TorchServe is a flexible and easy to use tool for serving PyTorch models.

For full documentation, see Model Server for PyTorch Documentation.

• Frontend: The request/response handling component of TorchServe. This portion of the serving component handles both request/response coming from clients and manages the lifecycles of the models.
• Model Workers: These workers are responsible for running the actual inference on the models. These are actual running instances of the models.
• Model: Models could be a script_module (JIT saved models) or eager_mode_models. These models can provide custom pre- and post-processing of data along with any other model artifacts such as state_dicts. Models can be loaded from cloud storage or from local hosts.
• Plugins: These are custom endpoints or authz/authn or batching algorithms that can be dropped into TorchServe at startup time.
• Model Store: This is a directory in which all the loadable models exist.

• Install TorchServe
• Install TorchServe on windows subsystem for linux
• Serve a Model
• Quick start with docker
• Contributing

1. Install Java 11

   For Ubuntu

   sudo apt-get install openjdk-11-jdk

   For Mac

   brew tap AdoptOpenJDK/openjdk
   brew cask install adoptopenjdk11

2. Install python pre-requisite packages

• For CPU or GPU-Cuda 10.2

  pip install -U -r requirements.txt

• For GPU with Cuda 10.1

  pip install -U -r requirements_gpu.txt

3. Install torchserve and torch-model-archiver

   For Conda

   conda install torchserve torch-model-archiver
   

   For Pip

   pip install torchserve torch-model-archiver
   

   Note: For Conda, Python 3.8 is required to run Torchserve

Now you are ready to package and serve models with TorchServe.

If you plan to develop with TorchServe and change some of the source code, you must install it from source code.

Please deactivate any conda env that you might be within. Run the following script from the top of the source directory.

NOTE: This script uninstalls existing torchserve and torch-model-archiver installations

./scripts/install_from_src

For information about the model archiver, see detailed documentation.

This section shows a simple example of serving a model with TorchServe. To complete this example, you must have already installed TorchServe and the model archiver.

To run this example, clone the TorchServe repository:

git clone https://github.com/pytorch/serve.git

Then run the following steps from the parent directory of the root of the repository. For example, if you cloned the repository into /home/my_path/serve, run the steps from /home/my_path.

To serve a model with TorchServe, first archive the model as a MAR file. You can use the model archiver to package a model. You can also create model stores to store your archived models.

1. Create a directory to store your models.

   mkdir model_store

2. Download a trained model.

   wget https://download.pytorch.org/models/densenet161-8d451a50.pth

3. Archive the model by using the model archiver. The extra-files param uses fa file from the TorchServe repo, so update the path if necessary.

   torch-model-archiver --model-name densenet161 --version 1.0 --model-file ./serve/examples/image_classifier/densenet_161/model.py --serialized-file densenet161-8d451a50.pth --export-path model_store --extra-files ./serve/examples/image_classifier/index_to_name.json --handler image_classifier

For more information about the model archiver, see Torch Model archiver for TorchServe

After you archive and store the model, use the torchserve command to serve the model.

torchserve --start --ncs --model-store model_store --models densenet161.mar

After you execute the torchserve command above, TorchServe runs on your host, listening for inference requests.

Note: If you specify model(s) when you run TorchServe, it automatically scales backend workers to the number equal to available vCPUs (if you run on a CPU instance) or to the number of available GPUs (if you run on a GPU instance). In case of powerful hosts with a lot of compute resoures (vCPUs or GPUs), this start up and autoscaling process might take considerable time. If you want to minimize TorchServe start up time you should avoid registering and scaling the model during start up time and move that to a later point by using corresponding Management API, which allows finer grain control of the resources that are allocated for any particular model).

To test the model server, send a request to the server's predictions API.

Complete the following steps:

• Open a new terminal window (other than the one running TorchServe).
• Use curl to download one of these cute pictures of a kitten and use the -o flag to name it kitten.jpg for you.
• Use curl to send POST to the TorchServe predict endpoint with the kitten's image.

The following code completes all three steps:

curl -O https://s3.amazonaws.com/model-server/inputs/kitten.jpg
curl http://127.0.0.1:8080/predictions/densenet161 -T kitten.jpg

The predict endpoint returns a prediction response in JSON. It will look something like the following result:

[
  {
    "tiger_cat": 0.46933549642562866
  },
  {
    "tabby": 0.4633878469467163
  },
  {
    "Egyptian_cat": 0.06456148624420166
  },
  {
    "lynx": 0.0012828214094042778
  },
  {
    "plastic_bag": 0.00023323034110944718
  }
]

You will see this result in the response to your curl call to the predict endpoint, and in the server logs in the terminal window running TorchServe. It's also being logged locally with metrics.

Now you've seen how easy it can be to serve a deep learning model with TorchServe! Would you like to know more?

To stop the currently running TorchServe instance, run the following command:

torchserve --stop

You see output specifying that TorchServe has stopped.

TorchServe exposes configurations that allow the user to configure the number of worker threads on CPU and GPUs. There is an important config property that can speed up the server depending on the workload. Note: the following property has bigger impact under heavy workloads. If TorchServe is hosted on a machine with GPUs, there is a config property called number_of_gpu that tells the server to use a specific number of GPUs per model. In cases where we register multiple models with the server, this will apply to all the models registered. If this is set to a low value (ex: 0 or 1), it will result in under-utilization of GPUs. On the contrary, setting to a high value (>= max GPUs available on the system) results in as many workers getting spawned per model. Clearly, this will result in unnecessary contention for GPUs and can result in sub-optimal scheduling of threads to GPU.

ValueToSet = (Number of Hardware GPUs) / (Number of Unique Models)

Refer torchserve docker for details.

• Full documentation on TorchServe
• Manage models API
• Inference API
• Package models for use with TorchServe
• TorchServe model zoo for pre-trained and pre-packaged models-archives

We welcome all contributions!

To learn more about how to contribute, see the contributor guide here.

To file a bug or request a feature, please file a GitHub issue. For filing pull requests, please use the template here. Cheers!

This repository is jointly operated and maintained by Amazon, Facebook and a number of individual contributors listed in the CONTRIBUTORS file. For questions directed at Facebook, please send an email to [email protected]. For questions directed at Amazon, please send an email to [email protected]. For all other questions, please open up an issue in this repository here.

TorchServe acknowledges the Multi Model Server (MMS) project from which it was derived