本文属于机器翻译版本。若本译文内容与英语原文存在差异，则一律以英文原文为准。

# 先决条件
<a name="model-customize-mtrl-prereqs"></a>

Multi-turn RL 使用 SageMaker AI `CreateJob` API，除了现有的模型自定义先决条件外，还需要其他权限。下表汇总了新增内容与已涵盖内容。


| 内容 | Status | 注意 | 
| --- | --- | --- | 
| 信任政策：job.sagemaker.amazonaws.com | New |  CreateJob API 是必需的 | 
| 信任政策：bedrock-agentcore.amazonaws.com | New | 在代理运行时角色上是必需的（仅限AgentCore 路径） | 
| PassRole: job.sagemaker.amazonaws.com | New | 现有的先决条件仅涵盖 sagemaker.amazonaws.com | 
| Job 操作（CreateJob DescribeJob、等） | New | 不在 AmazonSageMakerFullAccess | 
| bedrock-agentcore: ListAgentRuntimes | New | Studio 运行时选择器是必需的 | 
| bedrock-agentcore: ListAgentRuntimeVersions | New | Studio 版本选择器必填项 | 
| AmazonSageMakerJobFullAccess 托管策略 | New | 附加到任务执行角色 | 
| AmazonSageMakerJobRuntimeAccess 托管策略 | New | 附加到代理运行时角色 | 
| Lambda、Bedrock 部署、S3、KMS、mlFlow、 CloudWatch | 已经涵盖了 | 通过现有的先决条件或托管策略 | 

## 来电者角色
<a name="model-customize-mtrl-prereqs-caller-role"></a>

```
{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Sid": "PassRoleForCreateJob",
      "Effect": "Allow",
      "Action": "iam:PassRole",
      "Resource": "arn:aws:iam::<account-id>:role/<ExecutionRoleName>",
      "Condition": {
        "StringEquals": {
          "iam:PassedToService": "job.sagemaker.amazonaws.com"
        }
      }
    },
    {
      "Sid": "MTRLJobActions",
      "Effect": "Allow",
      "Action": [
        "sagemaker:CreateJob",
        "sagemaker:DescribeJob",
        "sagemaker:StopJob",
        "sagemaker:DeleteJob"
      ],
      "Resource": "arn:aws:sagemaker:*:*:job/*",
      "Condition": {
        "StringEquals": {
          "aws:ResourceAccount": "${aws:PrincipalAccount}"
        }
      }
    },
    {
      "Sid": "MTRLListJobsPermission",
      "Effect": "Allow",
      "Action": [
        "sagemaker:ListJobs",
        "sagemaker:ListJobSchemaVersions",
        "sagemaker:DescribeJobSchemaVersion"
      ],
      "Resource": "*",
      "Condition": {
        "StringEquals": {
          "aws:ResourceAccount": "${aws:PrincipalAccount}"
        }
      }
    }
  ]
}
```

PassRole — 让调用者将执行角色委托给`job.sagemaker.amazonaws.com`。

## 托管策略
<a name="model-customize-mtrl-prereqs-managed-policies"></a>

Multi-turn RL 使用两个 Amazon 托管策略。在创建任务之前，将每个人都分配到相应的角色。

**AmazonSageMakerJobFullAccess**— 附加到 A SageMaker I 任务执行角色（`RoleArn`您传递给的`CreateJob`）。

此策略向作业授予运行时所需的权限： read/write 访问 S3 中的训练数据和检查点、KMS S3-side 加密权限、访问 SageMaker AI 中心内容和模型包、mlFlow 实验跟踪、通过 Bedrock 或 L AgentCore ambda 调用代理、VPC 网络接口管理和日志写入。 CloudWatch 参考：[AmazonSageMakerJobFullAccess](https://docs.amazonaws.cn/aws-managed-policy/latest/reference/AmazonSageMakerJobFullAccess.html)。

**AmazonSageMakerJobRuntimeAccess**— 附加到代理运行时角色（您的代理所处的角色，无论是在 Bedrock AgentCore 上还是 Lambda 转发器后面）。此策略授予您的代理在推出期间调用的运行时 API：`sagemaker:Sample`、`sagemaker:SampleWithResponseStream`、`sagemaker:CompleteRollout``sagemaker:UpdateReward`、和`sagemaker:CallWithBearerToken`。参考：[AmazonSageMakerJobRuntimeAccess](https://docs.amazonaws.cn/aws-managed-policy/latest/reference/AmazonSageMakerJobRuntimeAccess.html)。

## 信任政策更新
<a name="model-customize-mtrl-prereqs-trust-policy"></a>

将以下`job.sagemaker.amazonaws.com`内容添加到您的执行角色的信任策略中：

```
{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Effect": "Allow",
      "Principal": {
        "Service": "job.sagemaker.amazonaws.com"
      },
      "Action": ["sts:AssumeRole", "sts:TagSession"]
    }
  ]
}
```

如果使用 Bedrock AgentCore，则您的代理运行时角色还需要自己的信任策略：

```
{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Effect": "Allow",
      "Principal": {
        "Service": "bedrock-agentcore.amazonaws.com"
      },
      "Action": "sts:AssumeRole"
    }
  ]
}
```

## Studio 用户界面的其他内联政策
<a name="model-customize-mtrl-prereqs-studio-policy"></a>

如果您通过 SageMaker AI Studio 提交作业，请将以下内联策略添加到您的 SageMaker AI 域执行角色中：

```
{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Sid": "MTRLJobActions",
      "Effect": "Allow",
      "Action": [
        "sagemaker:CreateJob",
        "sagemaker:DescribeJob",
        "sagemaker:StopJob",
        "sagemaker:DeleteJob",
        "sagemaker:ListJobs",
        "sagemaker:ListJobSchemaVersions",
        "sagemaker:DescribeJobSchemaVersion"
      ],
      "Resource": "arn:aws:sagemaker:*:*:job/*",
      "Condition": {
        "StringEquals": {
          "aws:ResourceAccount": "${aws:PrincipalAccount}"
        }
      }
    },
    {
      "Sid": "BedrockAgentCoreListPermissions",
      "Effect": "Allow",
      "Action": [
        "bedrock-agentcore:ListAgentRuntimes",
        "bedrock-agentcore:ListAgentRuntimeVersions"
      ],
      "Resource": "*",
      "Condition": {
        "StringEquals": {
          "aws:ResourceAccount": "${aws:PrincipalAccount}"
        }
      }
    },
    {
      "Sid": "PassRoleForCreateJob",
      "Effect": "Allow",
      "Action": "iam:PassRole",
      "Resource": "arn:aws:iam::*:role/*",
      "Condition": {
        "StringEquals": {
          "iam:PassedToService": "job.sagemaker.amazonaws.com",
          "aws:ResourceAccount": "${aws:PrincipalAccount}"
        }
      }
    }
  ]
}
```

## 自定义策略选项
<a name="model-customize-mtrl-prereqs-custom-policy"></a>

如果您更喜欢定义自己的策略而不是使用 Amazon 托管策略，请使用以下内容作为起点。

**SageMaker AI 任务执行角色策略**（等同于`AmazonSageMakerJobFullAccess`）：

```
{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Sid": "S3Permissions",
      "Effect": "Allow",
      "Action": [
        "s3:GetObject",
        "s3:PutObject",
        "s3:ListBucket"
      ],
      "Resource": "*",
      "Condition": {
        "StringEquals": {
          "s3:ResourceAccount": "${aws:PrincipalAccount}"
        }
      }
    },
    {
      "Sid": "KMSPermissions",
      "Effect": "Allow",
      "Action": [
        "kms:Decrypt",
        "kms:GenerateDataKey"
      ],
      "Resource": "arn:aws:kms:*:*:key/*",
      "Condition": {
        "StringEquals": {
          "aws:ResourceAccount": "${aws:PrincipalAccount}"
        },
        "StringLike": {
          "kms:ViaService": "s3.*.amazonaws.com"
        }
      }
    },
    {
      "Sid": "KMSDescribeKey",
      "Effect": "Allow",
      "Action": "kms:DescribeKey",
      "Resource": "arn:aws:kms:*:*:key/*",
      "Condition": {
        "StringEquals": {
          "aws:ResourceAccount": "${aws:PrincipalAccount}"
        }
      }
    },
    {
      "Sid": "SageMakerHubPermissions",
      "Effect": "Allow",
      "Action": [
        "sagemaker:DescribeHubContent"
      ],
      "Resource": [
        "arn:aws:sagemaker:*:*:hub/*",
        "arn:aws:sagemaker:*:*:hub-content/*"
      ],
      "Condition": {
        "StringEquals": {
          "aws:ResourceAccount": "${aws:PrincipalAccount}"
        }
      }
    },
    {
      "Sid": "SageMakerModelPackagePermissions",
      "Effect": "Allow",
      "Action": [
        "sagemaker:AccessModelPackage",
        "sagemaker:CreateModelPackage",
        "sagemaker:DescribeModelPackage",
        "sagemaker:DescribeModelPackageGroup"
      ],
      "Resource": [
        "arn:aws:sagemaker:*:*:model-package/*",
        "arn:aws:sagemaker:*:*:model-package-group/*"
      ],
      "Condition": {
        "StringEquals": {
          "aws:ResourceAccount": "${aws:PrincipalAccount}"
        }
      }
    },
    {
      "Sid": "MLflowPermissions",
      "Effect": "Allow",
      "Action": [
        "sagemaker:DescribeMlflowApp",
        "sagemaker:CallMlflowAppApi",
        "sagemaker-mlflow:CreateExperiment",
        "sagemaker-mlflow:CreateRun",
        "sagemaker-mlflow:UpdateRun",
        "sagemaker-mlflow:LogBatch",
        "sagemaker-mlflow:GetExperimentByName",
        "sagemaker-mlflow:GetMetricHistory",
        "sagemaker-mlflow:GetRun",
        "sagemaker-mlflow:StartTrace",
        "sagemaker-mlflow:EndTrace",
        "sagemaker-mlflow:SearchTraces",
        "sagemaker-mlflow:ListArtifacts"
      ],
      "Resource": [
        "arn:aws:sagemaker:*:*:mlflow-app/*"
      ],
      "Condition": {
        "StringEquals": {
          "aws:ResourceAccount": "${aws:PrincipalAccount}"
        }
      }
    },
    {
      "Sid": "BedrockAgentCorePermissions",
      "Effect": "Allow",
      "Action": [
        "bedrock-agentcore:InvokeAgentRuntime",
        "bedrock-agentcore:StopRuntimeSession",
        "bedrock-agentcore:GetAgentRuntime"
      ],
      "Resource": "arn:aws:bedrock-agentcore:*:*:runtime/*",
      "Condition": {
        "StringEquals": {
          "aws:ResourceAccount": "${aws:PrincipalAccount}"
        }
      }
    },
    {
      "Sid": "EC2NetworkPermissions",
      "Effect": "Allow",
      "Action": [
        "ec2:CreateNetworkInterface",
        "ec2:CreateNetworkInterfacePermission",
        "ec2:DescribeNetworkInterfaces",
        "ec2:DescribeVpcs",
        "ec2:DescribeSubnets",
        "ec2:DescribeSecurityGroups",
        "ec2:DescribeDhcpOptions"
      ],
      "Resource": "*",
      "Condition": {
        "StringEquals": {
          "aws:ResourceAccount": "${aws:PrincipalAccount}"
        }
      }
    },
    {
      "Sid": "EC2NetworkInterfaceTagPermissions",
      "Effect": "Allow",
      "Action": [
        "ec2:CreateTags"
      ],
      "Resource": "arn:aws:ec2:*:*:network-interface/*",
      "Condition": {
        "StringEquals": {
          "aws:ResourceAccount": "${aws:PrincipalAccount}",
          "ec2:CreateAction": "CreateNetworkInterface"
        }
      }
    },
    {
      "Sid": "EC2NetworkInterfaceDeletePermissions",
      "Effect": "Allow",
      "Action": [
        "ec2:DeleteNetworkInterface",
        "ec2:DeleteNetworkInterfacePermission"
      ],
      "Resource": "*",
      "Condition": {
        "StringEquals": {
          "aws:ResourceAccount": "${aws:PrincipalAccount}"
        }
      }
    },
    {
      "Sid": "CloudWatchLogsPermissions",
      "Effect": "Allow",
      "Action": [
        "logs:CreateLogGroup",
        "logs:CreateLogStream",
        "logs:PutLogEvents",
        "logs:DescribeLogStreams"
      ],
      "Resource": "arn:aws:logs:*:*:log-group:/aws/sagemaker/*",
      "Condition": {
        "StringEquals": {
          "aws:ResourceAccount": "${aws:PrincipalAccount}"
        }
      }
    },
    {
      "Sid": "LambdaPermissions",
      "Effect": "Allow",
      "Action": [
        "lambda:InvokeFunction"
      ],
      "Resource": "arn:aws:lambda:*:*:function:*",
      "Condition": {
        "StringEquals": {
          "aws:ResourceAccount": "${aws:PrincipalAccount}"
        }
      }
    }
  ]
}
```

**执行角色的信任关系：**

```
{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Effect": "Allow",
      "Principal": {
        "Service": "job.sagemaker.amazonaws.com"
      },
      "Action": ["sts:AssumeRole", "sts:TagSession"]
    }
  ]
}
```

**AmazonSageMakerJobRuntimeAccess 等效 \+ BedrockAgentCoreFullAccess**

```
{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Sid": "SageMakerJobRuntimePermissions",
      "Effect": "Allow",
      "Action": [
        "sagemaker:Sample",
        "sagemaker:SampleWithResponseStream",
        "sagemaker:CompleteRollout",
        "sagemaker:UpdateReward"
      ],
      "Resource": "arn:aws:sagemaker:*:*:job/*",
      "Condition": {
        "StringEquals": {
          "aws:ResourceAccount": "${aws:PrincipalAccount}"
        }
      }
    },
    {
      "Sid": "BearerTokenPermissions",
      "Effect": "Allow",
      "Action": [
        "sagemaker:CallWithBearerToken"
      ],
      "Resource": "*",
      "Condition": {
        "StringEquals": {
          "aws:ResourceAccount": "${aws:PrincipalAccount}"
        }
      }
    }
  ]
}
```

**代理运行时角色的信任关系：**

```
{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Effect": "Allow",
      "Principal": {
        "Service": "bedrock-agentcore.amazonaws.com"
      },
      "Action": "sts:AssumeRole"
    }
  ]
}
```

## Lambda 转发器注意事项
<a name="model-customize-mtrl-prereqs-lambda-forwarder"></a>

现有的模型自定义 Lambda 权限涵盖名称中带有*SageMaker*或 *sagemaker* 的函数。如果您的 Lambda 转发器使用不同的命名约定，请明确添加其 ARN：

```
{
  "Sid": "CustomAgentLambdaPermission",
  "Effect": "Allow",
  "Action": ["lambda:InvokeFunction"],
  "Resource": "arn:aws:lambda:*:*:function:your-agent-forwarder-function-name",
  "Condition": {
    "StringEquals": {"aws:ResourceAccount": "${aws:PrincipalAccount}"}
  }
}
```

## 其他设置
<a name="model-customize-mtrl-prereqs-other-setup"></a>
+ 如果您使用客户托管的 VPC，请参阅[为多轮 RL 任务配置 VPC](model-customize-mtrl-vpc.md)。
+ 如果您使用 KMS 密钥加密任务输入和输出，则执行角色和调用者角色需要额外的权限。请参阅[用于多回合强化学习的静态加密](model-customize-mtrl-encryption-at-rest.md)。