简介

• 原始链接：https://huggingface.co/docs/sagemaker/inference#deploy-with-modeldata
• https://docs.datarobot.com/en/docs/more-info/how-to/aws/sagemaker/sagemaker-deploy.html
  • 这个可以是java环境或者python环境。

部署的都是从huggingface上的model或者根据huaggingface上的model进行fine-tune后的。

一般输入格式如下：

text-classification request body{"inputs": "Camera - You are awarded a SiPix Digital Camera! call 09061221066 fromm landline. Delivery within 28 days."
}
question-answering request body{"inputs": {"question": "What is used for inference?","context": "My Name is Philipp and I live in Nuremberg. This model is used with sagemaker for inference."}
}
zero-shot classification request body{"inputs": "Hi, I recently bought a device from your company but it is not working as advertised and I would like to get reimbursed!","parameters": {"candidate_labels": ["refund","legal","faq"]}
}

所有官方示例

• https://github.com/huggingface/notebooks/tree/main/sagemaker

推理工具

• https://github.com/aws/sagemaker-huggingface-inference-toolkit

使用model.tar.gz

1.从huggingface上下载模型

由于模型文件比较大，需要先安装git-lfs

CentOS7安装Git LFS的方法如下：# 安装必要的软件包：
sudo yum install curl-devel expat-devel gettext-devel openssl-devel zlib-devel
# 安装Git LFS：
curl -s https://packagecloud.io/install/repositories/github/git-lfs/script.rpm.sh | sudo bash
# 安装
sudo yum install git-lfs
# 配置Git LFS：
git lfs install
# 检测是否安装成功：
git lfs version
如果出现版本信息，说明安装成功。

从huaggingface上clone你想使用的模型，以https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2 为例子

git clone https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2

2.自定义代码

允许用户覆盖 HuggingFaceHandlerService 的默认方法。您需要创建一个名为 code/ 的文件夹，其中包含 inference.py 文件。

• HuggingFaceHandlerService

目录结构如下：

model.tar.gz/
|- pytorch_model.bin
|- ....
|- code/|- inference.py|- requirements.txt 

inference.py 文件包含自定义推理模块， requirements.txt 文件包含应添加的其他依赖项。自定义模块可以重写以下方法：

• model_fn(model_dir) 覆盖加载模型的默认方法。返回值 model 将在 predict 中用于预测。 predict 接收参数 model_dir ，即解压后的 model.tar.gz 的路径。
• transform_fn(model, data, content_type, accept_type) 使用您的自定义实现覆盖默认转换函数。您需要在 transform_fn 中实现您自己的 preprocess 、 predict 和 postprocess 步骤。此方法不能与下面提到的 input_fn 、 predict_fn 或 output_fn 组合使用。
• input_fn(input_data, content_type) 覆盖默认的预处理方法。返回值 data 将在 predict 中用于预测。输入是：
  • input_data 是您请求的原始正文。
  • content_type 是请求标头中的内容类型。
• predict_fn(processed_data, model) 覆盖默认的预测方法。返回值 predictions 将在 postprocess 中使用。输入是 processed_data ，即 preprocess 的结果。
• output_fn(prediction, accept) 覆盖后处理的默认方法。返回值 result 将是您请求的响应（例如 JSON ）。输入是：
  • predictions 是 predict 的结果。
  • accept 是 HTTP 请求的返回接受类型，例如 application/json 。

以下是包含 model_fn 、 input_fn 、 predict_fn 和 output_fn 的自定义推理模块的示例：

from sagemaker_huggingface_inference_toolkit import decoder_encoderdef model_fn(model_dir):# implement custom code to load the modelloaded_model = ...return loaded_model def input_fn(input_data, content_type):# decode the input data  (e.g. JSON string -> dict)data = decoder_encoder.decode(input_data, content_type)return datadef predict_fn(data, model):# call your custom model with the dataoutputs = model(data , ... )return predictionsdef output_fn(prediction, accept):# convert the model output to the desired output format (e.g. dict -> JSON string)response = decoder_encoder.encode(prediction, accept)return response

仅使用 model_fn 和 transform_fn 自定义推理模块：

from sagemaker_huggingface_inference_toolkit import decoder_encoderdef model_fn(model_dir):# implement custom code to load the modelloaded_model = ...return loaded_model def transform_fn(model, input_data, content_type, accept):# decode the input data (e.g. JSON string -> dict)data = decoder_encoder.decode(input_data, content_type)# call your custom model with the dataoutputs = model(data , ... ) # convert the model output to the desired output format (e.g. dict -> JSON string)response = decoder_encoder.encode(output, accept)return response

重点，这里的话我们 all-MiniLM-L6-v2的示例代码如下：

from transformers import AutoTokenizer, AutoModel
import torch
import torch.nn.functional as F#Mean Pooling - Take attention mask into account for correct averaging
def mean_pooling(model_output, attention_mask):token_embeddings = model_output[0] #First element of model_output contains all token embeddingsinput_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9)# Sentences we want sentence embeddings for
sentences = ['This is an example sentence', 'Each sentence is converted']# Load model from HuggingFace Hub
tokenizer = AutoTokenizer.from_pretrained('sentence-transformers/all-MiniLM-L6-v2')
model = AutoModel.from_pretrained('sentence-transformers/all-MiniLM-L6-v2')# Tokenize sentences
encoded_input = tokenizer(sentences, padding=True, truncation=True, return_tensors='pt')# Compute token embeddings
with torch.no_grad():model_output = model(**encoded_input)# Perform pooling
sentence_embeddings = mean_pooling(model_output, encoded_input['attention_mask'])# Normalize embeddings
sentence_embeddings = F.normalize(sentence_embeddings, p=2, dim=1)print("Sentence embeddings:")
print(sentence_embeddings)

我们需要改造下，改为我们自己需要的自定义代码：

from transformers import AutoTokenizer, AutoModel
import torch
import torch.nn.functional as F# 这个方法直接同上
def mean_pooling(model_output, attention_mask):token_embeddings = model_output[0] #First element of model_output contains all token embeddingsinput_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9)# 覆盖 -- 模型加载 参考all-MiniLM-L6-v2给出的示例代码
def model_fn(model_dir):# Load model from HuggingFace Hubtokenizer = AutoTokenizer.from_pretrained(model_dir)model = AutoModel.from_pretrained(model_dir)return model, tokenizer
# 覆盖 -- 预测方法 参考all-MiniLM-L6-v2给出的示例代码
def predict_fn(data, model_and_tokenizer):# destruct model and tokenizermodel, tokenizer = model_and_tokenizer# Tokenize sentencessentences = data.pop("inputs", data)encoded_input = tokenizer(sentences, padding=True, truncation=True, return_tensors='pt')# Compute token embeddingswith torch.no_grad():model_output = model(**encoded_input)# Perform poolingsentence_embeddings = mean_pooling(model_output, encoded_input['attention_mask'])# Normalize embeddingssentence_embeddings = F.normalize(sentence_embeddings, p=2, dim=1)# return dictonary, which will be json serializablereturn {"vectors": sentence_embeddings[0].tolist()}

3.打包为tar 文件

cd all-MiniLM-L6-v2
tar zcvf model.tar.gz *

4.上传model.tar.gz到S3

5.部署推理

这里有好几种方式可选。

第一种：在jupyterlab执行这个脚本，替换model等参数即可。

• https://github.com/huggingface/notebooks/blob/main/sagemaker/10_deploy_model_from_s3/deploy_transformer_model_from_s3.ipynb

第二种：这个是吧上面所有步骤都包含了，但是这种无法处理我们在私有环境fine-tune后的模型。

• https://github.com/huggingface/notebooks/blob/main/sagemaker/17_custom_inference_script/sagemaker-notebook.ipynb

第三种：可视化部署，我重点介绍下这个吧

入口如下：

注意下面的选项

• 容器框架根据实际情况选择，这里我们就选择如图
• S3 URI
• IAM role:
  • 可以去IAM创建角色
    • AmazonS3FullAccess
    • AmazonSageMakerFullAccess
  • 也可以去JumpStart中的model去复制过来。

使用hub

原文：https://huggingface.co/docs/sagemaker/inference#deploy-a-model-from-the–hub

这种方式没有上面的方式灵活度高，支持的model也没有上面的方式多。

1.在sagemaker上新建个jupyterlab

2.上传官方示例ipynb文件

• https://github.com/huggingface/notebooks/blob/main/sagemaker/11_deploy_model_from_hf_hub/deploy_transformer_model_from_hf_hub.ipynb

3.指定HF_MODEL_ID和HF_TASK进行部署和推理

inference.py官方示例

• https://sagemaker-examples.readthedocs.io/en/latest/sagemaker-script-mode/pytorch_bert/deploy_bert_outputs.html#Write-the-Inference-Script

import os
import json
from transformers import BertTokenizer, BertModeldef model_fn(model_dir):"""Load the model for inference"""model_path = os.path.join(model_dir, 'model/')# Load BERT tokenizer from disk.tokenizer = BertTokenizer.from_pretrained(model_path)# Load BERT model from disk.model = BertModel.from_pretrained(model_path)model_dict = {'model': model, 'tokenizer':tokenizer}return model_dictdef predict_fn(input_data, model):"""Apply model to the incoming request"""tokenizer = model['tokenizer']bert_model = model['model']encoded_input = tokenizer(input_data, return_tensors='pt')return bert_model(**encoded_input)def input_fn(request_body, request_content_type):"""Deserialize and prepare the prediction input"""if request_content_type == "application/json":request = json.loads(request_body)else:request = request_bodyreturn requestdef output_fn(prediction, response_content_type):"""Serialize and prepare the prediction output"""if response_content_type == "application/json":response = str(prediction)else:response = str(prediction)return response