Use PyTorch with the SageMaker Python SDK

With PyTorch Estimators and Models, you can train and host PyTorch models on Amazon SageMaker.

For information about supported versions of PyTorch, see the AWS documentation.

We recommend that you use the latest supported version because that’s where we focus our development efforts.

You can visit the PyTorch repository at https://github.com/pytorch/pytorch.

Train a Model with PyTorch

To train a PyTorch model by using the SageMaker Python SDK:

  1. Prepare a training script OR Choose an Amazon SageMaker HyperPod recipe

  2. Create a sagemaker.pytorch.PyTorch Estimator

  3. Call the estimator’s fit method or ModelTrainer’s train method

Prepare a PyTorch Training Script

Your PyTorch training script must be a Python 3.6 compatible source file.

Prepare your script in a separate source file than the notebook, terminal session, or source file you’re using to submit the script to SageMaker via a PyTorch Estimator. This will be discussed in further detail below.

The training script is very similar to a training script you might run outside of SageMaker, but you can access useful properties about the training environment through various environment variables. For example:

  • SM_NUM_GPUS: An integer representing the number of GPUs available to the host.

  • SM_MODEL_DIR: A string representing the path to the directory to write model artifacts to. These artifacts are uploaded to S3 for model hosting.

  • SM_OUTPUT_DATA_DIR: A string representing the filesystem path to write output artifacts to. Output artifacts may include checkpoints, graphs, and other files to save, not including model artifacts. These artifacts are compressed and uploaded to S3 to the same S3 prefix as the model artifacts.

  • SM_CHANNEL_XXXX: A string that represents the path to the directory that contains the input data for the specified channel. For example, if you specify two input channels in the PyTorch estimator’s fit call, named ‘train’ and ‘test’, the environment variables SM_CHANNEL_TRAIN and SM_CHANNEL_TEST are set.

A typical training script loads data from the input channels, configures training with hyperparameters, trains a model, and saves a model to model_dir so that it can be hosted later. Hyperparameters are passed to your script as arguments and can be retrieved with an argparse.ArgumentParser instance. For example, a training script might start with the following:

import argparse
import os

if __name__ =='__main__':

    parser = argparse.ArgumentParser()

    # hyperparameters sent by the client are passed as command-line arguments to the script.
    parser.add_argument('--epochs', type=int, default=50)
    parser.add_argument('--batch-size', type=int, default=64)
    parser.add_argument('--learning-rate', type=float, default=0.05)
    parser.add_argument('--use-cuda', type=bool, default=False)

    # Data, model, and output directories
    parser.add_argument('--output-data-dir', type=str, default=os.environ['SM_OUTPUT_DATA_DIR'])
    parser.add_argument('--model-dir', type=str, default=os.environ['SM_MODEL_DIR'])
    parser.add_argument('--train', type=str, default=os.environ['SM_CHANNEL_TRAIN'])
    parser.add_argument('--test', type=str, default=os.environ['SM_CHANNEL_TEST'])

    args, _ = parser.parse_known_args()

    # ... load from args.train and args.test, train a model, write model to args.model_dir.

Because SageMaker imports your training script, you should put your training code in a main guard (if __name__=='__main__':) if you are using the same script to host your model, so that SageMaker does not inadvertently run your training code at the wrong point in execution.

Note that SageMaker doesn’t support argparse actions. If you want to use, for example, boolean hyperparameters, you need to specify type as bool in your script and provide an explicit True or False value for this hyperparameter when instantiating PyTorch Estimator.

For more on training environment variables, see the SageMaker Training Toolkit.

Save the Model

In order to save your trained PyTorch model for deployment on SageMaker, your training script should save your model to a certain filesystem path called model_dir. This value is accessible through the environment variable SM_MODEL_DIR. The following code demonstrates how to save a trained PyTorch model named model as model.pth at the :

import argparse
import os
import torch

if __name__=='__main__':
    # default to the value in environment variable `SM_MODEL_DIR`. Using args makes the script more portable.
    parser.add_argument('--model-dir', type=str, default=os.environ['SM_MODEL_DIR'])
    args, _ = parser.parse_known_args()

    # ... train `model`, then save it to `model_dir`
    with open(os.path.join(args.model_dir, 'model.pth'), 'wb') as f:
        torch.save(model.state_dict(), f)

After your training job is complete, SageMaker compresses and uploads the serialized model to S3, and your model data will be available in the S3 output_path you specified when you created the PyTorch Estimator.

If you are using Elastic Inference, you must convert your models to the TorchScript format and use torch.jit.save to save the model. For example:

import os
import torch

# ... train `model`, then save it to `model_dir`
model_dir = os.path.join(model_dir, "model.pt")
torch.jit.save(model, model_dir)

Using third-party libraries

When running your training script on SageMaker, it will have access to some pre-installed third-party libraries including torch, torchvision, and numpy.

If there are other packages you want to use with your script, you can include a requirements.txt file in the same directory as your training script to install other dependencies at runtime. Both requirements.txt and your training script should be put in the same folder. You must specify this folder in source_dir argument when creating PyTorch estimator.

The function of installing packages using requirements.txt is supported for all PyTorch versions during training. When serving a PyTorch model, support for this function varies with PyTorch versions. For PyTorch 1.3.1 or newer, requirements.txt must be under folder code. The SageMaker PyTorch Estimator will automatically save code in model.tar.gz after training (assuming you set up your script and requirements.txt correctly as stipulated in the previous paragraph). In the case of bringing your own trained model for deployment, you must save requirements.txt under folder code in model.tar.gz yourself or specify it through dependencies. For PyTorch 1.2.0, requirements.txt is not supported for inference. For PyTorch 0.4.0 to 1.1.0, requirements.txt must be in source_dir.

A requirements.txt file is a text file that contains a list of items that are installed by using pip install. You can also specify the version of an item to install. For information about the format of a requirements.txt file, see Requirements Files in the pip documentation.

If you were to use your own custom Docker Image, the SageMaker Python SDK and the SageMaker Training Toolkit need to be installed.

To do so, you can add the following lines to your requirements.txt file:

sagemaker
sagemaker-training

Deep Learning Framework-Specific SageMaker Toolkits and Containers

Framework-specific Toolkits exist. You might want to use them in your applications for framework-specific features.

For Training Toolkits, see:

For Inference Toolkits, see:

Moreover, for more information on the container runtime environment, including specific framework versions and configurations, see AWS Deep Learning Containers. More specifically, see:

Choose an Amazon Sagemaker HyperPod recipe

Alternatively, instead of using your own training script, you can choose an Amazon SageMaker HyperPod recipe to launch training for a supported model. If using a recipe, you do not need to provide your own training script. You only need to determine which recipe you want to run. You can modify a recipe as explained in the next section.

Create an Estimator

You run PyTorch training scripts on SageMaker by creating PyTorch Estimators. SageMaker training of your script is invoked when you call fit on a PyTorch Estimator. The following code sample shows how you train a custom PyTorch script “pytorch-train.py”, passing in three hyperparameters (‘epochs’, ‘batch-size’, and ‘learning-rate’), and using two input channel directories (‘train’ and ‘test’).

pytorch_estimator = PyTorch('pytorch-train.py',
                            instance_type='ml.p3.2xlarge',
                            instance_count=1,
                            framework_version='1.8.0',
                            py_version='py3',
                            hyperparameters = {'epochs': 20, 'batch-size': 64, 'learning-rate': 0.1})
pytorch_estimator.fit({'train': 's3://my-data-bucket/path/to/my/training/data',
                       'test': 's3://my-data-bucket/path/to/my/test/data'})

Amazon Sagemaker HyperPod recipes

Alternatively, if you are using Amazon SageMaker HyperPod recipes, you can follow the following instructions:

Prerequisites: you need git installed on your client to access Amazon SageMaker HyperPod recipes code.

When using a recipe, you must set the training_recipe arg in place of providing a training script. This can be a recipe from here or a local file or a custom url. Please note that you must override the following using recipe_overrides:

  • directory paths for the local container in the recipe as appropriate for Python SDK

  • the output s3 URIs

  • Huggingface access token

  • any other recipe fields you wish to edit

The code snippet below shows an example. Please refer to SageMaker docs for more details about the expected local paths in the container and the Amazon SageMaker HyperPod recipes tutorial for more examples. You can override the fields by either setting recipe_overrides or providing a modified training_recipe through a local file or a custom url. When using the recipe, any provided entry_point will be ignored.

SageMaker will automatically set up the distribution args. It will also determine the image to use for your model and device type, but you can override this with the image_uri arg.

You can also override the number of nodes in the recipe with the instance_count arg to estimator. source_dir will default to current working directory unless specified. A local copy of training scripts and recipe will be saved in the source_dir. You can specify any additional packages you want to install for training in an optional requirements.txt in the source_dir.

Note for llama3.2 multi-modal models, you need to upgrade transformers library by providing a requirements.txt in the source file with transformers==4.45.2. Please refer to the Amazon SageMaker HyperPod recipes documentation for more details.

Here is an example usage for recipe hf_llama3_8b_seq8k_gpu_p5x16_pretrain.

overrides = {
    "run": {
        "results_dir": "/opt/ml/model",
    },
    "exp_manager": {
        "exp_dir": "",
        "explicit_log_dir": "/opt/ml/output/tensorboard",
        "checkpoint_dir": "/opt/ml/checkpoints",
    },
    "model": {
        "data": {
            "train_dir": "/opt/ml/input/data/train",
            "val_dir": "/opt/ml/input/data/val",
        },
    },
}
pytorch_estimator = PyTorch(
  output_path=output_path,
  base_job_name=f"llama-recipe",
  role=role,
  instance_type="ml.p5.48xlarge",
  training_recipe="hf_llama3_8b_seq8k_gpu_p5x16_pretrain",
  recipe_overrides=recipe_overrides,
  sagemaker_session=sagemaker_session,
  tensorboard_output_config=tensorboard_output_config,
)
pytorch_estimator.fit({'train': 's3://my-data-bucket/path/to/my/training/data',
                       'test': 's3://my-data-bucket/path/to/my/test/data'})

# Or alternatively with ModelTrainer
recipe_overrides = {
    "run": {
        "results_dir": "/opt/ml/model",
    },
    "exp_manager": {
        "exp_dir": "",
        "explicit_log_dir": "/opt/ml/output/tensorboard",
        "checkpoint_dir": "/opt/ml/checkpoints",
    },
    "model": {
        "data": {
            "train_dir": "/opt/ml/input/data/train",
            "val_dir": "/opt/ml/input/data/val",
        },
    },
}

model_trainer = ModelTrainer.from_recipe(
    output_path=output_path,
    base_job_name=f"llama-recipe",
    training_recipe="training/llama/hf_llama3_8b_seq8k_gpu_p5x16_pretrain",
    recipe_overrides=recipe_overrides,
    compute=Compute(instance_type="ml.p5.48xlarge"),
    sagemaker_session=sagemaker_session
).with_tensorboard_output_config(
    tensorboard_output_config=tensorboard_output_config
)

train_input = Input(
    channel_name="train",
    data_source="s3://my-data-bucket/path/to/my/training/data"
)

test_input = Input(
    channel_name="test",
    data_source="s3://my-data-bucket/path/to/my/test/data"
)

model_trainer.train(input_data_config=[train_input, test_input)

Call the estimator’s fit method or ModelTrainer’s train method

You start your training script by calling fit on a PyTorch Estimator. fit takes both required and optional arguments.

fit Required Arguments

  • inputs: This can take one of the following forms: A string S3 URI, for example s3://my-bucket/my-training-data. In this case, the S3 objects rooted at the my-training-data prefix will be available in the default train channel. A dict from string channel names to S3 URIs. In this case, the objects rooted at each S3 prefix will be available as files in each channel directory.

For example:

{'train':'s3://my-bucket/my-training-data',
 'eval':'s3://my-bucket/my-evaluation-data'}

fit Optional Arguments

  • wait: Defaults to True, whether to block and wait for the training script to complete before returning.

  • logs: Defaults to True, whether to show logs produced by training job in the Python session. Only meaningful when wait is True.

Distributed PyTorch Training

SageMaker supports the PyTorch DistributedDataParallel (DDP) package. You simply need to check the variables in your training script, such as the world size and the rank of the current host, when initializing process groups for distributed training. And then, launch the training job using the sagemaker.pytorch.estimator.PyTorch estimator class with the pytorchddp option as the distribution strategy.

Note

This PyTorch DDP support is available in the SageMaker PyTorch Deep Learning Containers v1.12 and later.

Adapt Your Training Script

To initialize distributed training in your script, call torch.distributed.init_process_group with the desired backend and the rank of the current host.

import torch.distributed as dist

if args.distributed:
    # Initialize the distributed environment.
    world_size = len(args.hosts)
    os.environ['WORLD_SIZE'] = str(world_size)
    host_rank = args.hosts.index(args.current_host)
    dist.init_process_group(backend=args.backend, rank=host_rank)

SageMaker sets 'MASTER_ADDR' and 'MASTER_PORT' environment variables for you, but you can also overwrite them.

Supported backends:

  • gloo and tcp for CPU instances

  • gloo and nccl for GPU instances

Launching a Distributed Training Job

You can run multi-node distributed PyTorch training jobs using the sagemaker.pytorch.estimator.PyTorch estimator class. With instance_count=1, the estimator submits a single-node training job to SageMaker; with instance_count greater than one, a multi-node training job is launched.

To run a distributed training script that adopts the PyTorch DistributedDataParallel (DDP) package, choose the pytorchddp as the distributed training option in the PyTorch estimator.

With the pytorchddp option, the SageMaker PyTorch estimator runs a SageMaker training container for PyTorch, sets up the environment for MPI, and launches the training job using the mpirun command on each worker with the given information during the PyTorch DDP initialization.

Note

The SageMaker PyTorch estimator can operate both mpirun (for PyTorch 1.12.0 and later) and torchrun (for PyTorch 1.13.1 and later) in the backend for distributed training.

For more information about setting up PyTorch DDP in your training script, see Getting Started with Distributed Data Parallel in the PyTorch documentation.

The following examples show how to set a PyTorch estimator to run a distributed training job on two ml.p4d.24xlarge instances.

Using PyTorch DDP with the mpirun backend

from sagemaker.pytorch import PyTorch

pt_estimator = PyTorch(
    entry_point="train_ptddp.py",
    role="SageMakerRole",
    framework_version="1.12.0",
    py_version="py38",
    instance_count=2,
    instance_type="ml.p4d.24xlarge",
    distribution={
        "pytorchddp": {
            "enabled": True
        }
    }
)

Using PyTorch DDP with the torchrun backend

from sagemaker.pytorch import PyTorch

pt_estimator = PyTorch(
    entry_point="train_ptddp.py",
    role="SageMakerRole",
    framework_version="1.13.1",
    py_version="py38",
    instance_count=2,
    instance_type="ml.p4d.24xlarge",
    distribution={
        "torch_distributed": {
            "enabled": True
        }
    }
)

Note

For more information about setting up torchrun in your training script, see torchrun (Elastic Launch) in the PyTorch documentation.

Distributed Training with PyTorch Neuron on Trn1 instances

SageMaker Training supports Amazon EC2 Trn1 instances powered by AWS Trainium device, the second generation purpose-built machine learning accelerator from AWS. Each Trn1 instance consists of up to 16 Trainium devices, and each Trainium device consists of two NeuronCores in the AWS Neuron Documentation.

You can run distributed training job on Trn1 instances. SageMaker supports the xla package through torchrun. With this, you do not need to manually pass RANK, WORLD_SIZE, MASTER_ADDR, and MASTER_PORT. You can launch the training job using the sagemaker.pytorch.estimator.PyTorch estimator class with the torch_distributed option as the distribution strategy.

Note

This torch_distributed support is available in the AWS Deep Learning Containers for PyTorch Neuron starting v1.11.0. To find a complete list of supported versions of PyTorch Neuron, see Neuron Containers in the AWS Deep Learning Containers GitHub repository.

Note

SageMaker Debugger is not compatible with Trn1 instances.

Adapt Your Training Script to Initialize with the XLA backend

To initialize distributed training in your script, call torch.distributed.init_process_group with the xla backend as shown below.

import torch.distributed as dist

dist.init_process_group('xla')

SageMaker takes care of 'MASTER_ADDR' and 'MASTER_PORT' for you via torchrun

For detailed documentation about modifying your training script for Trainium, see Multi-worker data-parallel MLP training using torchrun in the AWS Neuron Documentation.

Currently Supported backends:

  • xla for Trainium (Trn1) instances

For up-to-date information on supported backends for Trn1 instances, see AWS Neuron Documentation.

Launching a Distributed Training Job on Trainium

You can run multi-node distributed PyTorch training jobs on Trn1 instances using the sagemaker.pytorch.estimator.PyTorch estimator class. With instance_count=1, the estimator submits a single-node training job to SageMaker; with instance_count greater than one, a multi-node training job is launched.

With the torch_distributed option, the SageMaker PyTorch estimator runs a SageMaker training container for PyTorch Neuron, sets up the environment, and launches the training job using the torchrun command on each worker with the given information.

Examples

The following examples show how to run a PyTorch training using torch_distributed in SageMaker on one ml.trn1.2xlarge instance and two ml.trn1.32xlarge instances:

from sagemaker.pytorch import PyTorch

pt_estimator = PyTorch(
    entry_point="train_torch_distributed.py",
    role="SageMakerRole",
    framework_version="1.11.0",
    py_version="py38",
    instance_count=1,
    instance_type="ml.trn1.2xlarge",
    distribution={
        "torch_distributed": {
            "enabled": True
        }
    }
)

pt_estimator.fit("s3://bucket/path/to/training/data")

from sagemaker.pytorch import PyTorch

pt_estimator = PyTorch(
    entry_point="train_torch_distributed.py",
    role="SageMakerRole",
    framework_version="1.11.0",
    py_version="py38",
    instance_count=2,
    instance_type="ml.trn1.32xlarge",
    distribution={
        "torch_distributed": {
            "enabled": True
        }
    }
)

pt_estimator.fit("s3://bucket/path/to/training/data")

Deploy PyTorch Models

After a PyTorch Estimator has been fit, you can host the newly created model in SageMaker.

After calling fit, you can call deploy on a PyTorch Estimator to create a SageMaker Endpoint. The Endpoint runs a SageMaker-provided PyTorch model server and hosts the model produced by your training script, which was run when you called fit. This was the model you saved to model_dir.

deploy returns a Predictor object, which you can use to do inference on the Endpoint hosting your PyTorch model. Each Predictor provides a predict method which can do inference with numpy arrays or Python lists. Inference arrays or lists are serialized and sent to the PyTorch model server by an InvokeEndpoint SageMaker operation.

predict returns the result of inference against your model. By default, the inference result a NumPy array.

# Train my estimator
pytorch_estimator = PyTorch(entry_point='train_and_deploy.py',
                            instance_type='ml.p3.2xlarge',
                            instance_count=1,
                            framework_version='1.8.0',
                            py_version='py3')
pytorch_estimator.fit('s3://my_bucket/my_training_data/')

# Deploy my estimator to a SageMaker Endpoint and get a Predictor
predictor = pytorch_estimator.deploy(instance_type='ml.m4.xlarge',
                                     initial_instance_count=1)

# `data` is a NumPy array or a Python list.
# `response` is a NumPy array.
response = predictor.predict(data)

You use the SageMaker PyTorch model server to host your PyTorch model when you call deploy on an PyTorch Estimator. The model server runs inside a SageMaker Endpoint, which your call to deploy creates. You can access the name of the Endpoint by the name property on the returned Predictor.

Elastic Inference

PyTorch on Amazon SageMaker has support for Elastic Inference, which allows for inference acceleration to a hosted endpoint for a fraction of the cost of using a full GPU instance. In order to attach an Elastic Inference accelerator to your endpoint provide the accelerator type to accelerator_type to your deploy call.

predictor = pytorch_estimator.deploy(instance_type='ml.m4.xlarge',
                                     initial_instance_count=1,
                                     accelerator_type='ml.eia2.medium')

Model Directory Structure

In general, if you use the same version of PyTorch for both training and inference with the SageMaker Python SDK, the SDK should take care of ensuring that the contents of your model.tar.gz file are organized correctly.

For versions 1.2 and higher

For PyTorch versions 1.2 and higher, the contents of model.tar.gz should be organized as follows:

  • Model files in the top-level directory

  • Inference script (and any other source files) in a directory named code/ (for more about the inference script, see The SageMaker PyTorch Model Server)

  • Optional requirements file located at code/requirements.txt (for more about requirements files, see Using third-party libraries)

For example:

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

In this example, model.pth is the model file saved from training, inference.py is the inference script, and requirements.txt is a requirements file.

The PyTorch and PyTorchModel classes repack model.tar.gz to include the inference script (and related files), as long as the framework_version is set to 1.2 or higher.

For versions 1.1 and lower

For PyTorch versions 1.1 and lower, model.tar.gz should contain only the model files, while your inference script and optional requirements file are packed in a separate tarball, named sourcedir.tar.gz by default.

For example:

model.tar.gz/
|- model.pth

sourcedir.tar.gz/
|- script.py
|- requirements.txt

In this example, model.pth is the model file saved from training, script.py is the inference script, and requirements.txt is a requirements file.

The SageMaker PyTorch Model Server

The PyTorch Endpoint you create with deploy runs a SageMaker PyTorch model server. The model server loads the model that was saved by your training script and performs inference on the model in response to SageMaker InvokeEndpoint API calls.

You can configure two components of the SageMaker PyTorch model server: Model loading and model serving. Model loading is the process of deserializing your saved model back into a PyTorch model. Serving is the process of translating InvokeEndpoint requests to inference calls on the loaded model.

You configure the PyTorch model server by defining functions in the Python source file you passed to the PyTorch constructor.

Load a Model

Before a model can be served, it must be loaded. The SageMaker PyTorch model server loads your model by invoking a model_fn function that you must provide in your script when you are not using Elastic Inference. The model_fn should have the following signature:

def model_fn(model_dir, context)

context is an optional argument that contains additional serving information, such as the GPU ID and batch size. If specified in the function declaration, the context will be created and passed to the function by SageMaker. For more information about context, see the Serving Context class.

SageMaker will inject the directory where your model files and sub-directories, saved by save, have been mounted. Your model function should return a model object that can be used for model serving.

The following code-snippet shows an example model_fn implementation. It loads the model parameters from a model.pth file in the SageMaker model directory model_dir. As explained in the preceding example, context is an optional argument that passes additional information.

import torch
import os

def model_fn(model_dir, context):
    model = Your_Model()
    with open(os.path.join(model_dir, 'model.pth'), 'rb') as f:
        model.load_state_dict(torch.load(f))
    return model

However, if you are using PyTorch Elastic Inference 1.3.1, you do not have to provide a model_fn since the PyTorch serving container has a default one for you. But please note that if you are utilizing the default model_fn, please save your ScriptModule as model.pt. If you are implementing your own model_fn, please use TorchScript and torch.jit.save to save your ScriptModule, then load it in your model_fn with torch.jit.load(..., map_location=torch.device('cpu')).

If you are using PyTorch Elastic Inference 1.5.1, you should provide model_fn like below in your script to use new api attach_eia. Reference can be find in Elastic Inference documentation.

import torch


def model_fn(model_dir):
    model = torch.jit.load('model.pth', map_location=torch.device('cpu'))
    if torch.__version__ == '1.5.1':
        import torcheia
        model = model.eval()
        # attach_eia() is introduced in PyTorch Elastic Inference 1.5.1,
        model = torcheia.jit.attach_eia(model, 0)
    return model

The client-side Elastic Inference framework is CPU-only, even though inference still happens in a CUDA context on the server. Thus, the default model_fn for Elastic Inference loads the model to CPU. Tracing models may lead to tensor creation on a specific device, which may cause device-related errors when loading a model onto a different device. Providing an explicit map_location=torch.device('cpu') argument forces all tensors to CPU.

For more information on the default inference handler functions, please refer to: SageMaker PyTorch Default Inference Handler.

Serve a PyTorch Model

After the SageMaker model server has loaded your model by calling model_fn, SageMaker will serve your model. Model serving is the process of responding to inference requests, received by SageMaker InvokeEndpoint API calls. The SageMaker PyTorch model server breaks request handling into three steps:

  • input processing,

  • prediction, and

  • output processing.

In a similar way to model loading, you configure these steps by defining functions in your Python source file.

Each step involves invoking a python function, with information about the request and the return value from the previous function in the chain. Inside the SageMaker PyTorch model server, the process looks like:

# Deserialize the Invoke request body into an object we can perform prediction on
input_object = input_fn(request_body, request_content_type, context)

# Perform prediction on the deserialized object, with the loaded model
prediction = predict_fn(input_object, model, context)

# Serialize the prediction result into the desired response content type
output = output_fn(prediction, response_content_type, context)

The above code sample shows the three function definitions:

  • input_fn: Takes request data and deserializes the data into an object for prediction.

  • predict_fn: Takes the deserialized request object and performs inference against the loaded model.

  • output_fn: Takes the result of prediction and serializes this according to the response content type.

The SageMaker PyTorch model server provides default implementations of these functions. You can provide your own implementations for these functions in your hosting script. If you omit any definition then the SageMaker PyTorch model server will use its default implementation for that function. If you use PyTorch Elastic Inference 1.5.1, remember to implement predict_fn yourself.

The Predictor used by PyTorch in the SageMaker Python SDK serializes NumPy arrays to the NPY format by default, with Content-Type application/x-npy. The SageMaker PyTorch model server can deserialize NPY-formatted data (along with JSON and CSV data).

If you rely solely on the SageMaker PyTorch model server defaults, you get the following functionality:

  • Prediction on models that implement the __call__ method

  • Serialization and deserialization of torch.Tensor.

The default input_fn and output_fn are meant to make it easy to predict on torch.Tensors. If your model expects a torch.Tensor and returns a torch.Tensor, then these functions do not have to be overridden when sending NPY-formatted data.

In the following sections we describe the default implementations of input_fn, predict_fn, and output_fn. We describe the input arguments and expected return types of each, so you can define your own implementations.

Process Model Input

When an InvokeEndpoint operation is made against an Endpoint running a SageMaker PyTorch model server, the model server receives two pieces of information:

  • The request Content-Type, for example “application/x-npy”

  • The request data body, a byte array

The SageMaker PyTorch model server will invoke an input_fn function in your hosting script, passing in this information. If you define an input_fn function definition, it should return an object that can be passed to predict_fn and have the following signature:

def input_fn(request_body, request_content_type, context)

Where request_body is a byte buffer and request_content_type is a Python string.

context is an optional argument that contains additional serving information, such as the GPU ID and batch size. If specified in the function declaration, the context will be created and passed to the function by SageMaker. For more information about context, see the Serving Context class.

The SageMaker PyTorch model server provides a default implementation of input_fn. This function deserializes JSON, CSV, or NPY encoded data into a torch.Tensor.

Default NPY deserialization requires request_body to follow the NPY format. For PyTorch, the Python SDK defaults to sending prediction requests with this format.

Default JSON deserialization requires request_body contain a single json list. Sending multiple JSON objects within the same request_body is not supported. The list must have a dimensionality compatible with the model loaded in model_fn. The list’s shape must be identical to the model’s input shape, for all dimensions after the first (which first dimension is the batch size).

Default csv deserialization requires request_body contain one or more lines of CSV numerical data. The data is loaded into a two-dimensional array, where each line break defines the boundaries of the first dimension.

The example below shows a custom input_fn for preparing pickled torch.Tensor.

import numpy as np
import torch
from six import BytesIO

def input_fn(request_body, request_content_type):
    """An input_fn that loads a pickled tensor"""
    if request_content_type == 'application/python-pickle':
        return torch.load(BytesIO(request_body))
    else:
        # Handle other content-types here or raise an Exception
        # if the content type is not supported.
        pass
Get Predictions from a PyTorch Model

After the inference request has been deserialized by input_fn, the SageMaker PyTorch model server invokes predict_fn on the return value of input_fn.

As with input_fn, you can define your own predict_fn or use the SageMaker PyTorch model server default.

The predict_fn function has the following signature:

def predict_fn(input_object, model, context)

Where input_object is the object returned from input_fn and model is the model loaded by model_fn. If you are using multiple GPUs, then specify the context argument, which contains information such as the GPU ID for a dynamically-selected GPU and the batch size. One of the examples below demonstrates how to configure predict_fn with the context argument to handle multiple GPUs. For more information about context, see the Serving Context class. If you are using CPUs or a single GPU, then you do not need to specify the context argument.

The default implementation of predict_fn invokes the loaded model’s __call__ function on input_object, and returns the resulting value. The return-type should be a torch.Tensor to be compatible with the default output_fn.

The following example shows an overridden predict_fn:

import torch
import numpy as np

def predict_fn(input_data, model):
    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
    model.to(device)
    model.eval()
    with torch.no_grad():
        return model(input_data.to(device))

The following example is for use cases with multiple GPUs and shows an overridden predict_fn that uses the context argument to dynamically select a GPU device for making predictions:

import torch
import numpy as np

def predict_fn(input_data, model, context):
    device = torch.device("cuda:" + str(context.system_properties.get("gpu_id")) if torch.cuda.is_available() else "cpu")
    model.to(device)
    model.eval()
    with torch.no_grad():
        return model(input_data.to(device))

If you implement your own prediction function, you should take care to ensure that:

  • The first argument is expected to be the return value from input_fn. If you use the default input_fn, this will be a torch.Tensor.

  • The second argument is the loaded model.

  • The return value should be of the correct type to be passed as the first argument to output_fn. If you use the default output_fn, this should be a torch.Tensor.

The default Elastic Inference predict_fn is similar but runs the TorchScript model using torch.jit.optimized_execution. If you are implementing your own predict_fn, please also use the torch.jit.optimized_execution block, for example:

import torch
import numpy as np

def predict_fn(input_data, model):
    device = torch.device("cpu")
    model = model.to(device)
    input_data = data.to(device)
    model.eval()
    with torch.jit.optimized_execution(True, {"target_device": "eia:0"}):
        output = model(input_data)

If you use PyTorch Elastic Inference 1.5.1, please implement your own predict_fn like below.

import numpy as np
import torch


def predict_fn(input_data, model):
    device = torch.device("cpu")
    input_data = data.to(device)
    # make sure torcheia is imported so that Elastic Inference api call will be invoked
    import torcheia
    # we need to set the profiling executor for EIA
    torch._C._jit_set_profiling_executor(False)
    with torch.jit.optimized_execution(True):
        output = model.forward(input_data)
Process Model Output

After invoking predict_fn, the model server invokes output_fn, passing in the return value from predict_fn and the content type for the response, as specified by the InvokeEndpoint request.

The output_fn has the following signature:

def output_fn(prediction, content_type, context)

Where prediction is the result of invoking predict_fn and the content type for the response, as specified by the InvokeEndpoint request. The function should return a byte array of data serialized to content_type.

context is an optional argument that contains additional serving information, such as the GPU ID and batch size. If specified in the function declaration, the context will be created and passed to the function by SageMaker. For more information about context, see the Serving Context class.

The default implementation expects prediction to be a torch.Tensor and can serialize the result to JSON, CSV, or NPY. It accepts response content types of “application/json”, “text/csv”, and “application/x-npy”.

Bring your own model

You can deploy a PyTorch model that you trained outside of SageMaker by using the PyTorchModel class. Typically, you save a PyTorch model as a file with extension .pt or .pth. To do this, you need to:

  • Write an inference script.

  • Create the directory structure for your model files.

  • Create the PyTorchModel object.

Write an inference script

You must create an inference script that implements (at least) the model_fn function that calls the loaded model to get a prediction.

Note: If you use elastic inference with PyTorch, you can use the default model_fn implementation provided in the serving container.

Optionally, you can also implement input_fn and output_fn to process input and output, and predict_fn to customize how the model server gets predictions from the loaded model. For information about how to write an inference script, see Serve a PyTorch Model. Save the inference script in the same folder where you saved your PyTorch model. Pass the filename of the inference script as the entry_point parameter when you create the PyTorchModel object.

Create the directory structure for your model files

You have to create a directory structure and place your model files in the correct location. The PyTorchModel constructor packs the files into a tar.gz file and uploads it to S3.

The directory structure where you saved your PyTorch model should look something like the following:

Note: This directory struture is for PyTorch versions 1.2 and higher. For the directory structure for versions 1.1 and lower, see For versions 1.1 and lower.

|   my_model
|           |--model.pth
|
|           code
|               |--inference.py
|               |--requirements.txt

Where requirements.txt is an optional file that specifies dependencies on third-party libraries.

Create a PyTorchModel object

Now call the sagemaker.pytorch.model.PyTorchModel constructor to create a model object, and then call its deploy() method to deploy your model for inference.

from sagemaker import get_execution_role
role = get_execution_role()

pytorch_model = PyTorchModel(model_data='s3://my-bucket/my-path/model.tar.gz', role=role,
                             entry_point='inference.py')

predictor = pytorch_model.deploy(instance_type='ml.c4.xlarge', initial_instance_count=1)

Now you can call the predict() method to get predictions from your deployed model.

Attach an estimator to an existing training job

You can attach a PyTorch Estimator to an existing training job using the attach method.

my_training_job_name = 'MyAwesomePyTorchTrainingJob'
pytorch_estimator = PyTorch.attach(my_training_job_name)

After attaching, if the training job has finished with job status “Completed”, it can be deployed to create a SageMaker Endpoint and return a Predictor. If the training job is in progress, attach will block and display log messages from the training job, until the training job completes.

The attach method accepts the following arguments:

  • training_job_name: The name of the training job to attach to.

  • sagemaker_session: The Session used to interact with SageMaker

PyTorch Training Examples

Amazon provides several example Jupyter notebooks that demonstrate end-to-end training on Amazon SageMaker using PyTorch. Please refer to:

https://github.com/awslabs/amazon-sagemaker-examples/tree/master/sagemaker-python-sdk

These are also available in SageMaker Notebook Instance hosted Jupyter notebooks under the sample notebooks folder.

SageMaker PyTorch Classes

For information about the different PyTorch-related classes in the SageMaker Python SDK, see https://sagemaker.readthedocs.io/en/stable/frameworks/pytorch/sagemaker.pytorch.html.