基于昇腾通用推理镜像的自定义部署 四、总结 昇腾通用推理镜像自定义部署最佳实践,覆盖代码包准备、模型准备、部署流程全流程。该实践可适配多类场景,有效提升部署效率与服务稳定性。未来将进一步优化推理部署体系,为企业 AI 业务规模化落地提供有力支撑。 附录 实例中使用的demo代码 downloadmodelandexportmodelpath.sh plaintext mkdir Qwen20.5BInstruct cd Qwen20.5BInstruct wget q nocheckcertificate wget q nocheckcertificate wget q nocheckcertificate wget q nocheckcertificate wget q nocheckcertificate wget q nocheckcertificate wget q nocheckcertificate wget q nocheckcertificate exportMODELPATHpwd cd.. llm.py plaintext from fastapi import FastAPI, Request from transformers import pipeline import json import torch import torchnpu import os app FastAPI() generator pipeline('textgeneration', modelos.environ.get('MODELPATH'), devicetorch.device('npu:0')) @app.post("/v1/chat/completions") async def createcompletion(request: Request): data await request.json() response generator(data.get('messages'), maxnewtokens 500) return{ "choices":[ { "message":response[0]['generatedtext'][1], "finishreason":"stop", "index":0 } ], "id":"chatcmpl1234567890", "model":"Qwen20.5BInstruct", "object":"chat.completion" } if name "main": import uvicorn uvicorn.run(app, host"0.0.0.0", port8899) test.sh plaintext curl location ' header 'ContentType: application/json' data '{ "messages": [ {"role": "system","content": "你是个客服"}, {"role": "user","content": "在吗"} ] }'