846\u67b6\u3002\u4f60\u53ef\u4ee5\u628a\u5b83\u60f3\u8c61\u6210\u4e00\u4e2a\u201c\u6a21\u578b\u52a0\u901f\u5668\u201d，\u5b83\u80fd\u8ba9\u5927\u6a21\u578b\u8dd1\u5f97\u66f4\u5feb\u3001\u66f4\u7701\u5185\u5b58\u3002\u7279\u522b\u662f\u5728ARM\u67b6\u6784\u7684\u670d\u52a1\u5668\u4e0a，vLLM\u7684\u517c\u5bb9\u6027\u505a\u5f97\u5f88\u597d\u3002<\/p>\n
Chainlit：\u8fd9\u662f\u4e00\u4e2a\u8f7b\u91cf\u7ea7\u7684Web\u524d\u7aef\u6846\u67b6\u3002\u6211\u4eec\u7528\u5b83\u6765\u642d\u5efa\u4e00\u4e2a\u804a\u5929\u754c\u9762，\u8ba9\u4f60\u53ef\u4ee5\u901a\u8fc7\u7f51\u9875\u76f4\u63a5\u548c\u6a21\u578b\u5bf9\u8bdd，\u4e0a\u4f20\u56fe\u7247\u3001\u63d0\u95ee\u3001\u770b\u56de\u7b54，\u5c31\u50cf\u7528ChatGPT\u4e00\u6837\u7b80\u5355\u3002<\/p>\n
ARM\u67b6\u6784\u670d\u52a1\u5668：\u6211\u4eec\u4eca\u5929\u7684\u90e8\u7f72\u73af\u5883\u3002ARM\u67b6\u6784\u73b0\u5728\u5728\u670d\u52a1\u5668\u9886\u57df\u8d8a\u6765\u8d8a\u6d41\u884c，\u7279\u522b\u662fNVIDIA\u7684Grace\u5e73\u53f0，\u6027\u80fd\u5f88\u5f3a\u52b2\u3002\u5982\u679c\u4f60\u7528\u7684\u662f\u5176\u4ed6ARM\u670d\u52a1\u5668，\u6bd4\u5982AWS\u7684Graviton\u6216\u8005\u82f9\u679c\u7684M\u7cfb\u5217\u82af\u7247，\u65b9\u6cd5\u4e5f\u57fa\u672c\u901a\u7528\u3002<\/p>\n
\u5f00\u59cb\u90e8\u7f72\u524d，\u8bf7\u786e\u8ba4\u4f60\u7684\u73af\u5883\u6ee1\u8db3\u4ee5\u4e0b\u8981\u6c42：<\/p>\n
\u5982\u679c\u4f60\u7528\u7684\u662f\u4e91\u670d\u52a1\u5668，\u5efa\u8bae\u9009\u62e9\u6709GPU\u7684\u5b9e\u4f8b，\u867d\u7136\u8fd9\u4e2a\u6a21\u578b\u5728CPU\u4e0a\u4e5f\u80fd\u8dd1，\u4f46\u6709GPU\u4f1a\u5feb\u5f88\u591a\u3002ARM\u67b6\u6784\u7684GPU（\u6bd4\u5982NVIDIA Grace\u4e0a\u7684GPU）\u5b8c\u5168\u6ca1\u95ee\u9898\u3002<\/p>\n
\u6253\u5f00\u4f60\u7684\u7ec8\u7aef，\u6211\u4eec\u4ece\u5934\u5f00\u59cb：<\/p>\n
# \u521b\u5efa\u4e00\u4e2a\u4e13\u95e8\u7684\u9879\u76ee\u76ee\u5f55
\nmkdir ostrakon-vl-deployment
\ncd ostrakon-vl-deployment<\/p>\n
# \u521b\u5efaPython\u865a\u62df\u73af\u5883（\u63a8\u8350，\u907f\u514d\u5305\u51b2\u7a81）
\npython3 -m venv venv<\/p>\n
# \u6fc0\u6d3b\u865a\u62df\u73af\u5883
\nsource venv\/bin\/activate<\/p>\n
# \u5982\u679c\u4f60\u7684\u7cfb\u7edf\u662fARM\u67b6\u6784，\u786e\u8ba4\u4e00\u4e0b
\nuname -m
\n# \u5e94\u8be5\u663e\u793a aarch64 \u6216 arm64<\/p>\n
\u770b\u5230aarch64\u6216arm64\u5c31\u8bf4\u660e\u4f60\u786e\u5b9e\u5728ARM\u67b6\u6784\u4e0a\u3002\u522b\u62c5\u5fc3，vLLM\u5bf9ARM\u7684\u652f\u6301\u5f88\u597d，\u6211\u4eec\u7528\u7684\u90fd\u662f\u517c\u5bb9\u7684\u7248\u672c\u3002<\/p>\n
\u73b0\u5728\u5b89\u88c5\u6211\u4eec\u9700\u8981\u7684Python\u5305\u3002\u8fd9\u91cc\u6709\u4e2a\u5c0f\u6280\u5de7：\u56e0\u4e3a\u6211\u4eec\u8981\u5728ARM\u4e0a\u8dd1，\u6709\u4e9b\u5305\u9700\u8981\u4ece\u6e90\u7801\u7f16\u8bd1，\u6240\u4ee5\u5148\u786e\u4fdd\u7cfb\u7edf\u6709\u7f16\u8bd1\u5de5\u5177：<\/p>\n
# \u66f4\u65b0\u5305\u5217\u8868\u5e76\u5b89\u88c5\u7f16\u8bd1\u5de5\u5177
\nsudo apt update
\nsudo apt install -y build-essential python3-dev<\/p>\n
# \u5b89\u88c5PyTorch（ARM\u517c\u5bb9\u7248\u672c）
\n# \u8bbf\u95ee https:\/\/pytorch.org\/get-started\/locally\/ \u67e5\u770b\u6700\u65b0\u7684ARM\u7248\u672c
\n# \u8fd9\u91cc\u4ee5PyTorch 2.3\u4e3a\u4f8b
\npip install torch torchvision torchaudio –index-url https:\/\/download.pytorch.org\/whl\/cpu<\/p>\n
# \u5b89\u88c5vLLM（\u6ce8\u610f\u7248\u672c，\u8981\u517c\u5bb9ARM）
\npip install vllm==0.4.2<\/p>\n
# \u5b89\u88c5Chainlit\u548c\u5176\u4ed6\u4f9d\u8d56
\npip install chainlit transformers accelerate pillow<\/p>\n
\u5b89\u88c5\u8fc7\u7a0b\u53ef\u80fd\u9700\u8981\u51e0\u5206\u949f，\u7279\u522b\u662fvLLM，\u56e0\u4e3a\u5b83\u8981\u5728\u4f60\u7684\u673a\u5668\u4e0a\u7f16\u8bd1\u4e00\u4e9bC++\u6269\u5c55\u3002\u5982\u679c\u9047\u5230\u7f51\u7edc\u95ee\u9898，\u53ef\u4ee5\u8bd5\u8bd5\u56fd\u5185\u7684\u955c\u50cf\u6e90：<\/p>\n
pip install vllm==0.4.2 -i https:\/\/pypi.tuna.tsinghua.edu.cn\/simple<\/p>\n
\u6a21\u578b\u53ef\u4ee5\u4eceHugging Face\u4e0b\u8f7d\u3002\u6211\u4eec\u5148\u521b\u5efa\u4e00\u4e2aPython\u811a\u672c\u6765\u5904\u7406\u4e0b\u8f7d：<\/p>\n
# download_model.py
\nfrom transformers import AutoModelForCausalLM, AutoTokenizer
\nimport torch<\/p>\n
model_name = "OstrakonAI\/Ostrakon-VL-8B"
\nsave_path = ".\/ostrakon-vl-8b"<\/p>\n
print("\u5f00\u59cb\u4e0b\u8f7dOstrakon-VL-8B\u6a21\u578b…")
\nprint("\u8fd9\u53ef\u80fd\u9700\u8981\u4e00\u4e9b\u65f6\u95f4，\u6a21\u578b\u5927\u5c0f\u7ea615GB")<\/p>\n
# \u4e0b\u8f7d\u6a21\u578b\u548ctokenizer
\nmodel = AutoModelForCausalLM.from_pretrained(
\n model_name,
\n torch_dtype=torch.float16, # \u7528\u534a\u7cbe\u5ea6\u8282\u7701\u5185\u5b58
\n device_map="auto", # \u81ea\u52a8\u5206\u914d\u5230\u53ef\u7528\u8bbe\u5907
\n trust_remote_code=True
\n)<\/p>\n
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)<\/p>\n
# \u4fdd\u5b58\u5230\u672c\u5730
\nmodel.save_pretrained(save_path)
\ntokenizer.save_pretrained(save_path)<\/p>\n
print(f"\u6a21\u578b\u5df2\u4fdd\u5b58\u5230: {save_path}")<\/p>\n
\u8fd0\u884c\u8fd9\u4e2a\u811a\u672c：<\/p>\n
python download_model.py<\/p>\n
\u4e0b\u8f7d\u65f6\u95f4\u53d6\u51b3\u4e8e\u4f60\u7684\u7f51\u901f，\u5927\u6982\u9700\u898130\u5206\u949f\u52301\u5c0f\u65f6\u3002\u5982\u679c\u4e0b\u8f7d\u4e2d\u65ad，\u53ef\u4ee5\u91cd\u65b0\u8fd0\u884c，\u5b83\u4f1a\u81ea\u52a8\u7eed\u4f20\u3002<\/p>\n
\u6a21\u578b\u4e0b\u8f7d\u597d\u540e，\u6211\u4eec\u521b\u5efa\u4e00\u4e2a\u542f\u52a8\u811a\u672c\u3002vLLM\u7684\u597d\u5904\u662f\u5b83\u5185\u7f6e\u4e86\u4e00\u4e2aOpenAI\u517c\u5bb9\u7684API\u670d\u52a1\u5668，\u6211\u4eec\u76f4\u63a5\u7528\u5b83\u5c31\u884c：<\/p>\n
# start_server.py
\nfrom vllm import LLM, SamplingParams
\nfrom vllm.entrypoints.openai import api_server
\nimport argparse
\nimport uvicorn<\/p>\n
def main():
\n parser = argparse.ArgumentParser()
\n parser.add_argument("–model", type=str, default=".\/ostrakon-vl-8b")
\n parser.add_argument("–host", type=str, default="0.0.0.0")
\n parser.add_argument("–port", type=int, default=8000)
\n parser.add_argument("–gpu-memory-utilization", type=float, default=0.9)
\n args = parser.parse_args()<\/p>\n
print("\u6b63\u5728\u52a0\u8f7dOstrakon-VL-8B\u6a21\u578b…")
\n print(f"\u6a21\u578b\u8def\u5f84: {args.model}")
\n print(f"\u670d\u52a1\u5730\u5740: http:\/\/{args.host}:{args.port}")<\/p>\n
# \u521b\u5efaLLM\u5b9e\u4f8b
\n llm = LLM(
\n model=args.model,
\n tokenizer=args.model,
\n tensor_parallel_size=1, # \u5982\u679c\u6709\u591aGPU\u53ef\u4ee5\u8c03\u6574
\n gpu_memory_utilization=args.gpu_memory_utilization,
\n trust_remote_code=True,
\n max_model_len=8192 # \u6700\u5927\u4e0a\u4e0b\u6587\u957f\u5ea6
\n )<\/p>\n
# \u542f\u52a8API\u670d\u52a1\u5668
\n app = api_server.create_app(llm=llm)<\/p>\n
print("\u6a21\u578b\u52a0\u8f7d\u5b8c\u6210！\u670d\u52a1\u5df2\u542f\u52a8\u3002")
\n print("\u4f60\u53ef\u4ee5\u901a\u8fc7\u4ee5\u4e0b\u65b9\u5f0f\u6d4b\u8bd5:")
\n print(f"curl http:\/\/{args.host}:{args.port}\/v1\/completions \\\\\\\\")
\n print(' -H "Content-Type: application\/json" \\\\\\\\')
\n print(' -d \\\\'{"model": "ostrakon-vl-8b", "prompt": "Hello", "max_tokens": 100}\\\\'')<\/p>\n
uvicorn.run(app, host=args.host, port=args.port)<\/p>\n
if __name__ == "__main__":
\n main()<\/p>\n
\u4fdd\u5b58\u4e3astart_server.py，\u7136\u540e\u8fd0\u884c：<\/p>\n
# \u5728\u540e\u53f0\u8fd0\u884c\u670d\u52a1，\u8f93\u51fa\u91cd\u5b9a\u5411\u5230\u65e5\u5fd7\u6587\u4ef6
\nnohup python start_server.py > llm.log 2>&1 &<\/p>\n
# \u67e5\u770b\u65e5\u5fd7，\u786e\u8ba4\u670d\u52a1\u542f\u52a8
\ntail -f llm.log<\/p>\n
\u4f60\u5e94\u8be5\u80fd\u770b\u5230\u7c7b\u4f3c\u8fd9\u6837\u7684\u8f93\u51fa：<\/p>\n
\u6b63\u5728\u52a0\u8f7dOstrakon-VL-8B\u6a21\u578b…
\n\u6a21\u578b\u8def\u5f84: .\/ostrakon-vl-8b
\n\u670d\u52a1\u5730\u5740: http:\/\/0.0.0.0:8000
\n\u6a21\u578b\u52a0\u8f7d\u5b8c\u6210！\u670d\u52a1\u5df2\u542f\u52a8\u3002<\/p>\n
\u670d\u52a1\u542f\u52a8\u540e，\u6a21\u578b\u9700\u8981\u4e00\u4e9b\u65f6\u95f4\u52a0\u8f7d\u5230\u5185\u5b58（\u6216GPU\u663e\u5b58）\u3002\u4f60\u53ef\u4ee5\u7528\u4e0b\u9762\u7684\u547d\u4ee4\u68c0\u67e5\u8fdb\u5ea6：<\/p>\n
# \u67e5\u770b\u6a21\u578b\u670d\u52a1\u662f\u5426\u90e8\u7f72\u6210\u529f
\ncat llm.log<\/p>\n
\u5f53\u770b\u5230"\u6a21\u578b\u52a0\u8f7d\u5b8c\u6210！\u670d\u52a1\u5df2\u542f\u52a8\u3002"\u65f6，\u5c31\u8bf4\u660e\u51c6\u5907\u597d\u4e86\u3002<\/p>\n
\u73b0\u5728\u6a21\u578b\u670d\u52a1\u8dd1\u8d77\u6765\u4e86，\u6211\u4eec\u521b\u5efa\u4e00\u4e2a\u7b80\u5355\u7684Web\u754c\u9762\u6765\u8c03\u7528\u5b83\u3002Chainlit\u8ba9\u8fd9\u4ef6\u4e8b\u53d8\u5f97\u7279\u522b\u7b80\u5355：<\/p>\n
# app.py – Chainlit\u5e94\u7528
\nimport chainlit as cl
\nimport requests
\nimport base64
\nfrom PIL import Image
\nimport io
\nimport json<\/p>\n
# vLLM\u670d\u52a1\u5668\u7684\u5730\u5740
\nVLLM_SERVER = "http:\/\/localhost:8000"<\/p>\n
@cl.on_chat_start
\nasync def start_chat():
\n await cl.Message(
\n content="\u4f60\u597d！\u6211\u662fOstrakon-VL-8B\u52a9\u624b，\u4e13\u95e8\u5904\u7406\u96f6\u552e\u548c\u9910\u996e\u573a\u666f\u7684\u56fe\u6587\u95ee\u9898\u3002\u4f60\u53ef\u4ee5\u4e0a\u4f20\u5e97\u94fa\u7167\u7247\u3001\u5546\u54c1\u56fe\u7247\u7b49，\u6211\u4f1a\u5e2e\u4f60\u5206\u6790\u3002"
\n ).send()<\/p>\n
@cl.on_message
\nasync def handle_message(message: cl.Message):
\n # \u68c0\u67e5\u6d88\u606f\u4e2d\u662f\u5426\u6709\u56fe\u7247
\n images = []
\n if message.elements:
\n for element in message.elements:
\n if "image" in element.mime:
\n # \u5c06\u56fe\u7247\u8f6c\u6362\u4e3abase64
\n img_bytes = element.content
\n img_base64 = base64.b64encode(img_bytes).decode('utf-8')
\n images.append(f"data:image\/jpeg;base64,{img_base64}")<\/p>\n
if images:
\n # \u5982\u679c\u6709\u56fe\u7247，\u6784\u5efa\u591a\u6a21\u6001\u8bf7\u6c42
\n user_message = message.content<\/p>\n
# \u6784\u5efa\u7b26\u5408Ostrakon-VL\u683c\u5f0f\u7684prompt
\n # \u6ce8\u610f：Ostrakon-VL\u4f7f\u7528\u7279\u6b8a\u7684\u56fe\u6587\u683c\u5f0f
\n prompt_parts = []<\/p>\n
for img in images:
\n prompt_parts.append(f'<img src="{img}" \/>')<\/p>\n
prompt_parts.append(f"\\\\n\u7528\u6237\u95ee\u9898: {user_message}")
\n prompt_parts.append("\\\\n\u52a9\u624b\u56de\u7b54:")<\/p>\n
full_prompt = "".join(prompt_parts)<\/p>\n
# \u53d1\u9001\u5230vLLM\u670d\u52a1\u5668
\n response = requests.post(
\n f"{VLLM_SERVER}\/v1\/completions",
\n json={
\n "model": "ostrakon-vl-8b",
\n "prompt": full_prompt,
\n "max_tokens": 500,
\n "temperature": 0.1, # \u4f4e\u6e29\u5ea6\u8ba9\u56de\u7b54\u66f4\u786e\u5b9a
\n "stop": ["<|endoftext|>", "\u7528\u6237\u95ee\u9898:", "\u52a9\u624b\u56de\u7b54:"]
\n }
\n )<\/p>\n
if response.status_code == 200:
\n result = response.json()
\n answer = result["choices"][0]["text"].strip()
\n await cl.Message(content=answer).send()
\n else:
\n await cl.Message(
\n content=f"\u8bf7\u6c42\u5931\u8d25: {response.status_code}\\\\n{response.text}"
\n ).send()
\n else:
\n # \u5982\u679c\u6ca1\u6709\u56fe\u7247，\u53ea\u5904\u7406\u6587\u672c
\n response = requests.post(
\n f"{VLLM_SERVER}\/v1\/completions",
\n json={
\n "model": "ostrakon-vl-8b",
\n "prompt": f"\u7528\u6237\u95ee\u9898: {message.content}\\\\n\u52a9\u624b\u56de\u7b54:",
\n "max_tokens": 300,
\n "temperature": 0.7
\n }
\n )<\/p>\n
if response.status_code == 200:
\n result = response.json()
\n answer = result["choices"][0]["text"].strip()
\n await cl.Message(content=answer).send()
\n else:
\n await cl.Message(
\n content=f"\u8bf7\u6c42\u5931\u8d25: {response.status_code}"
\n ).send()<\/p>\n
@cl.on_stop
\ndef on_stop():
\n print("\u804a\u5929\u4f1a\u8bdd\u7ed3\u675f")<\/p>\n
\u4fdd\u5b58\u4e3aapp.py，\u7136\u540e\u542f\u52a8Chainlit\u670d\u52a1：<\/p>\n
# \u542f\u52a8Chainlit，\u6307\u5b9a\u7aef\u53e3
\nchainlit run app.py -p 7860<\/p>\n
\u73b0\u5728\u6253\u5f00\u6d4f\u89c8\u5668，\u8bbf\u95ee http:\/\/\u4f60\u7684\u670d\u52a1\u5668IP:7860，\u5c31\u80fd\u770b\u5230\u804a\u5929\u754c\u9762\u4e86\u3002<\/p>\n
\u8ba9\u6211\u4eec\u8bd5\u8bd5\u6a21\u578b\u7684\u57fa\u672c\u80fd\u529b\u3002\u5728Chainlit\u754c\u9762\u4e2d：<\/p>\n
\u6a21\u578b\u5e94\u8be5\u80fd\u8bc6\u522b\u51fa\u5e97\u94fa\u62db\u724c\u4e0a\u7684\u6587\u5b57\u3002\u5982\u679c\u7167\u7247\u6e05\u6670，\u8bc6\u522b\u51c6\u786e\u7387\u5f88\u9ad8\u3002<\/p>\n
\u8bd5\u8bd5\u66f4\u4e13\u4e1a\u7684\u95ee\u9898：<\/p>\n
Ostrakon-VL\u5728\u8fd9\u65b9\u9762\u7279\u522b\u5f3a，\u5b83\u80fd\u6570\u5546\u54c1\u6570\u91cf\u3001\u5224\u65ad\u6446\u653e\u662f\u5426\u6574\u9f50，\u751a\u81f3\u80fd\u8bc6\u522b\u67d0\u4e9b\u5546\u54c1\u662f\u5426\u7f3a\u8d27\u3002<\/p>\n
\u5bf9\u4e8e\u9910\u996e\u573a\u666f：<\/p>\n
\u6a21\u578b\u7ecf\u8fc7\u4e13\u4e1a\u8bad\u7ec3，\u80fd\u8bc6\u522b\u4e00\u4e9b\u5e38\u89c1\u7684\u98df\u54c1\u5b89\u5168\u95ee\u9898，\u6bd4\u5982\u751f\u719f\u98df\u6df7\u653e\u3001\u5de5\u4f5c\u4eba\u5458\u6ca1\u6234\u624b\u5957\u7b49\u3002<\/p>\n
\u5982\u679c\u4f60\u66f4\u559c\u6b22\u7528\u4ee3\u7801\u8c03\u7528，\u8fd9\u91cc\u6709\u4e2aPython\u793a\u4f8b：<\/p>\n
import requests
\nimport base64<\/p>\n
# \u8bfb\u53d6\u56fe\u7247\u5e76\u8f6c\u6362\u4e3abase64
\ndef image_to_base64(image_path):
\n with open(image_path, "rb") as image_file:
\n return base64.b64encode(image_file.read()).decode('utf-8')<\/p>\n
# \u51c6\u5907\u8bf7\u6c42
\nimage_base64 = image_to_base64("your_shop_image.jpg")
\nprompt = f'<img src="data:image\/jpeg;base64,{image_base64}" \/>\\\\n\u7528\u6237\u95ee\u9898: \u56fe\u7247\u4e2d\u7684\u5e97\u94fa\u540d\u662f\u4ec0\u4e48？\\\\n\u52a9\u624b\u56de\u7b54:'<\/p>\n
response = requests.post(
\n "http:\/\/localhost:8000\/v1\/completions",
\n json={
\n "model": "ostrakon-vl-8b",
\n "prompt": prompt,
\n "max_tokens": 100,
\n "temperature": 0.1
\n }
\n)<\/p>\n
if response.status_code == 200:
\n result = response.json()
\n print("\u6a21\u578b\u56de\u7b54:", result["choices"][0]["text"])
\nelse:
\n print("\u8bf7\u6c42\u5931\u8d25:", response.text)<\/p>\n
\u5982\u679c\u542f\u52a8\u65f6\u9047\u5230\u9519\u8bef，\u5148\u68c0\u67e5：<\/p>\n
# \u68c0\u67e5\u65e5\u5fd7\u4e2d\u7684\u9519\u8bef\u4fe1\u606f
\ntail -100 llm.log<\/p>\n
# \u5e38\u89c1\u95ee\u98981：\u5185\u5b58\u4e0d\u8db3
\n# \u89e3\u51b3\u65b9\u6848：\u51cf\u5c11gpu_memory_utilization\u53c2\u6570\u503c，\u6bd4\u5982\u4ece0.9\u964d\u52300.7<\/p>\n
# \u5e38\u89c1\u95ee\u98982：\u6a21\u578b\u8def\u5f84\u9519\u8bef
\n# \u89e3\u51b3\u65b9\u6848：\u786e\u8ba4\u6a21\u578b\u6587\u4ef6\u786e\u5b9e\u5728.\/ostrakon-vl-8b\u76ee\u5f55\u4e0b
\nls -la .\/ostrakon-vl-8b\/<\/p>\n
# \u5e38\u89c1\u95ee\u98983：\u7aef\u53e3\u88ab\u5360\u7528
\n# \u89e3\u51b3\u65b9\u6848：\u6362\u4e00\u4e2a\u7aef\u53e3，\u6bd4\u5982–port 8001<\/p>\n
Ostrakon-VL-8B\u662f\u4e2a\u5927\u6a21\u578b，\u7b2c\u4e00\u6b21\u63a8\u7406\u4f1a\u6bd4\u8f83\u6162（\u9700\u8981\u52a0\u8f7d\u6743\u91cd）\u3002\u540e\u7eed\u8bf7\u6c42\u4f1a\u5feb\u5f88\u591a\u3002\u5982\u679c\u8fd8\u662f\u6162，\u53ef\u4ee5：<\/p>\n
\u786e\u4fdd\u7528\u4e86GPU：\u68c0\u67e5vLLM\u662f\u5426\u68c0\u6d4b\u5230\u4e86GPU<\/p>\n
python -c "import torch; print(torch.cuda.is_available())"\n <\/li>\n
\u8c03\u6574\u6279\u5904\u7406\u5927\u5c0f：vLLM\u652f\u6301\u6279\u5904\u7406，\u53ef\u4ee5\u540c\u65f6\u5904\u7406\u591a\u4e2a\u8bf7\u6c42<\/p>\n
# \u5728start_server.py\u4e2d\u8c03\u6574
\nllm = LLM(
\n model=args.model,
\n max_num_batched_tokens=4096, # \u589e\u52a0\u6279\u5904\u7406\u5927\u5c0f
\n # … \u5176\u4ed6\u53c2\u6570
\n)\n <\/li>\n
Chainlit\u5bf9\u56fe\u7247\u5927\u5c0f\u6709\u9650\u5236\u3002\u5982\u679c\u56fe\u7247\u592a\u5927：<\/p>\n
\u538b\u7f29\u56fe\u7247：\u7528PIL\u5e93\u5728\u4e0a\u4f20\u524d\u538b\u7f29<\/p>\n
from PIL import Image
\nimport io<\/p>\n
def compress_image(image_path, max_size=1024):
\n img = Image.open(image_path)
\n img.thumbnail((max_size, max_size))
\n byte_arr = io.BytesIO()
\n img.save(byte_arr, format='JPEG', quality=85)
\n return byte_arr.getvalue()\n <\/li>\n
\u8c03\u6574Chainlit\u914d\u7f6e：\u5728.chainlit\/config.toml\u4e2d\u589e\u52a0\u5927\u5c0f\u9650\u5236<\/p>\n
[features]
\nmax_upload_size = 10 # MB\n <\/li>\n
Ostrakon-VL\u5728\u96f6\u552e\u9910\u996e\u573a\u666f\u5f88\u4e13\u4e1a，\u4f46\u4e5f\u4e0d\u662f\u4e07\u80fd\u7684\u3002\u5982\u679c\u56de\u7b54\u4e0d\u51c6\u786e：<\/p>\n
\u5982\u679c\u4f60\u6709\u5f88\u591a\u56fe\u7247\u9700\u8981\u5206\u6790，\u53ef\u4ee5\u7528\u811a\u672c\u6279\u91cf\u5904\u7406：<\/p>\n
import os
\nimport requests
\nimport base64
\nfrom concurrent.futures import ThreadPoolExecutor<\/p>\n
def analyze_image(image_path, question):
\n """\u5206\u6790\u5355\u5f20\u56fe\u7247"""
\n with open(image_path, "rb") as f:
\n img_base64 = base64.b64encode(f.read()).decode()<\/p>\n
prompt = f'<img src="data:image\/jpeg;base64,{img_base64}" \/>\\\\n\u7528\u6237\u95ee\u9898: {question}\\\\n\u52a9\u624b\u56de\u7b54:'<\/p>\n
response = requests.post(
\n "http:\/\/localhost:8000\/v1\/completions",
\n json={
\n "model": "ostrakon-vl-8b",
\n "prompt": prompt,
\n "max_tokens": 200,
\n "temperature": 0.1
\n },
\n timeout=30
\n )<\/p>\n
if response.status_code == 200:
\n return response.json()["choices"][0]["text"].strip()
\n else:
\n return f"\u9519\u8bef: {response.status_code}"<\/p>\n
# \u6279\u91cf\u5904\u7406
\nimage_dir = ".\/shop_images\/"
\nquestions = "\u8d27\u67b6\u4e0a\u7684\u5546\u54c1\u6446\u653e\u6574\u9f50\u5417？"<\/p>\n
results = []
\nwith ThreadPoolExecutor(max_workers=4) as executor:
\n futures = []
\n for img_file in os.listdir(image_dir):
\n if img_file.endswith(('.jpg', '.jpeg', '.png')):
\n img_path = os.path.join(image_dir, img_file)
\n future = executor.submit(analyze_image, img_path, questions)
\n futures.append((img_file, future))<\/p>\n
for img_file, future in futures:
\n try:
\n result = future.result(timeout=60)
\n results.append((img_file, result))
\n print(f"{img_file}: {result[:50]}…")
\n except Exception as e:
\n print(f"{img_file} \u5904\u7406\u5931\u8d25: {e}")<\/p>\n
\u4f60\u53ef\u4ee5\u628aOstrakon-VL\u96c6\u6210\u5230\u73b0\u6709\u7cfb\u7edf\u4e2d：<\/p>\n
# \u793a\u4f8b：\u96f6\u552e\u5de1\u68c0\u7cfb\u7edf\u96c6\u6210
\nclass RetailInspectionSystem:
\n def __init__(self, model_endpoint="http:\/\/localhost:8000\/v1\/completions"):
\n self.endpoint = model_endpoint<\/p>\n
def inspect_shelf(self, image_path, shelf_id):
\n """\u68c0\u67e5\u8d27\u67b6\u72b6\u6001"""
\n analysis = self._analyze_image(image_path, "\u8bf7\u5206\u6790\u8d27\u67b6\u72b6\u6001：\u5546\u54c1\u662f\u5426\u9f50\u5168？\u6446\u653e\u662f\u5426\u6574\u9f50？")<\/p>\n
# \u89e3\u6790\u7ed3\u679c，\u751f\u6210\u62a5\u544a
\n report = {
\n "shelf_id": shelf_id,
\n "timestamp": datetime.now().isoformat(),
\n "analysis": analysis,
\n "issues": self._extract_issues(analysis),
\n "recommendations": self._generate_recommendations(analysis)
\n }<\/p>\n
return report<\/p>\n
def check_compliance(self, image_path, regulation_type):
\n """\u68c0\u67e5\u5408\u89c4\u6027"""
\n question = f"\u8fd9\u5f20\u56fe\u7247\u662f\u5426\u7b26\u5408{regulation_type}\u89c4\u5b9a？\u8bf7\u6307\u51fa\u95ee\u9898\u3002"
\n return self._analyze_image(image_path, question)<\/p>\n
def _analyze_image(self, image_path, question):
\n """\u8c03\u7528\u6a21\u578b\u5206\u6790\u56fe\u7247"""
\n # … \u5b9e\u73b0\u56fe\u7247\u5206\u6790\u548cAPI\u8c03\u7528
\n pass<\/p>\n
\u5728\u751f\u4ea7\u73af\u5883\u4e2d，\u4f60\u9700\u8981\u76d1\u63a7\u6a21\u578b\u670d\u52a1：<\/p>\n
# monitoring.py
\nimport time
\nimport requests
\nimport psutil
\nfrom datetime import datetime<\/p>\n
class ModelMonitor:
\n def __init__(self, endpoint="http:\/\/localhost:8000\/health"):
\n self.endpoint = endpoint<\/p>\n
def check_health(self):
\n """\u68c0\u67e5\u670d\u52a1\u5065\u5eb7\u72b6\u6001"""
\n try:
\n start = time.time()
\n response = requests.get(self.endpoint, timeout=5)
\n latency = (time.time() – start) * 1000 # \u6beb\u79d2<\/p>\n
return {
\n "timestamp": datetime.now().isoformat(),
\n "status": "healthy" if response.status_code == 200 else "unhealthy",
\n "latency_ms": round(latency, 2),
\n "memory_usage": self._get_memory_usage()
\n }
\n except Exception as e:
\n return {
\n "timestamp": datetime.now().isoformat(),
\n "status": "error",
\n "error": str(e)
\n }<\/p>\n
def _get_memory_usage(self):
\n """\u83b7\u53d6\u5185\u5b58\u4f7f\u7528\u60c5\u51b5"""
\n process = psutil.Process()
\n memory_info = process.memory_info()
\n return {
\n "rss_mb": memory_info.rss \/ 1024 \/ 1024,
\n "vms_mb": memory_info.vms \/ 1024 \/ 1024
\n }<\/p>\n
# \u5b9a\u65f6\u76d1\u63a7
\nmonitor = ModelMonitor()
\nwhile True:
\n status = monitor.check_health()
\n print(f"\u72b6\u6001\u68c0\u67e5: {status}")
\n time.sleep(60) # \u6bcf\u5206\u949f\u68c0\u67e5\u4e00\u6b21<\/p>\n
\u8d70\u5230\u8fd9\u91cc，\u4f60\u5df2\u7ecf\u6210\u529f\u5728ARM\u670d\u52a1\u5668\u4e0a\u90e8\u7f72\u4e86Ostrakon-VL-8B\u6a21\u578b，\u5e76\u4e14\u6709\u4e86\u4e00\u4e2a\u53ef\u7528\u7684\u56fe\u6587\u5bf9\u8bdd\u7cfb\u7edf\u3002\u8ba9\u6211\u4eec\u56de\u987e\u4e00\u4e0b\u4eca\u5929\u7684\u6210\u679c：<\/p>\n
\u4f60\u5f97\u5230\u4e86\u4ec0\u4e48？<\/p>\n
\u8fd9\u4e2a\u65b9\u6848\u6709\u4ec0\u4e48\u4f18\u52bf？<\/p>\n
\u63a5\u4e0b\u6765\u53ef\u4ee5\u505a\u4ec0\u4e48？<\/p>\n
\u6700\u540e\u7684\u5c0f\u5efa\u8bae：<\/p>\n
\u96f6\u552e\u548c\u9910\u996e\u884c\u4e1a\u7684\u6570\u5b57\u5316\u6b63\u5728\u52a0\u901f，AI\u89c6\u89c9\u5206\u6790\u662f\u4e00\u4e2a\u5f3a\u5927\u7684\u5de5\u5177\u3002\u73b0\u5728\u4f60\u6709\u4e86\u81ea\u5df1\u7684AI\u52a9\u624b，\u53ef\u4ee5\u5f00\u59cb\u63a2\u7d22\u5404\u79cd\u5e94\u7528\u573a\u666f\u4e86\u3002\u4ece\u8d27\u67b6\u5de1\u68c0\u5230\u98df\u54c1\u5b89\u5168\u68c0\u67e5，\u4ece\u987e\u5ba2\u884c\u4e3a\u5206\u6790\u5230\u667a\u80fd\u63a8\u8350，\u53ef\u80fd\u6027\u662f\u65e0\u9650\u7684\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":"
Ostrakon-VL-8B\u90e8\u7f72\u6559\u7a0b：ARM\u67b6\u6784\u670d\u52a1\u5668（\u5982NVIDIA Grace）\u4e0a\u7684vLLM\u517c\u5bb9\u90e8\u7f72
\n1. \u5f00\u7bc7：\u4e3a\u4ec0\u4e48\u4f60\u9700\u8981\u5173\u6ce8\u8fd9\u4e2a\u6a21\u578b？
\n\u5982\u679c\u4f60\u6b63\u5728\u5bfb\u627e\u4e00\u4e2a\u80fd\u771f\u6b63\u770b\u61c2\u96f6\u552e\u548c\u9910\u996e\u573a\u666f\u56fe\u7247\u7684AI\u52a9\u624b，\u90a3\u4e48Ostrakon-VL-8B\u7edd\u5bf9\u503c\u5f97\u4f60\u82b1\u65f6\u95f4\u4e86\u89e3\u4e00\u4e0b\u3002
\n\u60f3\u8c61\u4e00\u4e0b\u8fd9\u6837\u7684\u573a\u666f：\u4f60\u6709\u4e00\u5bb6\u8fde\u9501\u8d85\u5e02，\u6bcf\u5929\u8981\u5904\u7406\u6210\u5343\u4e0a\u4e07\u7684\u8d27\u67b6\u7167\u7247，\u9700\u8981\u68c0\u67e5\u5546\u54c1\u6446\u653e<\/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":[9300,9301,3781],"topic":[],"class_list":["post-83227","post","type-post","status-publish","format-standard","hentry","category-server","tag-ai","tag-3781"],"yoast_head":"\n