Amazon SageMaker Serverless Inference – Machine Learning Inference without Worrying about Servers

In December 2021, we introduced Amazon SageMaker Serverless Inference (in preview) as a new option in Amazon SageMaker to deploy machine learning (ML) models for inference without having to configure or manage the underlying infrastructure. Today, I’m happy to announce that Amazon SageMaker Serverless Inference is now generally available (GA).

Different ML inference use cases pose different requirements on your model hosting infrastructure. If you work on use cases such as ad serving, fraud detection, or personalized product recommendations, you are most likely looking for API-based, online inference with response times as low as a few milliseconds. If you work with large ML models, such as in computer vision (CV) applications, you might require infrastructure that is optimized to run inference on larger payload sizes in minutes. If you want to run predictions on an entire dataset, or larger batches of data, you might want to run an on-demand, one-time batch inference job instead of hosting a model-serving endpoint. And what if you have an application with intermittent traffic patterns, such as a chatbot service or an application to process forms or analyze data from documents? In this case, you might want an online inference option that is able to automatically provision and scale compute capacity based on the volume of inference requests. And during idle time, it should be able to turn off compute capacity completely so that you are not charged.

Amazon SageMaker, our fully managed ML service, offers different model inference options to support all of those use cases:

Amazon SageMaker Serverless Inference in More Detail
In a lot of conversations with ML practitioners, I’ve picked up the ask for a fully managed ML inference option that lets you focus on developing the inference code while managing all things infrastructure for you. SageMaker Serverless Inference now delivers this ease of deployment.

Based on the volume of inference requests your model receives, SageMaker Serverless Inference automatically provisions, scales, and turns off compute capacity. As a result, you pay for only the compute time to run your inference code and the amount of data processed, not for idle time.

You can use SageMaker’s built-in algorithms and ML framework-serving containers to deploy your model to a serverless inference endpoint or choose to bring your own container. If traffic becomes predictable and stable, you can easily update from a serverless inference endpoint to a SageMaker real-time endpoint without the need to make changes to your container image. Using Serverless Inference, you also benefit from SageMaker’s features, including built-in metrics such as invocation count, faults, latency, host metrics, and errors in Amazon CloudWatch.

Since its preview launch, SageMaker Serverless Inference has added support for the SageMaker Python SDK and model registry. SageMaker Python SDK is an open-source library for building and deploying ML models on SageMaker. SageMaker model registry lets you catalog, version, and deploy models to production.

New for the GA launch, SageMaker Serverless Inference has increased the maximum concurrent invocations per endpoint limit to 200 (from 50 during preview), allowing you to use Amazon SageMaker Serverless Inference for high-traffic workloads. Amazon SageMaker Serverless Inference is now available in all the AWS Regions where Amazon SageMaker is available, except for the AWS GovCloud (US) and AWS China Regions.

Several customers have already started enjoying the benefits of SageMaker Serverless Inference:

“Bazaarvoice leverages machine learning to moderate user-generated content to enable a seamless shopping experience for our clients in a timely and trustworthy manner. Operating at a global scale over a diverse client base, however, requires a large variety of models, many of which are either infrequently used or need to scale quickly due to significant bursts in content. Amazon SageMaker Serverless Inference provides the best of both worlds: it scales quickly and seamlessly during bursts in content and reduces costs for infrequently used models.” — Lou Kratz, PhD, Principal Research Engineer, Bazaarvoice

“Transformers have changed machine learning, and Hugging Face has been driving their adoption across companies, starting with natural language processing and now with audio and computer vision. The new frontier for machine learning teams across the world is to deploy large and powerful models in a cost-effective manner. We tested Amazon SageMaker Serverless Inference and were able to significantly reduce costs for intermittent traffic workloads while abstracting the infrastructure. We’ve enabled Hugging Face models to work out of the box with SageMaker Serverless Inference, helping customers reduce their machine learning costs even further.” — Jeff Boudier, Director of Product, Hugging Face

Now, let’s see how you can get started on SageMaker Serverless Inference.

For this demo, I’ve built a text classifier to turn e-commerce customer reviews, such as “I love this product!” into positive (1), neutral (0), and negative (-1) sentiments. I’ve used the Women’s E-Commerce Clothing Reviews dataset to fine-tune a RoBERTa model from the Hugging Face Transformers library and model hub. I will now show you how to deploy the trained model to an Amazon SageMaker Serverless Inference Endpoint.

Deploy Model to an Amazon SageMaker Serverless Inference Endpoint
You can create, update, describe, and delete a serverless inference endpoint using the SageMaker console, the AWS SDKs, the SageMaker Python SDK, the AWS CLI, or AWS CloudFormation. In this first example, I will use the SageMaker Python SDK as it simplifies the model deployment workflow through its abstractions. You can also use the SageMaker Python SDK to invoke the endpoint by passing the payload in line with the request. I will show you this in a bit.

First, let’s create the endpoint configuration with the desired serverless configuration. You can specify the memory size and maximum number of concurrent invocations. SageMaker Serverless Inference auto-assigns compute resources proportional to the memory you select. If you choose a larger memory size, your container has access to more vCPUs. As a general rule of thumb, the memory size should be at least as large as your model size. The memory sizes you can choose are 1024 MB, 2048 MB, 3072 MB, 4096 MB, 5120 MB, and 6144 MB. For my RoBERTa model, let’s configure a memory size of 5120 MB and a maximum of five concurrent invocations.

import sagemaker
from sagemaker.serverless import ServerlessInferenceConfig

serverless_config = ServerlessInferenceConfig(
	memory_size_in_mb=5120, 
	max_concurrency=5
)

Now let’s deploy the model. You can use the estimator.deploy() method to deploy the model directly from the SageMaker training estimator, together with the serverless inference endpoint configuration. I also provide my custom inference code in this example.

endpoint_name="roberta-womens-clothing-serverless-1"

estimator.deploy(
	endpoint_name = endpoint_name, 
	entry_point="inference.py",
	serverless_inference_config=serverless_config
)

SageMaker Serverless Inference also supports model registry when you use the AWS SDK for Python (Boto3). I will show you how to deploy the model from the model registry later in this post.

Let’s check the serverless inference endpoint settings and deployment status. Go to the SageMaker console and browse to the deployed inference endpoint:

From the SageMaker console, you can also create, update, or delete serverless inference endpoints if needed. In Amazon SageMaker Studio, select the endpoint tab and your serverless inference endpoint to review the endpoint configuration details.

Once the endpoint status shows InService, you can start sending inference requests.

Now, let’s run a few sample predictions. My fine-tuned RoBERTa model expects the inference requests in JSON Lines format with the review text to classify as the input feature. A JSON Lines text file comprises several lines where each individual line is a valid JSON object, delimited by a newline character. This is an ideal format for storing data that is processed one record at a time, such as in model inference. You can learn more about JSON Lines and other common data formats for inference in the Amazon SageMaker Developer Guide. Note that the following code might look different depending on your model’s accepted inference request format.

from sagemaker.predictor import Predictor
from sagemaker.serializers import JSONLinesSerializer
from sagemaker.deserializers import JSONLinesDeserializer

sess = sagemaker.Session(sagemaker_client=sm)

inputs = [
    {"features": ["I love this product!"]},
    {"features": ["OK, but not great."]},
    {"features": ["This is not the right product."]},
]

predictor = Predictor(
    endpoint_name=endpoint_name,
    serializer=JSONLinesSerializer(),
    deserializer=JSONLinesDeserializer(),
    sagemaker_session=sess
)

predicted_classes = predictor.predict(inputs)

for predicted_class in predicted_classes:
    print("Predicted class {} with probability {}".format(predicted_class['predicted_label'], predicted_class['probability']))

The result will look similar to this, classifying the sample reviews into the corresponding sentiment classes.

Predicted class 1 with probability 0.9495596289634705
Predicted class 0 with probability 0.5395089387893677
Predicted class -1 with probability 0.7887083292007446

You can also deploy your model from the model registry to a SageMaker Serverless Inference endpoint. This is currently only supported through the AWS SDK for Python (Boto3). Let me walk you through another quick demo.

Deploy Model from the SageMaker Model Registry
To deploy the model from the model registry using Boto3, let’s first create a model object from the model version by calling the create_model() method. Then, I pass the Amazon Resource Name (ARN) of the model version as part of the containers for the model object.

import boto3
import sagemaker

sm = boto3.client(service_name="sagemaker")
role = sagemaker.get_execution_role()
model_name="roberta-womens-clothing-serverless"

container_list = 
	[{'ModelPackageName': <MODEL_PACKAGE_ARN>}]

create_model_response = sm.create_model(
    ModelName = model_name,
    ExecutionRoleArn = role,
    Containers = container_list
)

Next, I create the serverless inference endpoint. Remember that you can create, update, describe, and delete a serverless inference endpoint using the SageMaker console, the AWS SDKs, the SageMaker Python SDK, the AWS CLI, or AWS CloudFormation. For consistency, I keep using Boto3 in this second example.

Similar to the first example, I start by creating the endpoint configuration with the desired serverless configuration. I specify the memory size of 5120 MB and a maximum number of five concurrent invocations for my endpoint.

endpoint_config_name="roberta-womens-clothing-serverless-ep-config"

create_endpoint_config_response = sm.create_endpoint_config(
    EndpointConfigName = endpoint_config_name,
    ProductionVariants=[{
        'ServerlessConfig':{
            'MemorySizeInMB' : 5120,
            'MaxConcurrency' : 5
        },
        'ModelName':model_name,
        'VariantName':'AllTraffic'}])

Next, I create the SageMaker Serverless Inference endpoint by calling the create_endpoint() method.

endpoint_name="roberta-womens-clothing-serverless-2"

create_endpoint_response = sm.create_endpoint(
    EndpointName=endpoint_name,
    EndpointConfigName=endpoint_config_name)

Once the endpoint status shows InService, you can start sending inference requests. Again, for consistency, I choose to run the sample prediction using Boto3 and the SageMaker runtime client invoke_endpoint() method.

sm_runtime = boto3.client("sagemaker-runtime")
response = sm_runtime.invoke_endpoint(
    EndpointName=endpoint_name,
    ContentType="application/jsonlines",
    Accept="application/jsonlines",
    Body=bytes('{"features": ["I love this product!"]}', 'utf-8')
)

print(response['Body'].read().decode('utf-8'))
{"probability": 0.966135561466217, "predicted_label": 1}

How to Optimize Your Model for SageMaker Serverless Inference
SageMaker Serverless Inference automatically scales the underlying compute resources to process requests. If the endpoint does not receive traffic for a while, it scales down the compute resources. If the endpoint suddenly receives new requests, you might notice that it takes some time for the endpoint to scale up the compute resources to process the requests.

This cold-start time greatly depends on your model size and the start-up time of your container. To optimize cold-start times, you can try to minimize the size of your model, for example, by applying techniques such as knowledge distillation, quantization, or model pruning.

Knowledge distillation uses a larger model (the teacher model) to train smaller models (student models) to solve the same task. Quantization reduces the precision of the numbers representing your model parameters from 32-bit floating-point numbers down to either 16-bit floating-point or 8-bit integers. Model pruning removes redundant model parameters that contribute little to the training process.

Availability and Pricing
Amazon SageMaker Serverless Inference is now available in all the AWS Regions where Amazon SageMaker is available except for the AWS GovCloud (US) and AWS China Regions.

With SageMaker Serverless Inference, you only pay for the compute capacity used to process inference requests, billed by the millisecond, and the amount of data processed. The compute capacity charge also depends on the memory configuration you choose. For detailed pricing information, visit the SageMaker pricing page.

Get Started Today with Amazon SageMaker Serverless Inference
To learn more about Amazon SageMaker Serverless Inference, visit the Amazon SageMaker machine learning inference webpage. Here are SageMaker Serverless Inference example notebooks that will help you get started right away. Give them a try from the SageMaker console, and let us know what you think.

– Antje