Skip to content

Latest commit

 

History

History
359 lines (302 loc) · 15.3 KB

README.md

File metadata and controls

359 lines (302 loc) · 15.3 KB
Raw

Serverless Huggingface Demos with AWS Lambda

Note: This code is meant as a demo. You should secure access to demos once deployed.

This project builds off of the Zero administration inference with AWS Lambda. Our solution leverages the techniques used to package and deploy open source natural language processing models downloaded from Hugging Face and extends the approach to provide quick prototypes to share with stakeholders. At a high level, we'd like to be able to select a new open source model from Hugging Face and have a prototype that we could share with stakeholders as quickly as possible with minimal code development or changes.

Overview

Our solution consists of an AWS Cloud Development Kit (AWS CDK) script that automatically provisions container image-based Lambda functions that perform ML inference using pre-trained Hugging Face models. This solution also includes Amazon Elastic File System (EFS) storage that is attached to the Lambda functions to cache the pre-trained models and reduce inference latency.

Architecture diagram

In this architectural diagram:

  1. Serverless inference is achieved by using Lambda functions that are based on container images.
  2. The container image is stored in an Amazon Elastic Container Registry (ECR) repository within your account.
  3. Pre-trained models are automatically downloaded from Hugging Face the first time the function is invoked.
  4. Pre-trained models are cached within Amazon Elastic File System storage to improve inference latency.
  5. Another Lambda function is implemented to provide a web-based user interface and access each model.

Prerequisites

The following software is required to run and deploy this project:

Provided that the required software is in place, the following commands should allow you to deploy a new instance of this demo project.

Deploying the example application

  1. Clone the project to your development environment:
git clone <https://github.com/aws-samples/zero-administration-inference-with-aws-lambda-for-hugging-face.git>
  1. Set up and activate a virtual environment. This is optional but recommended.
python -m venv .venv

Activating your virtual environment will be different based on if you are using a MacOS, Linux, or Windows environment.

MacOS and Linux

source .venv/bin/activate

Windows

.venv\Scripts\activate
  1. Install the required dependencies:
pip install -r requirements.txt
  1. Bootstrap the CDK. This command provisions the initial resources needed by the CDK to perform deployments:
cdk bootstrap
  1. This command deploys the CDK application to its environment. During the deployment, the toolkit outputs progress indications:
cdk deploy

After running these steps, the Cloud Development Kit (CDK) will ask you to confirm that you would like to deploy all of the resources required for this project to the default region configured for the CDK. Once you've confirmed that you'd like to deploy the solution, the CDK will create and configure all the services necessary to support this solution.

Architecture diagram

Once the deployment has completed, the output of the stack deployment will be a public URL that you can share with stakeholders which will allow them to experiement with open source nlp models running serverlessly in a specific AWS account and region.

Why host open source models?

Before we dive into the details of the solution, let's consider why we might want to work with open source models hosted in our own AWS account rather than use models hosted by 3rd parties or models hosted serverlessly by AWS Bedrock. Here are a few reasons.

  1. We don't have to share our data with 3rd parties when we host our own models.
  2. Hugging Face hosts hundreds of thousands of models, and new models are uploaded every day. Deploying the latest models provides a mechanism to leverage the latest advancements.
  3. We have the flexibility to deploy fine-tuned models which are more accurate, faster, or cheaper to run for our proprietary data or unique use cases.

Architecture

Let's take a closer look at the resources the CDK provides for this application. Our recipe for this deployment can be found in the app.py file. At the top of the file, we can see the imports for the services the CDK project will use to deploy and configure this application.

from aws_cdk import (
    aws_lambda as lambda_,
    aws_efs as efs,
    aws_ec2 as ec2,
    aws_apigateway as apigateway
)
  • We will be using AWS Lambda to serverlessly host our models and the user interface for our app. Lambda can scale up quickly to support hundreds of concurrent requests and then scale back down to zero when there is no traffic.
  • This project uses the Elastic File System (EFS) to cache the files for any models that we download from Hugging Face. If you aren't familiar with EFS, it works as a network drive that we share with multiple lambda functions and other compute resources.
  • Next we will be using an instance on the Elastic Compute Cloud (EC2) to build the Docker images that will host our models. The Hugging Face transformer libray detects the runtime environment and downloads the appropriate files for the environment it will be running in.
  • And finally, the project will use the Amazon API Gateway to configure a public RESTful HTTP API to allow GET and POST requests to our application.

Unlike in the original Zero administration inference with AWS Lambda project, this project needs to keep track of the Amazon Resource Names (ARNs) for each Docker Lambda. ARNs uniquely identify AWS resources and the router Lambda function will use the ARNs to call each Docker lambda function that gets created. The CDK project saves the ARNs for each Docker lambda in a dictionary, which will later be passed to the router lambda as environment variables.

passed_environment_variables["functionARN"+str(count)] = docker_lambda.function_arn

Once we have collected all the ARNs for the Docker lambdas, we can create our router lambda function and pass all the ARNs in as environment variables.

        # Lambda Function to route requests to the Docker Lambda Functions
        self.router_lambda = lambda_.Function(
            scope=self,
            id="LambdaRouter",
            runtime=lambda_.Runtime.PYTHON_3_12,
            function_name="LambdaRouter",
            code=lambda_.Code.from_asset(
                path="lambda"
            ),
            handler="router.handler",
            # The environment variables for the Lambda function are passed here.
            environment=passed_environment_variables,
        )

The location of the router lambda code will be saved in the router.py file in the lambda folder. We will revisit this code in more detail once we have reviewed the remainder of the app.py CDK code.

Once we have defined our router lambda function, we can create a public API Gateway endpoint to handle HTTP GET and POST requests to our application.

        # API Gateway to provide a public endpoint for the router Lambda Function
        api = apigateway.LambdaRestApi(
            self,
            "LambdaRouterApi",
            handler = self.router_lambda,
            proxy = False,
        )
        api_resource = api.root.add_resource("go")
        api_resource.add_method("GET")
        api_resource.add_method("POST")

And finally, we need to grant the router lambda function permission to access each of the Docker lambda functions that we've created. We do this by calling the grant_invoke method on each docker_lambda instance and pass it the router_lambda instance we wan to provide invoke permission.

        #Grant invoke permissions to the router lambda
        for docker_lambda in docker_lambdas:
            docker_lambda.grant_invoke(self.router_lambda)

Infrastucture as Code

The app.py file defines all of the resources needed to run our application. Defining all of our infrastructure resources as code is called Infrastructure as Code (IaC). Using code and software development techniques, rather than manually configuring systems, enables developers and IT operations teams to provision resources automatically. This provides some key benefits such as:

Key Features of Infrastructure as Code:

  1. Automation: IaC automates the provisioning of infrastructure, reducing the human error associated with manual setups. It allows for the quick setup of environments from development to production through scripts.
  2. Consistency: By defining your infrastructure in source code, IaC helps maintain consistency across environments, reducing discrepancies and incompatibilities between development, testing, and production environments.
  3. Version Control: Infrastructure configurations can be versioned and tracked using the same version control systems used for software development (such as Git). This enables changes to be tracked over time, rollbacks to be made if necessary, and an understanding of who made changes and why.
  4. Reusability: Code used to define infrastructure can be reused for different environments or solutions, saving time and reducing the risk of errors when configuring new infrastructure.
  5. Cost-Effective: With IaC, organizations can quickly spin up and tear down environments, which optimizes cloud resource usage and keeps costs under control.

Hugging Face Models

Hugging Face Transformers is a popular open source project that provides pre-trained, natural language processing (NLP) models for a wide variety of use cases. Customers with minimal machine learning experience can use pre-trained models to enhance their applications quickly using NLP. This includes tasks such as text classification, language translation, summarization, and question answering - to name a few.

This application includes example Python scripts for two common NLP use cases:

  • Sentiment analysis: Identifying if a sentence indicates positive or negative sentiment. It uses a fine-tuned model on sst2, which is a GLUE task.
  • Summarization: Summarizing a body of text into a shorter, representative text. It uses a Bart model that was fine-tuned on the CNN / Daily Mail dataset. For simplicity, both of these use cases are implemented using Hugging Face pipelines.

Understanding the code structure

The code is organized using the following structure:

├── inference
│   ├── Dockerfile
│   ├── sentiment.py
│   └── summarization.py
├── app.py
└── ...

The inference directory contains:

  • The Dockerfile is used to build a custom image to be able to run PyTorch Hugging Face inference using Lambda functions.
  • The Python scripts that perform the actual ML inference

The sentiment.py script shows how to use a Hugging Face Transformers model:

import json
from transformers import pipeline

nlp = pipeline("sentiment-analysis")

def handler(event, context):
    response = {
        "statusCode": 200,
        "body": nlp(event['text'])[0]
    }
    return response

For each Python script in the inference directory, the CDK generates a Lambda function backed by a container image and a Python inference script.

Adding a translator

You could add more models by adding additional Python scripts in the inference directory. For example, you could add a new model to translate English text to French by adding the following code in a file called translate-en2fr.py to the inference directory:

import json
from transformers
import pipeline

en_fr_translator = pipeline('translation_en_to_fr')

def handler(event, context):
    response = {
        "statusCode": 200,
        "body": en_fr_translator(event['text'])[0]
    }
    return response

Then you would cdk deploy again to create a new Docker lambda function to perform English to French translation:

$ cdk deploy

The router lambda

The code for the router lambda can be found in the router.py file in the lambda directory. When the application is accessed via an HTTP GET, such as when the endpoint is accessed in a web browser, the router lambda returns a simple HTML form.

http_method = event.get('httpMethod')
    
if http_method == 'GET':
    return {"statusCode": 200,
        "headers": {"content-type": "text/html"},
        "body": html}

Architecture diagram

Our simple HTML form is then designed to POST the contents of a text box back to our router lambda whenever the submit button is pressed. When the router lamba detects POST request, it checks to see if a specific Docker lambda has been requested. If none is found in the request, it defaults to using the ARN for the first Docker lambda passed via the environment variables for the router lambda.

elif http_method == 'POST':
        body = json.loads(event['body'])
        text = body['text']
        functionARN = os.getenv("functionARN1")
        if "functionARN" in body:
            functionARN = body['functionARN']
    
        result = get_response(text, functionARN)

In the get_response method, the boto3 lambda client is used to call the Docker lambda identified by the ARN that is passed in.

def get_response(text, functionARN):
        lambda_client = boto3.client('lambda')
        try:
            # Parameters to invoke the other Lambda function
            response = lambda_client.invoke(
                FunctionName=functionARN,  
                InvocationType='RequestResponse', 
                Payload=json.dumps({'text': text}) 
            )

            # Read and parse the response payload
            response_payload = response['Payload'].read().decode('utf-8')
            response_data = json.loads(response_payload)
            return {
                    'message': 'Invocation successful',
                    'responseFromLambda': response_data
                }

Architecture diagram

Cleaning up

After you are finished experimenting with this project, run cdk destroy to remove all of the associated infrastructure.

License

This library is licensed under the MIT No Attribution License. See the LICENSE file. Disclaimer: Deploying the demo applications contained in this repository will potentially cause your AWS Account to be billed for services.

Links