如何通过Azure Python SDK v2部署自定义Hugging Face模型（如AudioGen）

问题背景 #

部分Hugging Face模型（如facebook/audiogen-medium）未收录在Azure模型目录，也无一键部署按钮。使用Azure Python SDK v2部署时，存在模型注册路径困惑、注册报错，以及端点部署步骤待确认的问题。

本地调用模型代码：

from audiocraft.models import AudioGen
    
model = AudioGen.get_pretrained("facebook/audiogen-medium")
model.set_generation_params(duration=5)  # 生成5秒音频
descriptions = ['dog barking', 'sirene of an emergency vehicle', 'footsteps in a corridor']
wav = model.generate(descriptions)  # 生成3个样本

注册模型时因path属性填写错误触发ValueError: No value for given attribute，报错堆栈见用户提供内容。已编写score.py推理脚本及端点部署代码。

解决方案 #

一、模型注册的正确方式 #

Azure ML注册模型时，path指向包含模型文件+推理脚本的完整目录，而非单独文件。针对audiocraft有两种可行方案：

方案1：提前下载模型并打包注册

1. 本地下载模型并保存：

from audiocraft.models import AudioGen
model = AudioGen.get_pretrained("facebook/audiogen-medium")
model.save_pretrained("./audiogen-model")  # 保存到本地目录

2. 整理目录结构：

audiogen-model/
├── 模型权重文件（由save_pretrained生成）
└── score.py  # 你的推理脚本

3. 注册模型：

from azure.ai.ml.entities import Model
from azure.ai.ml.constants import AssetTypes

model = Model(
    name="audiogen-medium",
    version="1",
    path="./audiogen-model",
    type=AssetTypes.CUSTOM_MODEL,
    description="Facebook AudioGen Medium模型"
)
ml_client.models.create_or_update(model)

方案2：依赖audiocraft自动下载（无需提前打包）

若不想提前下载模型，可仅注册score.py，但需在部署环境中配置所有依赖，确保audiocraft能自动拉取模型权重。

二、解决注册时的ValueError报错 #

从堆栈信息看，问题出在workspaceName序列化为空，检查ML客户端初始化代码，确保正确传入工作区名称：

from azure.ai.ml import MLClient
from azure.identity import DefaultAzureCredential

ml_client = MLClient(
    DefaultAzureCredential(),
    subscription_id="你的订阅ID",
    resource_group_name="你的资源组名",
    workspace_name="你的工作区名"  # 必须非空且正确
)

三、score.py优化（增加鲁棒性） #

import json
import os
import traceback
from audiocraft.models import AudioGen
from audiocraft.data.audio import audio_write
from azure.storage.blob import BlobServiceClient, BlobClient, ContainerClient

def init():
    global model, blob_service_client, container_name
    try:
        model = AudioGen.get_pretrained('facebook/audiogen-medium')
        model.set_generation_params(duration=5)
        
        connect_str = os.getenv('AZURE_STORAGE_CONNECTION_STRING')
        if not connect_str:
            raise ValueError("未设置AZURE_STORAGE_CONNECTION_STRING环境变量")
        blob_service_client = BlobServiceClient.from_connection_string(connect_str)
        container_name = 'audiogen_generated_files'
        # 自动创建不存在的容器
        if not blob_service_client.get_container_client(container_name).exists():
            blob_service_client.create_container(container_name)
    except Exception as e:
        traceback.print_exc()
        raise e

def run(payload: str):
    try:
        data = json.loads(payload)
        theme = data.get("theme")
        if not theme:
            return {"error": "缺少theme参数"}
        
        prompts = [f'{theme}, slow speed', f'{theme}, fast speed', f'{theme} door closing', f'{theme} starting']
        wavs = model.generate(prompts)
        
        urls = []
        os.makedirs('/tmp', exist_ok=True)  # 确保临时目录存在
        
        for idx, one_wav in enumerate(wavs):
            file_name = f'{theme}_{idx}.wav'
            local_file_path = f'/tmp/{file_name}'
            audio_write(local_file_path, one_wav.cpu(), model.sample_rate, strategy="loudness", loudness_compressor=True)
            
            blob_client = blob_service_client.get_blob_client(container=container_name, blob=file_name)
            with open(local_file_path, "rb") as f:
                blob_client.upload_blob(f, overwrite=True)
            urls.append(blob_client.url)
            os.remove(local_file_path)
        
        return {"urls": urls}
    except Exception as e:
        traceback.print_exc()
        return {"error": str(e)}

四、端点与部署步骤修正 #

原部署代码缺少环境配置（需安装audiocraft等依赖），且存在拼写错误（deployement改为deployment），完整代码如下：

1. 准备conda依赖文件conda.yml：

name: audiogen-env
channels:
  - defaults
  - conda-forge
dependencies:
  - python=3.10
  - pip=23.0.1
  - pip:
    - audiocraft==0.1.7
    - azure-storage-blob==12.19.0
    - torch==2.0.1
    - torchaudio==2.0.2

2. 完整部署代码：

import uuid
from azure.ai.ml.entities import ManagedOnlineEndpoint, ManagedOnlineDeployment, Environment
from azure.ai.ml.constants import EnvironmentType

# 创建在线端点
endpoint = ManagedOnlineEndpoint(
    name=f"audiogen-endpoint-{str(uuid.uuid4())[:8]}",
    description="AudioGen模型推理端点",
    auth_mode="key"
)
endpoint = ml_client.online_endpoints.begin_create_or_update(endpoint).result()

# 创建自定义环境
env = Environment(
    name="audiogen-env",
    version="1",
    description="AudioGen模型运行环境",
    conda_file="./conda.yml",
    image="mcr.microsoft.com/azureml/openmpi4.1.0-ubuntu20.04:latest"
)
ml_client.environments.create_or_update(env)

# 创建部署
deployment = ManagedOnlineDeployment(
    name="audiogen-deployment-mo",
    endpoint_name=endpoint.name,
    model=ml_client.models.get(name="audiogen-medium", version="1"),
    environment=env,
    instance_count=1,
    instance_type="Standard_NC6s_v3",  # 必须选择GPU实例，AudioGen依赖CUDA
    environment_variables={"AZURE_STORAGE_CONNECTION_STRING": "你的Blob存储连接字符串"}
)
deployment = ml_client.begin_create_or_update(deployment).result()

# 设置流量全部分配到该部署
ml_client.online_endpoints.begin_create_or_update(
    ManagedOnlineEndpoint(
        name=endpoint.name,
        traffic={"audiogen-deployment-mo": 100}
    )
).result()

关键注意事项 #

• AudioGen依赖GPU加速，必须选择支持GPU的实例类型（如Standard_NC6s_v3、Standard_NV6等）
• 确保Azure工作区有足够的GPU实例配额
• 若不需要持久化音频，可修改run函数直接返回音频字节流，无需上传Blob存储

内容的提问来源于stack exchange，提问作者Momo