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。

# 支持的框架， AWS 区域、实例类型和经过测试的模型
<a name="training-compiler-support"></a>

**重要**  
Amazon Web Services (AWS) 宣布， SageMaker 训练编译器将没有新版本或新版本。你可以继续通过现有的 Dee AWS p Lear SageMaker ning Containers (DLC) 使用 Training Compiler 进行 SageMaker 训练。值得注意的是，尽管现有的 DLC 仍然可以访问，但根据[AWS 深度学习容器框架支持政策 AWS](https://docs.aws.amazon.com/deep-learning-containers/latest/devguide/support-policy.html)，它们将不再收到来自的补丁或更新。

在使用 T SageMaker raining Compiler 之前，请检查您选择的框架是否受支持 AWS 账户，您的实例类型 AWS 账户 是否在支持的框架中可用 AWS 区域。

**注意**  
SageMaker 训练编译器在 SageMaker Python SDK v2.70.0 或更高版本中可用。

## 支持的框架
<a name="training-compiler-supported-frameworks"></a>

SageMaker Training Compiler 支持以下深度学习框架，可通过 Deep Learning C AWS ontainers 获得。

**Topics**
+ [PyTorch](#training-compiler-supported-frameworks-pytorch)
+ [TensorFlow](#training-compiler-supported-frameworks-tensorflow)

### PyTorch
<a name="training-compiler-supported-frameworks-pytorch"></a>



- **PyTorch**
  - **框架版本:** PyTorch v1.13.1 / **深度学习容器 URI:** 763104351884.dkr.ecr。 {{<region>}}.amazonaws。 com/pytorch-trcomp-training：1.12.0-gpu-py38-cu113-ubuntu20.04-sagemaker / **对 Docker 自定义可扩展:** 否
  - **框架版本:** PyTorch v1.12.0 / **深度学习容器 URI:** 763104351884.dkr.ecr。 {{<region>}}.amazonaws。 com/pytorch-trcomp-training：1.13.1-gpu-py39-cu117-ubuntu20.04-sagemaker / **对 Docker 自定义可扩展:** 否

- **PyTorch 用 Hugging Face 变形金刚**
  - **框架版本:** Transformers v4.21.1<br />PyTorch v1.11.0 / **深度学习容器 URI:** 763104351884.dkr.ecr。 {{<region>}}.amazonaws。 com/huggingface-pytorch-trcomp-training：1.11.0-transformers 4.21.1-gpu-py38-cu113-ubuntu20.04 / **对 Docker 自定义可扩展:** 否
  - **框架版本:** Transformers v4.17.0<br />PyTorch v1.10.2 / **深度学习容器 URI:** 763104351884.dkr.ecr。 {{<region>}}.amazonaws。 com/huggingface-pytorch-trcomp-training：1.10.2-transformers 4.17.0-gpu-py38-cu113-ubuntu20.04 / **对 Docker 自定义可扩展:** 否
  - **框架版本:** Transformers v4.11.0<br />PyTorch v1.9.0 / **深度学习容器 URI:** 763104351884.dkr.ecr。 {{<region>}}.amazonaws。 com/huggingface-pytorch-training-comp: 1.9.0-transformers 4.11.0-gpu-py38-cu111-ubuntu20.04 / **对 Docker 自定义可扩展:** 否



### TensorFlow
<a name="training-compiler-supported-frameworks-tensorflow"></a>



- **TensorFlow**
  - **框架版本:** TensorFlow v2.11.0 / **深度学习容器 URI:** 763104351884.dkr.ecr。 {{<region>}}.amazonaws。 com/tensorflow-training：2.11.0-gpu-py39-cu112-ubuntu20.04-sagemaker / **对 Docker 自定义可扩展:** 是
  - **框架版本:** TensorFlow v2.10.0 / **深度学习容器 URI:** 763104351884.dkr.ecr。 {{<region>}}.amazonaws。 com/tensorflow-training：2.10.0-gpu-py39-cu112-ubuntu20.04-sagemaker / **对 Docker 自定义可扩展:** 是
  - **框架版本:** TensorFlow v2.9.1 / **深度学习容器 URI:** 763104351884.dkr.ecr。 {{<region>}}.amazonaws。 com/tensorflow-training：2.9.1-gpu-py39-cu112-ubuntu20.04-sagemaker / **对 Docker 自定义可扩展:** 是

- **TensorFlow 用 Hugging Face 变形金刚**
  - **框架版本:** Transformers v4.17.0<br />TensorFlow v2.6.3 / **深度学习容器 URI:** 763104351884.dkr.ecr。 {{<region>}}.amazonaws。 com/huggingface-tensorflow-trcomp-training：2.6.3-transformers 4.17.0-gpu-py38-cu112-ubuntu20.04 / **对 Docker 自定义可扩展:** 否
  - **框架版本:** Transformers v4.11.0<br />TensorFlow v2.5.1 / **深度学习容器 URI:** 763104351884.dkr.ecr。 {{<region>}}.amazonaws。 com/huggingface-tensorflow-training-comp: 2.5.1-transformers 4.11.0-gpu-py37-cu112-ubuntu18.04 / **对 Docker 自定义可扩展:** 否



有关更多信息，请参阅 Dee *AWS p Learning Containers GitHub 存储库中的*[可用镜像](https://github.com/aws/deep-learning-containers/blob/master/available_images.md)。

## AWS 区域
<a name="training-compiler-availablity-zone"></a>

[SageMaker 训练编译器容器](https://github.com/aws/deep-learning-containers/blob/master/available_images.md#sagemaker-training-compiler-containers)可在使用 Dee [AWS p Le AWS 区域 arning](https://github.com/aws/deep-learning-containers/blob/master/available_images.md) Containers 的地区使用，但中国地区除外。

## 支持的实例类型
<a name="training-compiler-supported-instance-types"></a>

SageMaker 训练编译器经过测试并支持以下 ML 实例类型。
+ P4 实例
+ P3 实例
+ G4dn 实例
+ G5 实例

有关实例类型的规格，请参阅 [Amazon EC2 实例类型](https://aws.amazon.com/ec2/instance-types/)页面中的**加速计算**部分。有关实例定价的信息，请参阅 [Amazon SageMaker 定价](https://aws.amazon.com/sagemaker/pricing/)。

如果您遇到类似以下内容的错误消息，请按照[请求增加 SageMaker AI 资源的服务配额中的说明进行](https://docs.aws.amazon.com/sagemaker/latest/dg/regions-quotas.html#service-limit-increase-request-procedure)操作。

```
ResourceLimitExceeded: An error occurred (ResourceLimitExceeded) when calling
the CreateTrainingJob operation: The account-level service limit 'ml.p3dn.24xlarge
for training job usage' is 0 Instances, with current utilization of 0 Instances
and a request delta of 1 Instances.
Please contact AWS support to request an increase for this limit.
```

## 经过测试的模型
<a name="training-compiler-tested-models"></a>

下表列出了使用 SageMaker 训练编译器测试过的模型。作为参考，内存中能够容纳的最大批量也包含在其他训练参数旁边。 SageMaker Training Compiler 可以更改模型训练过程的内存占用；因此，在训练过程中通常可以使用更大的批次大小，从而进一步缩短总训练时间。在某些情况下，Tra SageMaker ining Compiler 会智能地促进缓存，从而减少可容纳 GPU 的最大批量大小。您必须重新调整模型超参数并找到最适合您的案例的批处理大小。为了节省时间，请使用以下参考表来查找批处理大小，这将是您的使用案例的良好起点。

**注意**  
批处理大小是适合相应实例类型中的每个 GPU 的本地批处理大小。在更改批处理大小时，您还应调整学习率。

### PyTorch 1.13.1
<a name="training-compiler-tested-models-pt1131"></a>

**自然语言处理 (NLP) 模型**

在单节点和多节点、单或多 GPU 核心以及所示自动混合精度 (AMP) 的所有组合下，针对训练作业测试了以下模型。


<table>
<thead>
  <tr><th colspan="7">Single-node/multi-node单-GPU/multi-GPU</th></tr>
  <tr><th>模型</th><th>数据集</th><th>实例类型</th><th>精度</th><th>序列长度</th><th>原生框架的批处理大小 </th><th> SageMaker 训练编译器的批次大小 </th></tr>
</thead>
<tbody>
  <tr><td>albert-base-v2</td><td>wikitext-2-raw-v1</td><td>g4dn.16xlarge</td><td>float16</td><td>128</td><td>80</td><td>192</td></tr>
  <tr><td>albert-base-v2</td><td>wikitext-2-raw-v1</td><td>g5.4xlarge</td><td>float16</td><td>128</td><td>128</td><td>332</td></tr>
  <tr><td>albert-base-v2</td><td>wikitext-2-raw-v1</td><td>p3.2xlarge</td><td>float16</td><td>128</td><td>80</td><td>224</td></tr>
  <tr><td>bert-base-uncased</td><td>wikitext-2-raw-v1</td><td>g5.4xlarge</td><td>float16</td><td>128</td><td>160</td><td>288</td></tr>
  <tr><td>camembert-base</td><td>wikitext-2-raw-v1</td><td>g5.4xlarge</td><td>float16</td><td>128</td><td>160</td><td>280</td></tr>
  <tr><td>distilbert-base-uncased</td><td>wikitext-2-raw-v1</td><td>g5.4xlarge</td><td>float16</td><td>128</td><td>240</td><td>472</td></tr>
  <tr><td>distilgpt2</td><td>wikitext-2-raw-v1</td><td>g4dn.16xlarge</td><td>float16</td><td>128</td><td>77</td><td>128</td></tr>
  <tr><td>distilgpt2</td><td>wikitext-2-raw-v1</td><td>g5.4xlarge</td><td>float16</td><td>128</td><td>138</td><td>390</td></tr>
  <tr><td>distilgpt2</td><td>wikitext-2-raw-v1</td><td>p3.2xlarge</td><td>float16</td><td>128</td><td>96</td><td>256</td></tr>
  <tr><td>distilroberta-base</td><td>wikitext-2-raw-v1</td><td>g4dn.16xlarge</td><td>float16</td><td>128</td><td>96</td><td>192</td></tr>
  <tr><td>distilroberta-base</td><td>wikitext-2-raw-v1</td><td>g5.4xlarge</td><td>float16</td><td>128</td><td>171</td><td>380</td></tr>
  <tr><td>distilroberta-base</td><td>wikitext-2-raw-v1</td><td>p3.2xlarge</td><td>float16</td><td>128</td><td>112</td><td>256</td></tr>
  <tr><td>gpt2</td><td>wikitext-2-raw-v1</td><td>g4dn.16xlarge</td><td>float16</td><td>128</td><td>52</td><td>152</td></tr>
  <tr><td>gpt2</td><td>wikitext-2-raw-v1</td><td>g5.4xlarge</td><td>float16</td><td>128</td><td>84</td><td>240</td></tr>
  <tr><td>gpt2</td><td>wikitext-2-raw-v1</td><td>p3.2xlarge</td><td>float16</td><td>128</td><td>58</td><td>164</td></tr>
  <tr><td>microsoft/deberta-base</td><td>wikitext-2-raw-v1</td><td>g4dn.16xlarge</td><td>float16</td><td>128</td><td>48</td><td>128</td></tr>
  <tr><td>microsoft/deberta-base</td><td>wikitext-2-raw-v1</td><td>g5.4xlarge</td><td>float16</td><td>128</td><td>84</td><td>207</td></tr>
  <tr><td>microsoft/deberta-base</td><td>wikitext-2-raw-v1</td><td>p3.2xlarge</td><td>float16</td><td>128</td><td>53</td><td>133</td></tr>
  <tr><td>roberta-base</td><td>wikitext-2-raw-v1</td><td>g5.4xlarge</td><td>float16</td><td>128</td><td>125</td><td>224</td></tr>
  <tr><td>xlm-roberta-base</td><td>wikitext-2-raw-v1</td><td>g4dn.16xlarge</td><td>float16</td><td>128</td><td>16</td><td>31</td></tr>
  <tr><td>xlm-roberta-base</td><td>wikitext-2-raw-v1</td><td>p3.2xlarge</td><td>float16</td><td>128</td><td>18</td><td>50</td></tr>
  <tr><td>xlnet-base-cased</td><td>wikitext-2-raw-v1</td><td>g5.4xlarge</td><td>float16</td><td>128</td><td>128</td><td>240</td></tr>
  <tr><td>bert-base-uncased</td><td>wikitext-103-v1</td><td>g5.48xlarge</td><td>float16</td><td>512</td><td>29</td><td>50</td></tr>
  <tr><td>distilbert-base-uncased</td><td>wikitext-103-v1</td><td>g5.48xlarge</td><td>float16</td><td>512</td><td>45</td><td>64</td></tr>
  <tr><td>gpt2</td><td>wikitext-103-v1</td><td>g5.48xlarge</td><td>float16</td><td>512</td><td>18</td><td>45</td></tr>
  <tr><td>roberta-base</td><td>wikitext-103-v1</td><td>g5.48xlarge</td><td>float16</td><td>512</td><td>23</td><td>44</td></tr>
  <tr><td>gpt2</td><td>wikitext-103-v1</td><td>p4d.24xlarge</td><td>float16</td><td>512</td><td>36</td><td>64</td></tr>
</tbody>
</table>


**计算机视觉 (CV) 模型**

如图所示，使用具有自动混合精度 (AMP) 的 M [TensorFlowodel Garden](https://github.com/tensorflow/models) 进行了测试。


<table>
<thead>
  <tr><th colspan="6">Single/multi-node single/multi-GPU</th></tr>
  <tr><th>模型</th><th>数据集</th><th>实例类型</th><th>精度</th><th>原生框架的批处理大小 </th><th> SageMaker 训练编译器的批次大小 </th></tr>
</thead>
<tbody>
  <tr><td>ResNet152</td><td>food101</td><td>g4dn.16xlarge</td><td>float16</td><td>128</td><td>144</td></tr>
  <tr><td>ResNet152</td><td>food101</td><td>g5.4xlarge</td><td>float16</td><td>128</td><td>192</td></tr>
  <tr><td>ResNet152</td><td>food101</td><td>p3.2xlarge</td><td>float16</td><td>152</td><td>156</td></tr>
  <tr><td>ViT</td><td>food101</td><td>g4dn.16xlarge</td><td>float16</td><td>512</td><td>512</td></tr>
  <tr><td>ViT</td><td>food101</td><td>g5.4xlarge</td><td>float16</td><td>992</td><td>768</td></tr>
  <tr><td>ViT</td><td>food101</td><td>p3.2xlarge</td><td>float16</td><td>848</td><td>768</td></tr>
</tbody>
</table>


### PyTorch 1.12.0
<a name="training-compiler-tested-models-pt1120"></a>

**自然语言处理 (NLP) 模型**

在单节点和多节点、单或多 GPU 核心以及所示自动混合精度 (AMP) 的所有组合下，针对训练作业测试了以下模型。


<table>
<thead>
  <tr><th colspan="7">Single-node/multi-node单-GPU/multi-GPU</th></tr>
  <tr><th>模型</th><th>数据集</th><th>实例类型</th><th>精度</th><th>序列长度</th><th>原生框架的批处理大小 </th><th> SageMaker 训练编译器的批次大小 </th></tr>
</thead>
<tbody>
  <tr><td>albert-base-v2</td><td>wikitext-2-raw-v1</td><td>ml.g5.2xlarge</td><td>float16</td><td>128</td><td>128</td><td>248</td></tr>
  <tr><td>bert-base-uncased</td><td>wikitext-2-raw-v1</td><td>ml.g5.2xlarge</td><td>float16</td><td>128</td><td>160</td><td>288</td></tr>
  <tr><td>camembert-base</td><td>wikitext-2-raw-v1</td><td>ml.g5.2xlarge</td><td>float16</td><td>128</td><td>160</td><td>279</td></tr>
  <tr><td>camembert-base</td><td>wikitext-2-raw-v1</td><td>ml.p3.2xlarge</td><td>float16</td><td>128</td><td>105</td><td>164</td></tr>
  <tr><td>distilgpt2</td><td>wikitext-2-raw-v1</td><td>ml.g5.2xlarge</td><td>float16</td><td>128</td><td>136</td><td>256</td></tr>
  <tr><td>distilgpt2</td><td>wikitext-2-raw-v1</td><td>ml.p3.2xlarge</td><td>float16</td><td>128</td><td>80</td><td>118</td></tr>
  <tr><td>gpt2</td><td>wikitext-2-raw-v1</td><td>ml.g5.2xlarge</td><td>float16</td><td>128</td><td>84</td><td>240</td></tr>
  <tr><td>gpt2</td><td>wikitext-2-raw-v1</td><td>ml.p3.2xlarge</td><td>float16</td><td>128</td><td>80</td><td>119</td></tr>
  <tr><td>microsoft/deberta-base</td><td>wikitext-2-raw-v1</td><td>ml.g5.2xlarge</td><td>float16</td><td>128</td><td>93</td><td>197</td></tr>
  <tr><td>microsoft/deberta-base</td><td>wikitext-2-raw-v1</td><td>ml.p3.2xlarge</td><td>float16</td><td>128</td><td>113</td><td>130</td></tr>
  <tr><td>roberta-base</td><td>wikitext-2-raw-v1</td><td>ml.g5.2xlarge</td><td>float16</td><td>128</td><td>125</td><td>224</td></tr>
  <tr><td>roberta-base</td><td>wikitext-2-raw-v1</td><td>ml.p3.2xlarge</td><td>float16</td><td>128</td><td>78</td><td>112</td></tr>
  <tr><td>xlnet-base-cased</td><td>wikitext-2-raw-v1</td><td>ml.g5.2xlarge</td><td>float16</td><td>128</td><td>138</td><td>240</td></tr>
  <tr><td>bert-base-uncased</td><td>wikitext-103-v1</td><td>ml.p4d.24xlarge</td><td>float16</td><td>512</td><td></td><td>52</td></tr>
  <tr><td>distilbert-base-uncased</td><td>wikitext-103-v1</td><td>ml.p4d.24xlarge</td><td>float16</td><td>512</td><td></td><td>160</td></tr>
  <tr><td>gpt2</td><td>wikitext-103-v1</td><td>ml.p4d.24xlarge</td><td>float16</td><td>512</td><td></td><td>25</td></tr>
  <tr><td>roberta-base</td><td>wikitext-103-v1</td><td>ml.p4d.24xlarge</td><td>float16</td><td>512</td><td></td><td>64</td></tr>
</tbody>
</table>


### TensorFlow 2.11.0
<a name="training-compiler-tested-models-tf2110"></a>

**计算机视觉 (CV) 模型**

如图所示，使用具有自动混合精度 (AMP) 的 M [TensorFlowodel Garden](https://github.com/tensorflow/models) 进行了测试。


<table>
<thead>
  <tr><th colspan="6">Single/multi-node single/multi-GPU</th></tr>
  <tr><th>模型</th><th>数据集</th><th>实例类型</th><th>精度</th><th>原生框架的批处理大小 </th><th> SageMaker 训练编译器的批次大小 </th></tr>
</thead>
<tbody>
  <tr><td>MaskRCNN-ResNet50-FPN</td><td>COCO-2017</td><td>ml.g5.2xlarge</td><td>float16</td><td>6</td><td>8</td></tr>
  <tr><td>MaskRCNN-ResNet50-FPN</td><td>COCO-2017</td><td>ml.p3.2xlarge</td><td>float16</td><td>4</td><td>6</td></tr>
  <tr><td>ResNet50</td><td>ImageNet</td><td>ml.g5.2xlarge</td><td>float16</td><td>192</td><td>256</td></tr>
  <tr><td>ResNet50</td><td>ImageNet</td><td>ml.p3.2xlarge</td><td>float16</td><td>256</td><td>256</td></tr>
  <tr><td>ResNet101</td><td>ImageNet</td><td>ml.g5.2xlarge</td><td>float16</td><td>128</td><td>256</td></tr>
  <tr><td>ResNet101</td><td>ImageNet</td><td>ml.p3.2xlarge</td><td>float16</td><td>128</td><td>128</td></tr>
  <tr><td>ResNet152</td><td>ImageNet</td><td>ml.g5.2xlarge</td><td>float16</td><td>128</td><td>224</td></tr>
  <tr><td>ResNet152</td><td>ImageNet</td><td>ml.p3.2xlarge</td><td>float16</td><td>128</td><td>128</td></tr>
  <tr><td>VisionTransformer</td><td>ImageNet</td><td>ml.g5.2xlarge</td><td>float16</td><td>112</td><td>144</td></tr>
  <tr><td>VisionTransformer</td><td>ImageNet</td><td>ml.p3.2xlarge</td><td>float16</td><td>96</td><td>128</td></tr>
</tbody>
</table>


**自然语言处理 (NLP) 模型**

已结合使用带 `Sequence_Len=128` 的[转换器模型](https://github.com/huggingface/transformers)和自动混合精度 (AMP) 进行测试，如下所示。


<table>
<thead>
  <tr><th colspan="6">Single/multi-node single/multi-GPU</th></tr>
  <tr><th>模型</th><th>数据集</th><th>实例类型</th><th>精度</th><th>原生框架的批处理大小 </th><th> SageMaker 训练编译器的批次大小 </th></tr>
</thead>
<tbody>
  <tr><td>albert-base-v2</td><td>wikitext-2-raw-v1</td><td>ml.g5.2xlarge</td><td>float16</td><td>160</td><td>197</td></tr>
  <tr><td>albert-base-v2</td><td>wikitext-2-raw-v1</td><td>ml.p3.2xlarge</td><td>float16</td><td>95</td><td>127</td></tr>
  <tr><td>bert-base-uncased</td><td>wikitext-2-raw-v1</td><td>ml.g5.2xlarge</td><td>float16</td><td>160</td><td>128</td></tr>
  <tr><td>bert-base-uncased</td><td>wikitext-2-raw-v1</td><td>ml.p3.2xlarge</td><td>float16</td><td>104</td><td>111</td></tr>
  <tr><td>bert-large-uncased</td><td>wikitext-2-raw-v1</td><td>ml.g5.2xlarge</td><td>float16</td><td>65</td><td>48</td></tr>
  <tr><td>bert-large-uncased</td><td>wikitext-2-raw-v1</td><td>ml.p3.2xlarge</td><td>float16</td><td>40</td><td>35</td></tr>
  <tr><td>camembert-base</td><td>wikitext-2-raw-v1</td><td>ml.g5.2xlarge</td><td>float16</td><td>128</td><td>162</td></tr>
  <tr><td>camembert-base</td><td>wikitext-2-raw-v1</td><td>ml.p3.2xlarge</td><td>float16</td><td>105</td><td>111</td></tr>
  <tr><td>distilbert-base-uncased</td><td>wikitext-2-raw-v1</td><td>ml.g5.2xlarge</td><td>float16</td><td>256</td><td>264</td></tr>
  <tr><td>distilbert-base-uncased</td><td>wikitext-2-raw-v1</td><td>ml.p3.2xlarge</td><td>float16</td><td>128</td><td>169</td></tr>
  <tr><td>gpt2</td><td>wikitext-2-raw-v1</td><td>ml.g5.2xlarge</td><td>float16</td><td>128</td><td>120</td></tr>
  <tr><td>gpt2</td><td>wikitext-2-raw-v1</td><td>ml.p3.2xlarge</td><td>float16</td><td>80</td><td>83</td></tr>
  <tr><td>jplu/tf-xlm-roberta-base</td><td>wikitext-2-raw-v1</td><td>ml.g5.2xlarge</td><td>float16</td><td>32</td><td>32</td></tr>
  <tr><td>jplu/tf-xlm-roberta-base</td><td>wikitext-2-raw-v1</td><td>ml.p3.2xlarge</td><td>float16</td><td>32</td><td>36</td></tr>
  <tr><td>microsoft/mpnet-base</td><td>wikitext-2-raw-v1</td><td>ml.g5.2xlarge</td><td>float16</td><td>144</td><td>160</td></tr>
  <tr><td>microsoft/mpnet-base</td><td>wikitext-2-raw-v1</td><td>ml.p3.2xlarge</td><td>float16</td><td>106</td><td>110</td></tr>
  <tr><td>roberta-base</td><td>wikitext-2-raw-v1</td><td>ml.g5.2xlarge</td><td>float16</td><td>128</td><td>128</td></tr>
  <tr><td>roberta-base</td><td>wikitext-2-raw-v1</td><td>ml.p3.2xlarge</td><td>float16</td><td>72</td><td>98</td></tr>
  <tr><td>albert-base-v2</td><td>wikitext-2-raw-v1</td><td>ml.g5.48xlarge</td><td>float16</td><td>128</td><td>192</td></tr>
  <tr><td>albert-base-v2</td><td>wikitext-2-raw-v1</td><td>ml.p3.16xlarge</td><td>float16</td><td>95</td><td>96</td></tr>
  <tr><td>distilbert-base-uncased</td><td>wikitext-2-raw-v1</td><td>ml.g5.48xlarge</td><td>float16</td><td>256</td><td>256</td></tr>
  <tr><td>distilbert-base-uncased</td><td>wikitext-2-raw-v1</td><td>ml.p3.16xlarge</td><td>float16</td><td>140</td><td>184</td></tr>
  <tr><td>google/electra-小鉴别器</td><td>wikitext-2-raw-v1</td><td>ml.g5.48xlarge</td><td>float16</td><td>256</td><td>384</td></tr>
  <tr><td>google/electra-小鉴别器</td><td>wikitext-2-raw-v1</td><td>ml.p3.16xlarge</td><td>float16</td><td>256</td><td>268</td></tr>
  <tr><td>gpt2</td><td>wikitext-2-raw-v1</td><td>ml.g5.48xlarge</td><td>float16</td><td>116</td><td>116</td></tr>
  <tr><td>gpt2</td><td>wikitext-2-raw-v1</td><td>ml.p3.16xlarge</td><td>float16</td><td>85</td><td>83</td></tr>
  <tr><td>gpt2</td><td>wikitext-2-raw-v1</td><td>ml.p4d.24xlarge</td><td>float16</td><td>94</td><td>110</td></tr>
  <tr><td>microsoft/mpnet-base</td><td>wikitext-2-raw-v1</td><td>ml.g5.48xlarge</td><td>float16</td><td>187</td><td>164</td></tr>
  <tr><td>microsoft/mpnet-base</td><td>wikitext-2-raw-v1</td><td>ml.p3.16xlarge</td><td>float16</td><td>106</td><td>111</td></tr>
</tbody>
</table>


### TensorFlow 2.10.0
<a name="training-compiler-tested-models-tf2100"></a>

**计算机视觉 (CV) 模型**

如图所示，使用具有自动混合精度 (AMP) 的 M [TensorFlowodel Garden](https://github.com/tensorflow/models) 进行了测试。


<table>
<thead>
  <tr><th colspan="6">Single-node单-GPU/multi-GPU</th></tr>
  <tr><th>模型</th><th>数据集</th><th>实例类型</th><th>精度</th><th>原生框架的批处理大小 </th><th> SageMaker 训练编译器的批次大小 </th></tr>
</thead>
<tbody>
  <tr><td>DetectionTransformer-ResNet50</td><td>COCO-2017</td><td>ml.g4dn.2xlarge</td><td>float32</td><td>2</td><td>4</td></tr>
  <tr><td>DetectionTransformer-ResNet50</td><td>COCO-2017</td><td>ml.g5.2xlarge</td><td>float32</td><td>3</td><td>6</td></tr>
  <tr><td>DetectionTransformer-ResNet50</td><td>COCO-2017</td><td>ml.p3.2xlarge</td><td>float32</td><td>2</td><td>4</td></tr>
  <tr><td>MaskRCNN-ResNet50-FPN</td><td>COCO-2017</td><td>ml.g4dn.2xlarge</td><td>float16</td><td>4</td><td>6</td></tr>
  <tr><td>MaskRCNN-ResNet50-FPN</td><td>COCO-2017</td><td>ml.g5.2xlarge</td><td>float16</td><td>6</td><td>8</td></tr>
  <tr><td>MaskRCNN-ResNet50-FPN</td><td>COCO-2017</td><td>ml.g5.48xlarge</td><td>float16</td><td>48</td><td>64</td></tr>
  <tr><td>MaskRCNN-ResNet50-FPN</td><td>COCO-2017</td><td>ml.p3.2xlarge</td><td>float16</td><td>4</td><td>6</td></tr>
  <tr><td>ResNet50</td><td>ImageNet</td><td>ml.g4dn.2xlarge</td><td>float16</td><td>224</td><td>256</td></tr>
  <tr><td>ResNet50</td><td>ImageNet</td><td>ml.g5.2xlarge</td><td>float16</td><td>192</td><td>160</td></tr>
  <tr><td>ResNet50</td><td>ImageNet</td><td>ml.g5.48xlarge</td><td>float16</td><td>2048</td><td>2048</td></tr>
  <tr><td>ResNet50</td><td>ImageNet</td><td>ml.p3.2xlarge</td><td>float16</td><td>224</td><td>160</td></tr>
  <tr><td>ResNet101</td><td>ImageNet</td><td>ml.g4dn.2xlarge</td><td>float16</td><td>160</td><td>128</td></tr>
  <tr><td>ResNet101</td><td>ImageNet</td><td>ml.g5.2xlarge</td><td>float16</td><td>192</td><td>256</td></tr>
  <tr><td>ResNet101</td><td>ImageNet</td><td>ml.g5.48xlarge</td><td>float16</td><td>2048</td><td>2048</td></tr>
  <tr><td>ResNet101</td><td>ImageNet</td><td>ml.p3.2xlarge</td><td>float16</td><td>160</td><td>224</td></tr>
  <tr><td>ResNet152</td><td>ImageNet</td><td>ml.g4dn.2xlarge</td><td>float16</td><td>128</td><td>128</td></tr>
  <tr><td>ResNet152</td><td>ImageNet</td><td>ml.g5.2xlarge</td><td>float16</td><td>192</td><td>224</td></tr>
  <tr><td>ResNet152</td><td>ImageNet</td><td>ml.g5.48xlarge</td><td>float16</td><td>1536</td><td>1792</td></tr>
  <tr><td>ResNet152</td><td>ImageNet</td><td>ml.p3.2xlarge</td><td>float16</td><td>128</td><td>160</td></tr>
  <tr><td>VisionTransformer</td><td>ImageNet</td><td>ml.g4dn.2xlarge</td><td>float16</td><td>80</td><td>128</td></tr>
  <tr><td>VisionTransformer</td><td>ImageNet</td><td>ml.g5.2xlarge</td><td>float16</td><td>112</td><td>144</td></tr>
  <tr><td>VisionTransformer</td><td>ImageNet</td><td>ml.g5.48xlarge</td><td>float16</td><td>896</td><td>1152</td></tr>
  <tr><td>VisionTransformer</td><td>ImageNet</td><td>ml.p3.2xlarge</td><td>float16</td><td>80</td><td>128</td></tr>
</tbody>
</table>


**自然语言处理 (NLP) 模型**

已结合使用带 `Sequence_Len=128` 的[转换器模型](https://github.com/huggingface/transformers)和自动混合精度 (AMP) 进行测试，如下所示。


<table>
<thead>
  <tr><th colspan="6">Single-node单-GPU/multi-GPU</th></tr>
  <tr><th>模型</th><th>数据集</th><th>实例类型</th><th>精度</th><th>原生框架的批处理大小 </th><th> SageMaker 训练编译器的批次大小 </th></tr>
</thead>
<tbody>
  <tr><td>albert-base-v2</td><td>wikitext-2-raw-v1</td><td>g4dn.16xlarge</td><td>float16</td><td>128</td><td>112</td></tr>
  <tr><td>albert-base-v2</td><td>wikitext-2-raw-v1</td><td>p3.2xlarge</td><td>float16</td><td>128</td><td>128</td></tr>
  <tr><td>albert-base-v2</td><td>wikitext-2-raw-v1</td><td>p3.8xlarge</td><td>float16</td><td>128</td><td>135</td></tr>
  <tr><td>albert-base-v2</td><td>wikitext-2-raw-v1</td><td>g5.4xlarge</td><td>float16</td><td>128</td><td>191</td></tr>
  <tr><td>bert-base-uncased</td><td>wikitext-2-raw-v1</td><td>g4dn.16xlarge</td><td>float16</td><td>64</td><td>94</td></tr>
  <tr><td>bert-base-uncased</td><td>wikitext-2-raw-v1</td><td>p3.2xlarge</td><td>float16</td><td>96</td><td>101</td></tr>
  <tr><td>bert-base-uncased</td><td>wikitext-2-raw-v1</td><td>p3.8xlarge</td><td>float16</td><td>96</td><td>96</td></tr>
  <tr><td>bert-base-uncased</td><td>wikitext-2-raw-v1</td><td>g5.4xlarge</td><td>float16</td><td>128</td><td>128</td></tr>
  <tr><td>bert-large-uncased</td><td>wikitext-2-raw-v1</td><td>g4dn.16xlarge</td><td>float16</td><td>35</td><td>21</td></tr>
  <tr><td>bert-large-uncased</td><td>wikitext-2-raw-v1</td><td>p3.2xlarge</td><td>float16</td><td>39</td><td>26</td></tr>
  <tr><td>bert-large-uncased</td><td>wikitext-2-raw-v1</td><td>g5.4xlarge</td><td>float16</td><td>60</td><td>50</td></tr>
  <tr><td>camembert-base</td><td>wikitext-2-raw-v1</td><td>g4dn.16xlarge</td><td>float16</td><td>96</td><td>90</td></tr>
  <tr><td>camembert-base</td><td>wikitext-2-raw-v1</td><td>p3.2xlarge</td><td>float16</td><td>96</td><td>98</td></tr>
  <tr><td>camembert-base</td><td>wikitext-2-raw-v1</td><td>p3.8xlarge</td><td>float16</td><td>96</td><td>96</td></tr>
  <tr><td>camembert-base</td><td>wikitext-2-raw-v1</td><td>g5.4xlarge</td><td>float16</td><td>128</td><td>128</td></tr>
  <tr><td>distilbert-base-uncased</td><td>wikitext-2-raw-v1</td><td>g4dn.16xlarge</td><td>float16</td><td>256</td><td>160</td></tr>
  <tr><td>distilbert-base-uncased</td><td>wikitext-2-raw-v1</td><td>p3.2xlarge</td><td>float16</td><td>128</td><td>176</td></tr>
  <tr><td>distilbert-base-uncased</td><td>wikitext-2-raw-v1</td><td>p3.8xlarge</td><td>float16</td><td>128</td><td>160</td></tr>
  <tr><td>distilbert-base-uncased</td><td>wikitext-2-raw-v1</td><td>g5.4xlarge</td><td>float16</td><td>256</td><td>258</td></tr>
  <tr><td>google\_electra-small-discriminator</td><td>wikitext-2-raw-v1</td><td>g4dn.16xlarge</td><td>float16</td><td>256</td><td>216</td></tr>
  <tr><td>google\_electra-small-discriminator</td><td>wikitext-2-raw-v1</td><td>p3.2xlarge</td><td>float16</td><td>256</td><td>230</td></tr>
  <tr><td>google\_electra-small-discriminator</td><td>wikitext-2-raw-v1</td><td>p3.8xlarge</td><td>float16</td><td>256</td><td>224</td></tr>
  <tr><td>google\_electra-small-discriminator</td><td>wikitext-2-raw-v1</td><td>g5.4xlarge</td><td>float16</td><td>256</td><td>320</td></tr>
  <tr><td>gpt2</td><td>wikitext-2-raw-v1</td><td>g4dn.16xlarge</td><td>float16</td><td>80</td><td>64</td></tr>
  <tr><td>gpt2</td><td>wikitext-2-raw-v1</td><td>p3.2xlarge</td><td>float16</td><td>80</td><td>77</td></tr>
  <tr><td>gpt2</td><td>wikitext-2-raw-v1</td><td>p3.8xlarge</td><td>float16</td><td>80</td><td>72</td></tr>
  <tr><td>gpt2</td><td>wikitext-2-raw-v1</td><td>g5.4xlarge</td><td>float16</td><td>128</td><td>120</td></tr>
  <tr><td>jplu\_tf-xlm-roberta-base</td><td>wikitext-2-raw-v1</td><td>g4dn.16xlarge</td><td>float16</td><td>28</td><td>24</td></tr>
  <tr><td>jplu\_tf-xlm-roberta-base</td><td>wikitext-2-raw-v1</td><td>p3.2xlarge</td><td>float16</td><td>32</td><td>24</td></tr>
  <tr><td>jplu\_tf-xlm-roberta-base</td><td>wikitext-2-raw-v1</td><td>p3.8xlarge</td><td>float16</td><td>32</td><td>26</td></tr>
  <tr><td>jplu\_tf-xlm-roberta-base</td><td>wikitext-2-raw-v1</td><td>g5.4xlarge</td><td>float16</td><td>66</td><td>52</td></tr>
  <tr><td>microsoft\_mpnet-base</td><td>wikitext-2-raw-v1</td><td>g4dn.16xlarge</td><td>float16</td><td>96</td><td>92</td></tr>
  <tr><td>microsoft\_mpnet-base</td><td>wikitext-2-raw-v1</td><td>p3.2xlarge</td><td>float16</td><td>96</td><td>101</td></tr>
  <tr><td>microsoft\_mpnet-base</td><td>wikitext-2-raw-v1</td><td>p3.8xlarge</td><td>float16</td><td>96</td><td>101</td></tr>
  <tr><td>microsoft\_mpnet-base</td><td>wikitext-2-raw-v1</td><td>g5.4xlarge</td><td>float16</td><td>128</td><td>152</td></tr>
  <tr><td>roberta-base</td><td>wikitext-2-raw-v1</td><td>g4dn.16xlarge</td><td>float16</td><td>64</td><td>72</td></tr>
  <tr><td>roberta-base</td><td>wikitext-2-raw-v1</td><td>p3.2xlarge</td><td>float16</td><td>64</td><td>84</td></tr>
  <tr><td>roberta-base</td><td>wikitext-2-raw-v1</td><td>p3.8xlarge</td><td>float16</td><td>64</td><td>86</td></tr>
  <tr><td>roberta-base</td><td>wikitext-2-raw-v1</td><td>g5.4xlarge</td><td>float16</td><td>128</td><td>128</td></tr>
</tbody>
</table>


### TensorFlow 2.9.1
<a name="training-compiler-tested-models-tf291"></a>

使用具有自动混合精度 (AMP) 的 [TensorFlowModel Garden](https://github.com/tensorflow/models) 进行了测试。


<table>
<thead>
  <tr><th colspan="5">Single-node单-GPU/multi-GPU</th></tr>
  <tr><th>模型</th><th>数据集</th><th>实例类型</th><th>原生框架的批处理大小 </th><th> SageMaker 训练编译器的批次大小 </th></tr>
</thead>
<tbody>
  <tr><td>ResNet50</td><td>ImageNet</td><td>ml.g4dn.2xlarge</td><td>192</td><td>256\*</td></tr>
  <tr><td rowspan="3">ResNet101</td><td rowspan="3">ImageNet</td><td>ml.g4dn.2xlarge</td><td>128</td><td>160</td></tr>
  <tr><td>ml.g5.2xlarge</td><td>224</td><td>256\*</td></tr>
  <tr><td>ml.p3.16xlarge</td><td>1536</td><td>1792</td></tr>
  <tr><td rowspan="3">ResNet152</td><td rowspan="3">ImageNet</td><td>ml.g5.2xlarge</td><td>192</td><td>224</td></tr>
  <tr><td>ml.p3.2xlarge</td><td>160</td><td>160</td></tr>
  <tr><td>ml.p3.16xlarge</td><td>1024</td><td>1 280</td></tr>
  <tr><td rowspan="4">VisionTransformer</td><td rowspan="4">ImageNet</td><td>ml.g4dn.2xlarge</td><td>80</td><td>128\*</td></tr>
  <tr><td>ml.g5.2xlarge</td><td>112</td><td>128\*</td></tr>
  <tr><td>ml.p3.2xlarge</td><td>56</td><td>128\*</td></tr>
  <tr><td>ml.p3.16xlarge</td><td>640</td><td>1024\*</td></tr>
  <tr><td rowspan="4">DetectionTransformer-ResNet50</td><td rowspan="4">COCO-2017</td><td>ml.g4dn.2xlarge</td><td>2</td><td>2</td></tr>
  <tr><td>ml.g5.2xlarge</td><td>3</td><td>6</td></tr>
  <tr><td>ml.p3.2xlarge</td><td>2</td><td>4</td></tr>
  <tr><td>ml.p3.16xlarge</td><td>8</td><td>32</td></tr>
  <tr><td rowspan="3">MaskRCNN-ResNet50-FPN</td><td rowspan="3">COCO-2017</td><td>ml.g4dn.2xlarge</td><td>4</td><td>4</td></tr>
  <tr><td>ml.g5.2xlarge</td><td>6</td><td>8</td></tr>
  <tr><td>ml.p3.2xlarge</td><td>4</td><td>6</td></tr>
</tbody>
</table>


\* 标有星号 (\*) 的批量大小表示 SageMaker 训练编译器开发团队测试的最大批量。对于已标记的单元格，该实例可能能够容纳比所示批处理大小更大的批处理大小。

### 变形金刚 4.21.1 和 1.11.0 PyTorch
<a name="training-compiler-tested-models-hf421-pt111"></a>

已通过 `Sequence_Len=512` 和自动混合精度 (AMP) 进行测试。


<table>
<thead>
  <tr><th colspan="6">Single-node 单 GPU</th></tr>
  <tr><th>模型 </th><th>数据集</th><th>实例类型</th><th>实例计数</th><th>原生框架的批处理大小</th><th>Training Compiler 的批处理大小</th></tr>
</thead>
<tbody>
  <tr><td rowspan="3">albert-base-v2</td><td rowspan="3">wikitext-2</td><td>ml.g4dn.2xlarge</td><td>1</td><td>14</td><td>28</td></tr>
  <tr><td>ml.g5.2xlarge</td><td>1</td><td>18</td><td>40</td></tr>
  <tr><td>ml.p3.2xlarge</td><td>1</td><td>14</td><td>32</td></tr>
  <tr><td rowspan="3">bert-base-cased</td><td rowspan="3">wikitext-2</td><td>ml.g4dn.2xlarge</td><td>1</td><td>12</td><td>24</td></tr>
  <tr><td>ml.g5.2xlarge</td><td>1</td><td>28</td><td>44</td></tr>
  <tr><td>ml.p3.2xlarge</td><td>1</td><td>16</td><td>20</td></tr>
  <tr><td rowspan="3">camembert-base</td><td rowspan="3">wikitext-2</td><td>ml.g4dn.2xlarge</td><td>1</td><td>16</td><td>28</td></tr>
  <tr><td>ml.g5.2xlarge</td><td>1</td><td>24</td><td>40</td></tr>
  <tr><td>ml.p3.2xlarge</td><td>1</td><td>16</td><td>24</td></tr>
  <tr><td rowspan="4">distilbert-base-uncased</td><td rowspan="3">wikitext-2</td><td>ml.g4dn.2xlarge</td><td>1</td><td>28</td><td>52</td></tr>
  <tr><td>ml.g5.2xlarge</td><td>1</td><td>40</td><td>76</td></tr>
  <tr><td>ml.p3.2xlarge</td><td>1</td><td>32</td><td>48</td></tr>
  <tr><td>wikitext-103-v1</td><td>ml.p4d.24xlarge</td><td>4</td><td>82</td><td>160</td></tr>
  <tr><td rowspan="3">distilgpt2</td><td rowspan="3">wikitext-2</td><td>ml.g4dn.2xlarge</td><td>1</td><td>6</td><td>18</td></tr>
  <tr><td>ml.g5.2xlarge</td><td>1</td><td>12</td><td>28</td></tr>
  <tr><td>ml.p3.2xlarge</td><td>1</td><td>6</td><td>16</td></tr>
  <tr><td rowspan="3">distilroberta-base</td><td rowspan="3">wikitext-2</td><td>ml.g4dn.2xlarge</td><td>1</td><td>20</td><td>40</td></tr>
  <tr><td>ml.g5.2xlarge</td><td>1</td><td>28</td><td>56</td></tr>
  <tr><td>ml.p3.2xlarge</td><td>1</td><td>24</td><td>40</td></tr>
  <tr><td rowspan="3">EleutherAI/gpt-neo-125M</td><td rowspan="3">wikitext-2</td><td>ml.g4dn.2xlarge</td><td>1</td><td>4</td><td>8</td></tr>
  <tr><td>ml.g5.2xlarge</td><td>1</td><td>6</td><td>14</td></tr>
  <tr><td>ml.p3.2xlarge</td><td>1</td><td>4</td><td>10</td></tr>
  <tr><td rowspan="4">gpt2</td><td rowspan="3">wikitext-2</td><td>ml.g4dn.2xlarge</td><td>1</td><td>4</td><td>8</td></tr>
  <tr><td>ml.g5.2xlarge</td><td>1</td><td>6</td><td>16</td></tr>
  <tr><td>ml.p3.2xlarge</td><td>1</td><td>4</td><td>10</td></tr>
  <tr><td>wikitext-103-v1</td><td>ml.p4d.24xlarge</td><td>4</td><td>13</td><td>25</td></tr>
  <tr><td rowspan="4">roberta-base</td><td rowspan="3">wikitext-2</td><td>ml.g4dn.2xlarge</td><td>1</td><td>12</td><td>20</td></tr>
  <tr><td>ml.g5.2xlarge</td><td>1</td><td>24</td><td>36</td></tr>
  <tr><td>ml.p3.2xlarge</td><td>1</td><td>12</td><td>20</td></tr>
  <tr><td>wikitext-103-v1</td><td>ml.p4d.24xlarge</td><td>4</td><td>36</td><td>64</td></tr>
  <tr><td rowspan="3">xlnet-base-cased</td><td rowspan="3">wikitext-2</td><td>ml.g4dn.2xlarge</td><td>1</td><td>2</td><td>6</td></tr>
  <tr><td>ml.g5.2xlarge</td><td>1</td><td>2</td><td>10</td></tr>
  <tr><td>ml.p3.2xlarge</td><td>1</td><td>2</td><td>8</td></tr>
  <tr><td rowspan="4">bert-base-uncased</td><td rowspan="4">wikitext-103-v1</td><td rowspan="4">ml.p4d.24xlarge</td><td>2</td><td>32</td><td>64</td></tr>
  <tr><td>4</td><td>32</td><td>64</td></tr>
  <tr><td>8</td><td>32</td><td>64</td></tr>
  <tr><td>16</td><td>32</td><td>64</td></tr>
  <tr><td>roberta-large</td><td>wikitext-103-v1</td><td>ml.p4d.24xlarge</td><td>4</td><td>16</td><td>24</td></tr>
  <tr><td>microsoft/deberta-v3 base</td><td>wikitext-103-v1</td><td>ml.p4d.24xlarge</td><td>16</td><td>9</td><td>23</td></tr>
</tbody>
</table>


### 变形金刚 4.17.0 和 1.10.2 PyTorch
<a name="training-compiler-tested-models-hf417-pt110"></a>

已通过 `Sequence_Len=512` 和自动混合精度 (AMP) 进行测试。


<table>
<thead>
  <tr><th colspan="4">Single-node 单 GPU</th></tr>
  <tr><th>模型 </th><th>实例类型</th><th>原生框架的批处理大小</th><th>Training Compiler 的批处理大小</th></tr>
</thead>
<tbody>
  <tr><td rowspan="2">albert-base-v2</td><td>ml.p3.2xlarge</td><td>14</td><td>28</td></tr>
  <tr><td>ml.g4dn.2xlarge</td><td>14</td><td>24</td></tr>
  <tr><td rowspan="2">bert-base-cased</td><td>ml.p3.2xlarge</td><td>16</td><td>24</td></tr>
  <tr><td>ml.g4dn.2xlarge</td><td>12</td><td>24</td></tr>
  <tr><td rowspan="2">bert-base-uncased</td><td>ml.p3.2xlarge</td><td>16</td><td>24</td></tr>
  <tr><td>ml.g4dn.2xlarge</td><td>12</td><td>28</td></tr>
  <tr><td rowspan="2">camembert-base</td><td>ml.p3.2xlarge</td><td>12</td><td>24</td></tr>
  <tr><td>ml.g4dn.2xlarge</td><td>12</td><td>28</td></tr>
  <tr><td rowspan="2">distilbert-base-uncased</td><td>ml.p3.2xlarge</td><td>28</td><td>48</td></tr>
  <tr><td>ml.g4dn.2xlarge</td><td>24</td><td>52</td></tr>
  <tr><td rowspan="2">distilgpt2</td><td>ml.p3.2xlarge</td><td>6</td><td>12</td></tr>
  <tr><td>ml.g4dn.2xlarge</td><td>6</td><td>14</td></tr>
  <tr><td rowspan="2">distilroberta-base</td><td>ml.p3.2xlarge</td><td>20</td><td>40</td></tr>
  <tr><td>ml.g4dn.2xlarge</td><td>12</td><td>40</td></tr>
  <tr><td rowspan="2">EleutherAI/gpt-neo-125M</td><td>ml.p3.2xlarge</td><td>2</td><td>10</td></tr>
  <tr><td>ml.g4dn.2xlarge</td><td>2</td><td>8</td></tr>
  <tr><td rowspan="2">facebook/bart-base</td><td>ml.p3.2xlarge</td><td>2</td><td>6</td></tr>
  <tr><td>ml.g4dn.2xlarge</td><td>2</td><td>6</td></tr>
  <tr><td rowspan="2">gpt2</td><td>ml.p3.2xlarge</td><td>4</td><td>8</td></tr>
  <tr><td>ml.g4dn.2xlarge</td><td>2</td><td>8</td></tr>
  <tr><td rowspan="2">roberta-base</td><td>ml.p3.2xlarge</td><td>12</td><td>20</td></tr>
  <tr><td>ml.g4dn.2xlarge</td><td>12</td><td>20</td></tr>
  <tr><td rowspan="2">xlnet-base-cased</td><td>ml.p3.2xlarge</td><td>2</td><td>8</td></tr>
  <tr><td>ml.g4dn.2xlarge</td><td>4</td><td>6</td></tr>
</tbody>
</table>


### 变形金刚 4.11.0 和 1.9.0 PyTorch
<a name="training-compiler-tested-models-hf411-pt190"></a>

已通过 `Sequence_Len=512` 和自动混合精度 (AMP) 进行测试。


<table>
<thead>
  <tr><th colspan="4">Single-node 单 GPU</th></tr>
  <tr><th>模型 </th><th>实例类型</th><th>本机的批处理大小</th><th>Training Compiler 的批处理大小</th></tr>
</thead>
<tbody>
  <tr><td>albert-base-v2 </td><td>ml.p3.2xlarge</td><td>12</td><td>32</td></tr>
  <tr><td>bert-base-cased </td><td>ml.p3.2xlarge</td><td>14</td><td>24</td></tr>
  <tr><td>bert-base-chinese</td><td>ml.p3.2xlarge</td><td>16</td><td>24</td></tr>
  <tr><td>bert-base-multilingual-cased </td><td>ml.p3.2xlarge</td><td>4</td><td>16</td></tr>
  <tr><td>bert-base-multilingual-uncased </td><td>ml.p3.2xlarge</td><td>8</td><td>16</td></tr>
  <tr><td>bert-base-uncased </td><td>ml.p3.2xlarge</td><td>12</td><td>24</td></tr>
  <tr><td>cl-base-japanes tohoku/bert e-word-masking</td><td>ml.p3.2xlarge</td><td>12</td><td>24</td></tr>
  <tr><td>cl--bas tohoku/bert e-japanese </td><td>ml.p3.2xlarge</td><td>12</td><td>24</td></tr>
  <tr><td>distilbert-base-uncased </td><td>ml.p3.2xlarge</td><td>28</td><td>32</td></tr>
  <tr><td>distilbert-base-uncased-finetuned-sst-2-english</td><td>ml.p3.2xlarge</td><td>28</td><td>32</td></tr>
  <tr><td>distilgpt2 </td><td>ml.p3.2xlarge</td><td>16</td><td>32</td></tr>
  <tr><td>facebook/bart-base </td><td>ml.p3.2xlarge</td><td>4</td><td>8</td></tr>
  <tr><td>gpt2</td><td>ml.p3.2xlarge</td><td>6</td><td>20</td></tr>
  <tr><td>nreimers/MiniLMv2-L6-H384-distilled-from-RoBERTa-Large </td><td>ml.p3.2xlarge</td><td>20</td><td>32</td></tr>
  <tr><td>roberta-base </td><td>ml.p3.2xlarge</td><td>12</td><td>20</td></tr>
</tbody>
</table>



<table>
<thead>
  <tr><th colspan="4">Single-node 多 GPU</th></tr>
  <tr><th>模型 </th><th>实例类型</th><th>本机的批处理大小</th><th>Training Compiler 的批处理大小</th></tr>
</thead>
<tbody>
  <tr><td>bert-base-chinese </td><td>ml.p3.8xlarge</td><td>16</td><td>26</td></tr>
  <tr><td>bert-base-multilingual-cased </td><td>ml.p3.8xlarge</td><td>6</td><td>16</td></tr>
  <tr><td>bert-base-multilingual-uncased</td><td>ml.p3.8xlarge</td><td>6</td><td>16</td></tr>
  <tr><td>bert-base-uncased </td><td>ml.p3.8xlarge</td><td>14</td><td>24</td></tr>
  <tr><td>distilbert-base-uncased </td><td>ml.p3.8xlarge</td><td>14</td><td>32</td></tr>
  <tr><td>distilgpt2</td><td>ml.p3.8xlarge</td><td>6</td><td>32</td></tr>
  <tr><td>facebook/bart-base</td><td>ml.p3.8xlarge</td><td>8</td><td>16</td></tr>
  <tr><td>gpt2 </td><td>ml.p3.8xlarge</td><td>8</td><td>20</td></tr>
  <tr><td>roberta-base </td><td>ml.p3.8xlarge</td><td>12</td><td>20</td></tr>
</tbody>
</table>


### 变形金刚 4.17.0 与 2.6.3 TensorFlow
<a name="training-compiler-tested-models-hf417-tf263"></a>

已通过 `Sequence_Len=128` 和自动混合精度 (AMP) 进行测试。


| 模型  | 实例类型 | 原生框架的批处理大小 | Training Compiler 的批处理大小 | 
| --- | --- | --- | --- | 
| albert-base-v2 | ml.g4dn.16xlarge | 136 | 208 | 
| albert-base-v2 | ml.g5.4xlarge | 219 | 312 | 
| albert-base-v2 | ml.p3.2xlarge | 152 | 208 | 
| albert-base-v2 | ml.p3.8xlarge | 152 | 192 | 
| bert-base-uncased | ml.g4dn.16xlarge | 120 | 101 | 
| bert-base-uncased | ml.g5.4xlarge | 184 | 160 | 
| bert-base-uncased | ml.p3.2xlarge | 128 | 108 | 
| bert-large-uncased | ml.g4dn.16xlarge | 37 | 28 | 
| bert-large-uncased | ml.g5.4xlarge | 64 | 55 | 
| bert-large-uncased | ml.p3.2xlarge | 40 | 32 | 
| camembert-base | ml.g4dn.16xlarge | 96 | 100 | 
| camembert-base | ml.g5.4xlarge | 190 | 160 | 
| camembert-base | ml.p3.2xlarge | 129 | 108 | 
| camembert-base | ml.p3.8xlarge | 128 | 104 | 
| distilbert-base-uncased | ml.g4dn.16xlarge | 210 | 160 | 
| distilbert-base-uncased | ml.g5.4xlarge | 327 | 288 | 
| distilbert-base-uncased | ml.p3.2xlarge | 224 | 196 | 
| distilbert-base-uncased | ml.p3.8xlarge | 192 | 182 | 
| google\_electra-small-discriminator | ml.g4dn.16xlarge | 336 | 288 | 
| google\_electra-small-discriminator | ml.g5.4xlarge | 504 | 384 | 
| google\_electra-small-discriminator | ml.p3.2xlarge | 352 | 323 | 
| gpt2 | ml.g4dn.16xlarge | 89 | 64 | 
| gpt2 | ml.g5.4xlarge | 140 | 146 | 
| gpt2 | ml.p3.2xlarge | 94 | 96 | 
| gpt2 | ml.p3.8xlarge | 96 | 88 | 
| jplu\_tf-xlm-roberta-base | ml.g4dn.16xlarge | 52 | 16 | 
| jplu\_tf-xlm-roberta-base | ml.g5.4xlarge | 64 | 44 | 
| microsoft\_mpnet-base | ml.g4dn.16xlarge | 120 | 100 | 
| microsoft\_mpnet-base | ml.g5.4xlarge | 192 | 160 | 
| microsoft\_mpnet-base | ml.p3.2xlarge | 128 | 104 | 
| microsoft\_mpnet-base | ml.p3.8xlarge | 130 | 92 | 
| roberta-base | ml.g4dn.16xlarge | 108 | 64 | 
| roberta-base | ml.g5.4xlarge | 176 | 142 | 
| roberta-base | ml.p3.2xlarge | 118 | 100 | 
| roberta-base | ml.p3.8xlarge | 112 | 88 | 

### 变形金刚 4.11.0 和 2.5.1 TensorFlow
<a name="training-compiler-tested-models-hf411-tf251"></a>

已通过 `Sequence_Len=128` 和自动混合精度 (AMP) 进行测试。


<table>
<thead>
  <tr><th colspan="4">Single-node 单 GPU</th></tr>
  <tr><th>模型 </th><th>实例类型</th><th>本机的批处理大小</th><th>Training Compiler 的批处理大小</th></tr>
</thead>
<tbody>
  <tr><td>albert-base-v2 </td><td>ml.p3.2xlarge</td><td>128</td><td>128</td></tr>
  <tr><td>bart-base </td><td>ml.p3.2xlarge</td><td>12</td><td>64</td></tr>
  <tr><td>bart-large </td><td>ml.p3.2xlarge</td><td>4</td><td>28</td></tr>
  <tr><td>bert-base-cased </td><td>ml.p3.2xlarge</td><td>16</td><td>128</td></tr>
  <tr><td>bert-base-chinese</td><td>ml.p3.2xlarge</td><td>16</td><td>128</td></tr>
  <tr><td>bert-base-multilingual-cased </td><td>ml.p3.2xlarge</td><td>12</td><td>64</td></tr>
  <tr><td>bert-base-multilingual-uncased </td><td>ml.p3.2xlarge</td><td>16</td><td>96</td></tr>
  <tr><td>bert-base-uncased</td><td>ml.p3.2xlarge</td><td>16</td><td>96</td></tr>
  <tr><td>bert-large-uncased </td><td>ml.p3.2xlarge</td><td>4</td><td>24</td></tr>
  <tr><td>cl--bas tohoku/bert e-japanese </td><td>ml.p3.2xlarge</td><td>16</td><td>128</td></tr>
  <tr><td>cl-base-japanes tohoku/bert e-word-masking </td><td>ml.p3.2xlarge</td><td>16</td><td>128</td></tr>
  <tr><td>distilbert-base-sst2 </td><td>ml.p3.2xlarge</td><td>32</td><td>128</td></tr>
  <tr><td>distilbert-base-uncased </td><td>ml.p3.2xlarge</td><td>32</td><td>128</td></tr>
  <tr><td>distilgpt2</td><td>ml.p3.2xlarge</td><td>32</td><td>128</td></tr>
  <tr><td>gpt2 </td><td>ml.p3.2xlarge</td><td>12</td><td>64</td></tr>
  <tr><td>gpt2-large </td><td>ml.p3.2xlarge</td><td>2</td><td>24</td></tr>
  <tr><td>jplu/tf-xlm-roberta-base </td><td>ml.p3.2xlarge</td><td>12</td><td>32</td></tr>
  <tr><td>roberta-base </td><td>ml.p3.2xlarge</td><td>4</td><td>64</td></tr>
  <tr><td>roberta-large </td><td>ml.p3.2xlarge</td><td>4</td><td>64</td></tr>
  <tr><td>t5-base </td><td>ml.p3.2xlarge</td><td>64</td><td>64</td></tr>
  <tr><td>t5-small </td><td>ml.p3.2xlarge</td><td>128</td><td>128</td></tr>
</tbody>
</table>