Model Hosting FAQs - Amazon SageMaker

Model Hosting FAQs

Refer to the following FAQ items for answers to commonly asked questions about SageMaker Inference Hosting.

General Hosting

The following FAQ items answer common general questions for SageMaker Inference.

A: After you build and train models, Amazon SageMaker provides four options to deploy them so you can start making predictions. Real-Time Inference is suitable for workloads with millisecond latency requirements, payload sizes up to 6 MB, and processing times of up to 60 seconds. Batch Transform is ideal for offline predictions on large batches of data that are available up front. Asynchronous Inference is designed for workloads that do not have sub-second latency requirements, payload sizes up to 1 GB, and processing times of up to 15 minutes. With Serverless Inference, you can quickly deploy machine learning models for inference without having to configure or manage the underlying infrastructure, and you pay only for the compute capacity used to process inference requests, which is ideal for intermittent workloads.

A: The following diagram can help you choose a SageMaker Hosting model deployment option.

Flowchart explaining how to choose a model deployment option in SageMaker, described in the following paragraph.

The preceding diagram walks you through the following decision process. If you want to process requests in batches, you might want to choose Batch Transform. Otherwise, if you want to receive inference for each request to your model, you might want to choose Asynchronous Inference, Serverless Inference, or Real-Time Inference. You can choose Asynchronous Inference if you have long processing times or large payloads and want to queue requests. You can choose Serverless Inference if your workload has unpredictable or intermittent traffic. You can choose Real-Time Inference if you have sustained traffic and need lower and consistent latency for your requests.

A: To optimize your costs with SageMaker Inference, you should choose the right hosting option for your use case. You can also use Inference features such as Amazon SageMaker Savings Plans, model optimization with SageMaker Neo, Multi-Model Endpoints and Multi-Container Endpoints, or autoscaling. For tips on how to optimize your Inference costs, see Inference cost optimization best practices.

A: You should use Amazon SageMaker Inference Recommender if you need recommendations for the right endpoint configuration to improve performance and reduce costs. Previously, data scientists who wanted to deploy their models had to run manual benchmarks to select the right endpoint configuration. First, they had to select the right machine learning instance type out of more than 70 available instance types based on the resource requirements of their models and sample payloads, and then optimize the model to account for differing hardware. Then, they had to conduct extensive load tests to validate that latency and throughput requirements were met and that the costs were low. Inference Recommender eliminates this complexity by helping you do the following:

  • Get started in minutes with an instance recommendation.

  • Conduct load tests across instance types to get recommendations on your endpoint configuration within hours.

  • Automatically tune container and model server parameters as well as perform model optimizations for a given instance type.

A: SageMaker endpoints are HTTP REST endpoints that use a containerized web server, which includes a model server. These containers are responsible for loading up and serving requests for a machine learning model. They implement a web server that responds to /invocations and /ping on port 8080.

Common model servers include TensorFlow Serving, TorchServe and Multi Model Server. SageMaker framework containers have these model servers built in.

A: Everything in SageMaker Inference is containerized. SageMaker provides managed containers for popular frameworks such as TensorFlow, SKlearn, and HuggingFace. For a comprehensive updated list of those images, see Available Images.

Sometimes there are custom frameworks for which you might need to build a container. This approach is known as Bring Your Own Container or BYOC. With the BYOC approach, you provide the Docker image to set up your framework or library. Then, you push the image to Amazon Elastic Container Registry (Amazon ECR) so that you can use the image with SageMaker. For an example of a BYOC approach, see Overivew of Containers for Amazon SageMaker.

Alternatively, instead of building an image from scratch, you can extend a container. You can take one of the base images that SageMaker provides and add your dependencies on top of it in your Dockerfile.

A: SageMaker offers the capacity to bring your own trained framework model that you've trained outside of SageMaker and deploy it on any of the SageMaker hosting options.

SageMaker requires you to package the model in a model.tar.gz file and have a specific directory structure. Each framework has its own model structure (see the following question for example structures). For more information, see the SageMaker Python SDK documentation for TensorFlow, PyTorch, and MXNet.

While you can choose from prebuilt framework images such as TensorFlow, PyTorch, and MXNet to host your trained model, you can also build your own container to host your trained models on SageMaker endpoints. For a walkthrough, see the example Jupyter notebook Building your own algorithm container.

A: SageMaker requires your model artifacts to be compressed in a .tar.gz file, or a tarball. SageMaker automatically extracts this .tar.gz file into the /opt/ml/model/ directory in your container. The tarball shouldn't contain any symlinks or unncessary files. If you are making use of one of the framework containers, such as TensorFlow, PyTorch, or MXNet, the container expects your TAR structure to be as follows:

TensorFlow

model.tar.gz/ |--[model_version_number]/ |--variables |--saved_model.pb code/ |--inference.py |--requirements.txt

PyTorch

model.tar.gz/ |- model.pth |- code/ |- inference.py |- requirements.txt # only for versions 1.3.1 and higher

MXNet

model.tar.gz/ |- model-symbol.json |- model-shapes.json |- model-0000.params |- code/ |- inference.py |- requirements.txt # only for versions 1.6.0 and higher

A: ContentType is the MIME type of the input data in the request body (the MIME type of the data you are sending to your endpoint). The model server uses the ContentType to determine if it can handle the type provided or not.

Accept is the MIME type of the inference response (the MIME type of the data your endpoint returns). The model server uses the Accept type to determine if it can handle returning the type provided or not.

Common MIME types include text/csv, application/json, and application/jsonlines.

A: SageMaker passes any request onto the model container without modification. The container must contain the logic to deserialize the request. For information about the formats defined for built-in algorithms, see Common Data Formats for Inference. If you are building your own container or using a SageMaker Framework container, you can include the logic to accept a request format of your choice.

Similarly, SageMaker also returns the response without modification, and then the client must deserialize the response. In case of the built-in algorithms, they return responses in specific formats. If you are building your own container or using a SageMaker Framework container, you can include the logic to return a response in the format you choose.

Use the Invoke Endpoint API call to make inference against your endpoint.

When passing your input as a payload to the InvokeEndpoint API, you must provide the correct type of input data that your model expects. When passing a payload in the InvokeEndpoint API call, the request bytes are forwarded directly to the model container. For example, for an image, you may use application/jpeg for the ContentType, and make sure that your model can perform inference on this type of data. This applies for JSON, CSV, video, or any other type of input with which you may be dealing.

Another factor to consider is payload size limits. In terms of real-time and serverless endpoints, the payload limit is 6 MB. You can split your video into multiple frames and invoke the endpoint with each frame individually. Alternatively, if your use case permits, you can send the whole video in the payload using an asynchronous endpoint, which supports up to 1 GB payloads.

For an example that showcases how to run computer vision inference on large videos with Asynchronous Inference, see this blog post.

Real-Time Inference

The following FAQ items answer common questions for SageMaker Real-Time Inference.

A: You can create a SageMaker endpoint through AWS-supported tooling such as the AWS SDKs, the SageMaker Python SDK, the AWS Management Console, AWS CloudFormation, and the AWS Cloud Development Kit (AWS CDK).

There are three key entities in endpoint creation: a SageMaker model, a SageMaker endpoint configuration, and a SageMaker endpoint. The SageMaker model points towards the model data and image you are using. The endpoint configuration defines your production variants, which might include the instance type and instance count. You can then use either the create_endpoint API call or the .deploy() call for SageMaker to create an endpoint using the metadata from your model and endpoint configuration.

A: No, you can use the various AWS SDKs (see Invoke/Create for available SDKs) or even call the corresponding web APIs directly.

A: A Multi-Model Endpoint is a Real-Time Inference option that SageMaker provides. With Multi-Model Endpoints, you can host thousands of models behind one endpoint. Multi Model Server is an open-source framework for serving machine learning models. It provides the HTTP front-end and model management capabilities required by multi-model endpoints to host multiple models within a single container, load models into and unload models out of the container dynamically, and perform inference on a specified loaded model.

A: SageMaker Real-Time Inference supports various model deployment architecture such as Multi-Model Endpoints, Multi-Container Endpoints, and Serial Inference Pipelines.

Multi-Model Endpoints (MME) – MME allows customers to deploy 1000s of hyper‐personalized models in a cost effective way. All the models are deployed on a shared‐resource fleet. MME works best when the models are of similar size and latency and belong to the same ML framework. These endpoints are ideal for when you have don’t need to call the same model at all times. You can dynamically load respective models onto the SageMaker endpoint to serve your request.

Multi-Container Endpoints (MCE) – MCE allows customers to deploy 15 different containers with diverse ML frameworks and functionalities with no cold starts while only using one SageMaker endpoint. You can directly invoke these containers. MCE is best for when you want to keep all the models in memory.

Serial Inference Pipelines (SIP) – You can use SIP to chain together 2‐15 containers on a single endpoint. SIP is mostly suitable for combining preprocessing and model inference in one endpoint and for low latency operations.

Serverless Inference

The following FAQ items answer common questions for Amazon SageMaker Serverless Inference.

A: Serverless Inference is a purpose-built serverless model serving option that makes it easy to deploy and scale ML models. Serverless Inference endpoints automatically start compute resources and scale them in and out depending on traffic, eliminating the need for you to choose instance type, run provisioned capacity, or manage scaling. You can optionally specify the memory requirements for your serverless endpoint. You pay only for the duration of running the inference code and the amount of data processed, not for idle periods.

A: Serverless Inference simplifies the developer experience by eliminating the need to provision capacity up front and manage scaling policies. Serverless Inference can scale instantly from tens to thousands of inferences within seconds based on the usage patterns, making it ideal for ML applications with intermittent or unpredictable traffic. For example, a chatbot service used by a payroll processing company experiences an increase in inquiries at the end of the month while traffic is intermittent for rest of the month. Provisioning instances for the entire month in such scenarios is not cost-effective, as you end up paying for idle periods.

Serverless Inference helps address these types of use cases by providing you automatic and fast scaling out of the box without the need for you to forecast traffic up front or manage scaling policies. Additionally, you pay only for the compute time to run your inference code and for data processing, making it ideal for workloads with intermittent traffic.

A: Your serverless endpoint has a minimum RAM size of 1024 MB (1 GB), and the maximum RAM size you can choose is 6144 MB (6 GB). The memory sizes you can choose are 1024 MB, 2048 MB, 3072 MB, 4096 MB, 5120 MB, or 6144 MB. 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.

Choose your endpoint’s memory size according to your model size. Generally, the memory size should be at least as large as your model size. You may need to benchmark in order to choose the right memory selection for your model based on your latency SLAs. The memory size increments have different pricing; see the Amazon SageMaker pricing page for more information.

Batch Transform

The following FAQ items answer common questions for SageMaker Batch Transform.

A: For specific file formats such as CSV, RecordIO and TFRecord, SageMaker can split your data into single-record or multi-record mini batches and send this as a payload to your model container. When the value of BatchStrategy is MultiRecord, SageMaker sends the maximum number of records in each request, up to the MaxPayloadInMB limit. When the value of BatchStrategy is SingleRecord, SageMaker sends individual records in each request.

A: The maximum timeout for Batch Transform is 3600 seconds. The maximum payload size for a record (per mini batch) is 100 MB.

A: If you are using the CreateTransformJob API, you can reduce the time it takes to complete batch transform jobs by using optimal values for parameters such as MaxPayloadInMB, MaxConcurrentTransforms, or BatchStrategy. The ideal value for MaxConcurrentTransforms is equal to the number of compute workers in the batch transform job. If you are using the SageMaker console, you can specify these optimal parameter values in the Additional configuration section of the Batch transform job configuration page. SageMaker automatically finds the optimal parameter settings for built-in algorithms. For custom algorithms, provide these values through an execution-parameters endpoint.

A: Batch Transform supports CSV and JSON.

Asynchronous Inference

The following FAQ items answer common general questions for SageMaker Asynchronous Inference.

A: Asynchronous Inference queues incoming requests and processes them asynchronously. This option is ideal for requests with large payload sizes or long processing times that need to be processed as they arrive. Optionally, you can configure auto-scaling settings to scale down the instance count to zero when not actively processing requests.

A: Amazon SageMaker supports automatic scaling (autoscaling) your asynchronous endpoint. Autoscaling dynamically adjusts the number of instances provisioned for a model in response to changes in your workload. Unlike other hosted models SageMaker supports, with Asynchronous Inference you can also scale down your asynchronous endpoints instances to zero. Requests that are received when there are zero instances are queued for processing once the endpoint scales up. For more information, see Autoscale an asynchronous endpoint.

Amazon SageMaker Serverless Inference also automatically scales down to zero. You won’t see this because SageMaker manages scaling your serverless endpoints, but if you are not experiencing any traffic, the same infrastructure applies.