# Amazon Nova 配方
<a name="nova-model-recipes"></a>

您可以从[配方](https://github.com/aws/sagemaker-hyperpod-recipes/tree/main/recipes_collection/recipes)存储库中获取 Amazon Nova 配方。Amazon Nova 配方是指 [YAML](https://yaml.org/) 配置文件，该文件向 SageMaker 提供有关如何运行模型自定义作业的详细信息。它提供基本模型名称、设置训练超参数、定义优化设置，并且包括成功微调或训练模型所需的任何其他选项。

您还可以通过 Amazon SageMaker Studio 访问 Amazon Nova 配方：进入模型中心，选择 AWS，浏览 Amazon Nova 模型，即可找到其关联的配方。Amazon SageMaker Studio 为每个配方都提供了示例笔记本，其中包括修改配方以及使用 SageMaker 训练作业或 SageMaker HyperPod 环境运行自定义作业的所有必要步骤。

要访问 Amazon SageMaker Studio 中的配方页面，执行角色必须具有以下权限。

------
#### [ JSON ]

****  

```
{
    "Version":"2012-10-17",		 	 	 
    "Statement": [
        {
            "Effect": "Allow",
            "Action": [
                "s3:GetObject"
            ],
            "Resource": [
                "arn:aws:s3:::*model-customization-recipes*"
            ]
        }
    ]
}
```

------

要在 SageMaker 训练作业上执行示例笔记本，请使用以下 SageMaker 分发映像版本之一：`2.7.1+`、`2.8.0+`、`3.2.1+`、`3.3.0+`。这适用于 Amazon SageMaker Studio。

**Topics**
+ [获取 Amazon Nova 配方](#nova-model-get-recipes)
+ [获取 Amazon Nova Forge 配方](#nova-model-get-forge-recipes)
+ [可用的模型和算法](#nova-model-algorithm)
+ [Amazon Nova Lite](#nova-model-recipes-reference-novalite)
+ [Amazon Nova Micro](#nova-model-recipes-reference-novamicro)
+ [Amazon Nova Pro](#nova-model-recipes-reference-novapro)

## 获取 Amazon Nova 配方
<a name="nova-model-get-recipes"></a>

要获取基本 Amazon Nova 配方，请运行以下命令克隆[配方](https://github.com/aws/sagemaker-hyperpod-recipes/tree/main/recipes_collection/recipes)存储库。

```
git clone https://github.com/aws/sagemaker-hyperpod-recipes.git
```

可通过 [https://github.com/aws/sagemaker-hyperpod-recipes/tree/main/recipes_collection/recipes/](https://github.com/aws/sagemaker-hyperpod-recipes/tree/main/recipes_collection/recipes/) 获得基本配方。

```
cd recipes_collection/recipes/
```

Amazon Nova 自定义配方位于以下文件夹中。


****  

| 食谱类型 | Folder | 
| --- | --- | 
| SFT（全秩和 PEFT）、PPO、DPO（全秩和 PEFT） | [fine-tuning/nova](https://github.com/aws/sagemaker-hyperpod-recipes/tree/main/recipes_collection/recipes/fine-tuning/nova) | 
| 评估 | [evaluation/nova](https://github.com/aws/sagemaker-hyperpod-recipes/tree/main/recipes_collection/recipes/evaluation/nova) | 
| CPT | [training/nova](https://github.com/aws/sagemaker-hyperpod-recipes/tree/main/recipes_collection/recipes/training) | 

## 获取 Amazon Nova Forge 配方
<a name="nova-model-get-forge-recipes"></a>

如需访问作业专用的 Amazon Nova Forge 配方，请先完成 [Amazon Nova Forge 访问权限设置](nova-forge-access.md)，然后按照[本讲习会](https://catalog.us-east-1.prod.workshops.aws/workshops/dcac6f7a-3c61-4978-8344-7535526bf743/en-US/02-smhp-rig/03-nova-forge)完成 Forge 版 CLI 的设置。对于 SageMaker 训练作业，AWS 管理控制台和 Nova Forge SDK 将自动访问 Forge 配方。

## 可用的模型和算法
<a name="nova-model-algorithm"></a>

下表汇总了 Amazon Nova 模型自定义功能的可用情况，以及使用 SageMaker 的支持算法。


| 模型名称 | 模型 ID | 微调 | 备注 | 
| --- |--- |--- |--- |
| Amazon Nova Lite 2.0 | amazon.nova-2-lite-v1:0:256k | 是 |    SFT：接受文本和/或图像或文本和/或视频作为输入，并生成文本作为输出。单个作业不能在同一次运行中合并文本、图像和视频。   DPO：接受文本和图像作为输入，并生成文本作为输出。   RFT：支持以文本和图像作为单轮对话的输入，并输出文本。具备更优的奖励优化能力。    | 

## Amazon Nova Lite
<a name="nova-model-recipes-reference-novalite"></a>

下表列出了 Amazon Nova Lite 配方参考的详细信息。


****  

| 模型 | 类别/子类别 | 方法 | 配方名称 | 映像 URI（训练作业） | 映像 URI（SageMaker HyperPod） | 计算实例 | 
| --- | --- | --- | --- | --- | --- | --- | 
| Nova Lite | 训练/微调 | 监督式微调（LoRA） | `nova_lite_1_0_g5_g6_12x_gpu_lora_sft.yaml`<br />`nova_lite_1_0_g5_g6_48x_gpu_lora_sft.yaml`<br />`nova_lite_1_0_p5_p4d_gpu_lora_sft.yaml` | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-TJ-SFT-latest | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-HP-SFT-latest | `ml.p5.48xlarge`, `ml.p5en.48xlarge`,<br />`ml.g5.12xlarge`, `ml.g6.12xlarge`,<br />`ml.g5.48xlarge`, `ml.g6.48xlarge` | 
| Nova Lite | 训练/微调 | 监督式微调（全秩） | nova\_lite\_1\_0\_p5\_p4d\_gpu\_sft.yaml | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-TJ-SFT-latest | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-HP-SFT-latest | ml.p5.48xlarge, ml.p5en.48xlarge | 
| Nova Lite | 训练/微调 | 直接偏好优化（全秩） | nova\_lite\_1\_0\_p5\_p4d\_gpu\_dpo.yaml | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-TJ-DPO-latest | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-HP-DPO-latest | ml.p5.48xlarge, ml.p5en.48xlarge | 
| Nova Lite | 训练/微调 | 直接偏好优化（LoRA） | `nova_lite_1_0_p5_p4d_gpu_lora_dpo.yaml`<br />`nova_lite_1_0_g5_g6_48x_gpu_lora_dpo.yaml` | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-TJ-DPO-latest | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-HP-DPO-latest | `ml.p5.48xlarge`, `ml.p5en.48xlarge`,<br />`ml.g5.48xlarge`, `ml.g6.48xlarge` | 
| Nova Lite | 训练/强化学习 | 强化微调（RFT） | nova\_lite\_1\_0\_p5\_gpu\_ppo.yaml | 不适用 | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SMHP-PPO-TRAIN-latest | ml.p5.48xlarge, ml.p5en.48xlarge | 
| Nova Lite | 训练/持续预训练 | 持续预训练（基本模型） | nova\_lite\_gpu\_p5x16\_pretrain.yaml | 不适用 | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:HP-CPT-latest | ml.p5.48xlarge | 
| Nova Lite | 评估/评估 | 标准文本基准测试 | nova\_lite\_p5\_48xl\_general\_text\_benchmark\_eval.yaml | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-TJ-Eval-latest | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-HP-Eval-latest | ml.p5.48xlarge | 
| Nova Lite | 评估/评估 | 自定义数据集评估 | nova\_lite\_p5\_48xl\_bring\_your\_own\_dataset\_eval.yaml | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-TJ-Eval-latest | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-HP-Eval-latest | ml.p5.48xlarge | 
| Nova Lite | 评估/评估 | 多模态基准测试 | nova\_lite\_p5\_48\_general\_multi\_modal\_benchmark\_eval.yaml | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-TJ-Eval-latest | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-HP-Eval-latest | ml.p5.48xlarge | 
| Nova Lite | 评估/评估 | LLM-as-a-Judge | nova\_lite\_p5\_48xl\_llm\_judge\_eval.yaml | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-TJ-Eval-latest | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-HP-Eval-latest | ml.p5.48xlarge | 

## Amazon Nova Micro
<a name="nova-model-recipes-reference-novamicro"></a>

下表列出了 Amazon Nova Micro 配方参考的详细信息。


****  

| 模型 | 类别/子类别 | 方法 | 配方名称 | 映像 URI（训练作业） | 映像 URI（SageMaker HyperPod） | 计算实例 | 
| --- | --- | --- | --- | --- | --- | --- | 
| Nova Micro | 训练/微调 | 监督式微调（LoRA） | `nova_micro_1_0_p5_p4d_gpu_lora_sft.yaml`<br />`nova_micro_1_0_g5_g6_12x_gpu_lora_sft.yaml`<br />`nova_micro_1_0_g5_g6_48x_gpu_lora_sft.yaml` | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-TJ-SFT-latest | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-HP-SFT-latest | `ml.p5.48xlarge`, `ml.p5en.48xlarge`, `ml.p4d.24xlarge`,<br />`ml.g5.12xlarge`, `ml.g6.12xlarge`,<br />`ml.g5.48xlarge`, `ml.g6.48xlarge` | 
| Nova Micro | 训练/微调 | 监督式微调（全秩） | `nova_micro_1_0_p5_p4d_gpu_sft.yaml`<br />`nova_micro_1_0_g5_g6_48x_gpu_sft.yaml` | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-TJ-SFT-latest | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-HP-SFT-latest | `ml.p5.48xlarge`, `ml.p5en.48xlarge`, `ml.p4d.24xlarge`,<br />`ml.g5.48xlarge`, `ml.g6.48xlarge` | 
| Nova Micro | 训练/微调 | 直接偏好优化（LoRA） | `nova_micro_1_0_g5_g6_12x_gpu_lora_dpo.yaml`<br />`nova_micro_1_0_g5_g6_48x_gpu_lora_dpo.yaml`<br />`nova_micro_1_0_p5_p4d_gpu_lora_dpo.yaml` | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-TJ-DPO-latest | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-HP-DPO-latest | `ml.p5.48xlarge`, `ml.p5en.48xlarge`, `ml.p4d.24xlarge`,<br />`ml.g5.12xlarge`, `ml.g6.12xlarge`,<br />`ml.g5.48xlarge`, `ml.g6.48xlarge` | 
| Nova Micro | 训练/微调 | 直接偏好优化（全秩） | nova\_micro\_1\_0\_p5\_p4d\_gpu\_dpo.yaml | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-TJ-DPO-latest | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-HP-DPO-latest | ml.p5.48xlarge, ml.p5en.48xlarge, ml.p4d.24xlarge | 
| Nova Micro | 训练/强化学习 | 强化微调（RFT） | nova\_micro\_1\_0\_p5\_gpu\_ppo.yaml | 不适用 | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SMHP-PPO-TRAIN-latest | ml.p5.48xlarge, ml.p5en.48xlarge | 
| Nova Micro | 训练/持续预训练 | 持续预训练（基本模型） | nova\_micro\_gpu\_p5x8\_pretrain.yaml | 不适用 | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:HP-CPT-latest | ml.p5.48xlarge | 
| Nova Micro | 评估/评估 | 一般文本基准测试 | nova\_micro\_p5\_48xl\_general\_text\_benchmark\_eval.yaml | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-TJ-Eval-latest | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-HP-Eval-latest | ml.p5.48xlarge | 
| Nova Micro | 评估/评估 | 自带数据集（gen\_qa）基准测试 | nova\_micro\_p5\_48xl\_bring\_your\_own\_dataset\_eval.yaml | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-TJ-Eval-latest | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-HP-Eval-latest | ml.p5.48xlarge | 
| Nova Micro | 评估/评估 | LLM-as-a-Judge | nova\_micro\_p5\_48xl\_llm\_judge\_eval.yaml | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-TJ-Eval-latest | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-HP-Eval-latest | ml.p5.48xlarge | 

## Amazon Nova Pro
<a name="nova-model-recipes-reference-novapro"></a>

下表列出了 Amazon Nova Pro 配方参考的详细信息。


****  

| 模型 | 类别/子类别 | 方法 | 配方名称 | 映像 URI（训练作业） | 映像 URI（SageMaker HyperPod） | 计算实例 | 
| --- | --- | --- | --- | --- | --- | --- | 
| Nova Pro | 训练/微调 | 监督式微调（LoRA） | nova\_pro\_1\_0\_p5\_p4d\_gpu\_lora\_sft.yaml | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-TJ-SFT-latest | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-HP-SFT-latest | ml.p5.48xlarge, ml.p5en.48xlarge, ml.p4d.24xlarge | 
| Nova Pro | 训练/微调 | 监督式微调（全秩） | nova\_pro\_1\_0\_p5\_p4d\_gpu\_sft.yaml | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-TJ-SFT-latest | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-HP-SFT-latest | ml.p5.48xlarge, ml.p5en.48xlarge | 
| Nova Pro | 训练/微调 | 直接偏好优化（全秩） | nova\_pro\_1\_0\_p5\_gpu\_dpo.yaml | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-TJ-DPO-latest | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-HP-DPO-latest | ml.p5.48xlarge, ml.p5en.48xlarge | 
| Nova Pro | 训练/微调 | 直接偏好优化（LoRA） | nova\_pro\_1\_0\_p5\_p4d\_gpu\_lora\_dpo.yaml | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-TJ-DPO-latest | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SM-HP-DPO-latest | ml.p5.48xlarge, ml.p5en.48xlarge, ml.p4d.24xlarge | 
| Nova Pro | 训练/强化学习 | 强化微调（RFT） | nova\_pro\_1\_0\_p5\_gpu\_ppo.yaml | 不适用 | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:SMHP-PPO-TRAIN-latest | ml.p5.48xlarge, ml.p5en.48xlarge | 
| Nova Pro | 训练/持续预训练 | 持续预训练（基本模型） | nova\_pro\_gpu\_p5x24\_pretrain.yaml | 不适用 | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-fine-tune-repo:HP-CPT-latest | ml.p5.48xlarge | 
| Nova Pro | 训练/数据增强 | 针对后训练的模型蒸馏 | nova\_pro\_r5\_cpu\_distill.yaml | 不适用 | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-distillation-repo:SM-TJ-DISTILL-LATEST | ml.r5.24xlarge | 
| Nova Pro | 评估/评估 | 标准文本基准测试 | nova\_pro\_p5\_48xl\_general\_text\_benchmark\_eval.yaml | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-TJ-Eval-latest | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-HP-Eval-latest | ml.p5.48xlarge | 
| Nova Pro | 评估/评估 | 自定义数据集评估 | nova\_pro\_p5\_48xl\_bring\_your\_own\_dataset\_eval.yaml | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-TJ-Eval-latest | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-HP-Eval-latest | ml.p5.48xlarge | 
| Nova Pro | 评估/评估 | 多模态基准测试 | nova\_pro\_p5\_48xl\_general\_multi\_modal\_benchmark\_eval.yaml | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-TJ-Eval-latest | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-HP-Eval-latest | ml.p5.48xlarge | 
| Nova Pro | 评估/评估 | LLM-as-a-Judge | nova\_pro\_p5\_48xl\_llm\_judge\_eval.yaml | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-TJ-Eval-latest | 708977205387.dkr.ecr.us-east-1.amazonaws.com/nova-evaluation-repo:SM-HP-Eval-latest | ml.p5.48xlarge |