Python Backend

The Triton backend for Python. The goal of Python backend is to let you serve models written in Python by Triton Inference Server without having to write any C++ code.

User Documentation

Python Backend
Business Logic Scripting
Interoperability and GPU Support
Frameworks
- PyTorch
  - PyTorch Determinism
- TensorFlow
  - TensorFlow Determinism
Custom Metrics
Examples
Running with Inferentia
Logging
Reporting problems, asking questions

Quick Start

Run the Triton Inference Server container.

docker run --shm-size=1g --ulimit memlock=-1 -p 8000:8000 -p 8001:8001 -p 8002:8002 --ulimit stack=67108864 -ti nvcr.io/nvidia/tritonserver:<xx.yy>-py3

Replace <xx.yy> with the Triton version (e.g. 21.05).

Inside the container, clone the Python backend repository.

git clone https://github.com/triton-inference-server/python_backend -b r<xx.yy>

Install example model.

cd python_backend
mkdir -p models/add_sub/1/
cp examples/add_sub/model.py models/add_sub/1/model.py
cp examples/add_sub/config.pbtxt models/add_sub/config.pbtxt

Start the Triton server.

tritonserver --model-repository `pwd`/models

In the host machine, start the client container.

docker run -ti --net host nvcr.io/nvidia/tritonserver:<xx.yy>-py3-sdk /bin/bash

In the client container, clone the Python backend repository.

git clone https://github.com/triton-inference-server/python_backend -b r<xx.yy>

Run the example client.

python3 python_backend/examples/add_sub/client.py

Building from Source

Requirements

cmake >= 3.17
numpy
rapidjson-dev
libarchive-dev
zlib1g-dev

pip3 install numpy

On Ubuntu or Debian you can use the command below to install rapidjson, libarchive, and zlib:

sudo apt-get install rapidjson-dev libarchive-dev zlib1g-dev

Build Python backend. Replace <GIT_BRANCH_NAME> with the GitHub branch that you want to compile. For release branches it should be r<xx.yy> (e.g. r21.06).

mkdir build
cd build
cmake -DTRITON_ENABLE_GPU=ON -DTRITON_BACKEND_REPO_TAG=<GIT_BRANCH_NAME> -DTRITON_COMMON_REPO_TAG=<GIT_BRANCH_NAME> -DTRITON_CORE_REPO_TAG=<GIT_BRANCH_NAME> -DCMAKE_INSTALL_PREFIX:PATH=`pwd`/install ..
make install

The following required Triton repositories will be pulled and used in the build. If the CMake variables below are not specified, "main" branch of those repositories will be used. <GIT_BRANCH_NAME> should be the same as the Python backend repository branch that you are trying to compile.

triton-inference-server/backend: -DTRITON_BACKEND_REPO_TAG=<GIT_BRANCH_NAME>
triton-inference-server/common: -DTRITON_COMMON_REPO_TAG=<GIT_BRANCH_NAME>
triton-inference-server/core: -DTRITON_CORE_REPO_TAG=<GIT_BRANCH_NAME>

Set -DCMAKE_INSTALL_PREFIX to the location where the Triton Server is installed. In the released containers, this location is /opt/tritonserver.

Copy example model and configuration

mkdir -p models/add_sub/1/
cp examples/add_sub/model.py models/add_sub/1/model.py
cp examples/add_sub/config.pbtxt models/add_sub/config.pbtxt

Start the Triton Server

/opt/tritonserver/bin/tritonserver --model-repository=`pwd`/models

Use the client app to perform inference

python3 examples/add_sub/client.py

Usage

In order to use the Python backend, you need to create a Python file that has a structure similar to below:

import triton_python_backend_utils as pb_utils


class TritonPythonModel:
    """Your Python model must use the same class name. Every Python model
    that is created must have "TritonPythonModel" as the class name.
    """

    @staticmethod
    def auto_complete_config(auto_complete_model_config):
        """`auto_complete_config` is called only once when loading the model
        assuming the server was not started with
        `--disable-auto-complete-config`. Implementing this function is
        optional. No implementation of `auto_complete_config` will do nothing.
        This function can be used to set `max_batch_size`, `input` and `output`
        properties of the model using `set_max_batch_size`, `add_input`, and
        `add_output`. These properties will allow Triton to load the model with
        minimal model configuration in absence of a configuration file. This
        function returns the `pb_utils.ModelConfig` object with these
        properties. You can use the `as_dict` function to gain read-only access
        to the `pb_utils.ModelConfig` object. The `pb_utils.ModelConfig` object
        being returned from here will be used as the final configuration for
        the model.

        Note: The Python interpreter used to invoke this function will be
        destroyed upon returning from this function and as a result none of the
        objects created here will be available in the `initialize`, `execute`,
        or `finalize` functions.

        Parameters
        ----------
        auto_complete_model_config : pb_utils.ModelConfig
          An object containing the existing model configuration. You can build
          upon the configuration given by this object when setting the
          properties for this model.

        Returns
        -------
        pb_utils.ModelConfig
          An object containing the auto-completed model configuration
        """
        inputs = [{
            'name': 'INPUT0',
            'data_type': 'TYPE_FP32',
            'dims': [4]
        }, {
            'name': 'INPUT1',
            'data_type': 'TYPE_FP32',
            'dims': [4]
        }]
        outputs = [{
            'name': 'OUTPUT0',
            'data_type': 'TYPE_FP32',
            'dims': [4]
        }, {
            'name': 'OUTPUT1',
            'data_type': 'TYPE_FP32',
            'dims': [4]
        }]

        # Demonstrate the usage of `as_dict`, `add_input`, `add_output`,
        # `set_max_batch_size`, and `set_dynamic_batching` functions.
        # Store the model configuration as a dictionary.
        config = auto_complete_model_config.as_dict()
        input_names = []
        output_names = []
        for input in config['input']:
            input_names.append(input['name'])
        for output in config['output']:
            output_names.append(output['name'])

        for input in inputs:
            # The name checking here is only for demonstrating the usage of
            # `as_dict` function. `add_input` will check for conflicts and
            # raise errors if an input with the same name already exists in
            # the configuration but has different data_type or dims property.
            if input['name'] not in input_names:
                auto_complete_model_config.add_input(input)
        for output in outputs:
            # The name checking here is only for demonstrating the usage of
            # `as_dict` function. `add_output` will check for conflicts and
            # raise errors if an output with the same name already exists in
            # the configuration but has different data_type or dims property.
            if output['name'] not in output_names:
                auto_complete_model_config.add_output(output)

        auto_complete_model_config.set_max_batch_size(0)

        # To enable a dynamic batcher with default settings, you can use
        # auto_complete_model_config set_dynamic_batching() function. It is
        # commented in this example because the max_batch_size is zero.
        #
        # auto_complete_model_config.set_dynamic_batching()

        return auto_complete_model_config

    def initialize(self, args):
        """`initialize` is called only once when the model is being loaded.
        Implementing `initialize` function is optional. This function allows
        the model to initialize any state associated with this model.

        Parameters
        ----------
        args : dict
          Both keys and values are strings. The dictionary keys and values are:
          * model_config: A JSON string containing the model configuration
          * model_instance_kind: A string containing model instance kind
          * model_instance_device_id: A string containing model instance device
            ID
          * model_repository: Model repository path
          * model_version: Model version
          * model_name: Model name
        """
        print('Initialized...')

    def execute(self, requests):
        """`execute` must be implemented in every Python model. `execute`
        function receives a list of pb_utils.InferenceRequest as the only
        argument. This function is called when an inference is requested
        for this model.

        Parameters
        ----------
        requests : list
          A list of pb_utils.InferenceRequest

        Returns
        -------
        list
          A list of pb_utils.InferenceResponse. The length of this list must
          be the same as `requests`
        """

        responses = []

        # Every Python backend must iterate through list of requests and create
        # an instance of pb_utils.InferenceResponse class for each of them.
        # Reusing the same pb_utils.InferenceResponse object for multiple
        # requests may result in segmentation faults. You should avoid storing
        # any of the input Tensors in the class attributes as they will be
        # overridden in subsequent inference requests. You can make a copy of
        # the underlying NumPy array and store it if it is required.
        for request in requests:
            # Perform inference on the request and append it to responses
            # list...

        # You must return a list of pb_utils.InferenceResponse. Length
        # of this list must match the length of `requests` list.
        return responses

    def finalize(self):
        """`finalize` is called only once when the model is being unloaded.
        Implementing `finalize` function is optional. This function allows
        the model to perform any necessary clean ups before exit.
        """
        print('Cleaning up...')

Every Python backend can implement four main functions:

`auto_complete_config`

auto_complete_config is called only once when loading the model assuming the server was not started with --disable-auto-complete-config.

Implementing this function is optional. No implementation of auto_complete_config will do nothing. This function can be used to set max_batch_size, dynamic_batching, input and output properties of the model using set_max_batch_size, set_dynamic_batching, add_input, and add_output. These properties will allow Triton to load the model with minimal model configuration in absence of a configuration file. This function returns the pb_utils.ModelConfig object with these properties. You can use the as_dict function to gain read-only access to the pb_utils.ModelConfig object. The pb_utils.ModelConfig object being returned from here will be used as the final configuration for the model.

Note: The Python interpreter used to invoke this function will be destroyed upon returning from this function and as a result none of the objects created here will be available in the initialize, execute, or finalize functions.

`initialize`

initialize is called once the model is being loaded. Implementing initialize is optional. initialize allows you to do any necessary initializations before execution. In the initialize function, you are given an args variable. args is a Python dictionary. Both keys and values for this Python dictionary are strings. You can find the available keys in the args dictionary along with their description in the table below:

key	description
model_config	A JSON string containing the model configuration
model_instance_kind	A string containing model instance kind
model_instance_device_id	A string containing model instance device ID
model_repository	Model repository path
model_version	Model version
model_name	Model name

`execute`

execute function is called whenever an inference request is made. Every Python model must implement execute function. In the execute function you are given a list of InferenceRequest objects. There are two modes of implementing this function. The mode you choose should depend on your use case. That is whether or not you want to return decoupled responses from this model or not.

Default Mode

This is the most generic way you would like to implement your model and requires the execute function to return exactly one response per request. This entails that in this mode, your execute function must return a list of InferenceResponse objects that has the same length as requests. The work flow in this mode is:

execute function receives a batch of pb_utils.InferenceRequest as a length N array.
Perform inference on the pb_utils.InferenceRequest and append the corresponding pb_utils.InferenceResponse to a response list.
Return back the response list.
- The length of response list being returned must be N.
- Each element in the list should be the response for the corresponding element in the request array.
- Each element must contain a response (a response can be either output tensors or an error); an element cannot be None.

Triton checks to ensure that these requirements on response list are satisfied and if not returns an error response for all inference requests. Upon return from the execute function all tensor data associated with the InferenceRequest objects passed to the function are deleted, and so InferenceRequest objects should not be retained by the Python model.

Error Handling

In case one of the requests has an error, you can use the TritonError object to set the error message for that specific request. Below is an example of setting errors for an InferenceResponse object:

import triton_python_backend_utils as pb_utils


class TritonPythonModel:
    ...

    def execute(self, requests):
        responses = []

        for request in requests:
            if an_error_occurred:
              # If there is an error, there is no need to pass the
              # "output_tensors" to the InferenceResponse. The "output_tensors"
              # that are passed in this case will be ignored.
              responses.append(pb_utils.InferenceResponse(
                error=pb_utils.TritonError("An Error Occurred")))

        return responses

Decoupled mode

This mode allows user to send multiple responses for a request or not send any responses for a request. A model may also send responses out-of-order relative to the order that the request batches are executed. Such models are called decoupled models. In order to use this mode, the transaction policy in the model configuration must be set to decoupled.

In decoupled mode, model must use InferenceResponseSender object per request to keep creating and sending any number of responses for the request. The workflow in this mode may look like:

execute function receives a batch of pb_utils.InferenceRequest as a length N array.
Iterate through each pb_utils.InferenceRequest and perform for the following steps for each pb_utils.InferenceRequest object:
1. Get InferenceResponseSender object for the InferenceRequest using InferenceRequest.get_response_sender().
2. Create and populate pb_utils.InferenceResponse to be sent back.
3. Use InferenceResponseSender.send() to send the above response. If this is the last request then pass pb_utils.TRITONSERVER_RESPONSE_COMPLETE_FINAL as a flag with InferenceResponseSender.send(). Otherwise continue with Step 1 for sending next request.
The return value for execute function in this mode should be None.

Similar to above, in case one of the requests has an error, you can use the TritonError object to set the error message for that specific request. After setting errors for an pb_utils.InferenceResponse object, use InferenceResponseSender.send() to send response with the error back to the user.

Use Cases

The decoupled mode is powerful and supports various other use cases:

If the model should not send any response for the request, then call InferenceResponseSender.send() with no response but flag parameter set to pb_utils.TRITONSERVER_RESPONSE_COMPLETE_FINAL.
The model can also send responses out-of-order in which it received requests.
The request data and InferenceResponseSender object can be passed to a separate thread in the model. This means main caller thread can exit from execute function and the model can still continue generating responses as long as it holds InferenceResponseSender object.

The decoupled examples demonstrate full power of what can be achieved from decoupled API. Read Decoupled Backends and Models for more details on how to host a decoupled model.

Known Issues

Currently, async BLS requests are not supported in decoupled mode.

`finalize`

Implementing finalize is optional. This function allows you to do any clean ups necessary before the model is unloaded from Triton server.

You can look at the add_sub example which contains a complete example of implementing all these functions for a Python model that adds and subtracts the inputs given to it. After implementing all the necessary functions, you should save this file as model.py.

Model Config File

Every Python Triton model must provide a config.pbtxt file describing the model configuration. In order to use this backend you must set the backend field of your model config.pbtxt file to python. You shouldn't set platform field of the configuration.

Your models directory should look like below:

models
└── add_sub
    ├── 1
    │   └── model.py
    └── config.pbtxt

Inference Request Parameters

You can retrieve the parameters associated with an inference request using the inference_request.parameters() function. This function returns a JSON string where the keys are the keys of the parameters object and the values are the values for the parameters field. Note that you need to parse this string using json.loads to convert it to a dictionary.

You can read more about the inference request parameters in the parameters extension documentation.

Managing Python Runtime and Libraries

Python backend shipped in the NVIDIA GPU Cloud containers uses Python 3.10. Python backend is able to use the libraries that exist in the current Python environment. These libraries can be installed in a virtualenv, conda environment, or the global system Python. These libraries will only be used if the Python version matches the Python version of the Python backend's stub executable. For example, if you install a set of libraries in a Python 3.9 environment and your Python backend stub is compiled with Python 3.10 these libraries will NOT be available in your Python model served using Triton. You would need to compile the stub executable with Python 3.9 using the instructions in Building Custom Python Backend Stub section.

Building Custom Python Backend Stub

Important Note: You only need to compile a custom Python backend stub if the Python version is different from Python 3.10 which is shipped by default in the Triton containers.

Python backend uses a stub process to connect your model.py file to the Triton C++ core. This stub process has an embedded Python interpreter with a fixed Python version. If you intend to use a Python interpreter with different version from the default Python backend stub, you need to compile your own Python backend stub by following the steps below:

Install the software packages below:

cmake
rapidjson and libarchive (instructions for installing these packages in Ubuntu or Debian are included in Building from Source Section)

Make sure that the expected Python version is available in your environment.

If you are using conda, you should make sure to activate the environment by conda activate <conda-env-name>. Note that you don't have to use conda and can install Python however you wish. Python backend relies on pybind11 to find the correct Python version. If you noticed that the correct Python version is not picked up, you can read more on how pybind11 decides which Python to use.

Clone the Python backend repository and compile the Python backend stub (replace <GIT_BRANCH_NAME> with the branch name that you want to use, for release branches it should be r<xx.yy>):

git clone https://github.com/triton-inference-server/python_backend -b
<GIT_BRANCH_NAME>
cd python_backend
mkdir build && cd build
cmake -DTRITON_ENABLE_GPU=ON -DTRITON_BACKEND_REPO_TAG=<GIT_BRANCH_NAME> -DTRITON_COMMON_REPO_TAG=<GIT_BRANCH_NAME> -DTRITON_CORE_REPO_TAG=<GIT_BRANCH_NAME> -DCMAKE_INSTALL_PREFIX:PATH=`pwd`/install ..
make triton-python-backend-stub

Now, you have access to a Python backend stub with your Python version. You can verify that using ldd:

ldd triton_python_backend_stub
...
libpython3.6m.so.1.0 => /home/ubuntu/envs/miniconda3/envs/python-3-6/lib/libpython3.6m.so.1.0 (0x00007fbb69cf3000)
...

There are many other shared libraries printed in addition to the library posted above. However, it is important to see libpython<major>.<minor>m.so.1.0 in the list of linked shared libraries. If you use a different Python version, you should see that version instead. You need to copy the triton_python_backend_stub to the model directory of the models that want to use the custom Python backend stub. For example, if you have model_a in your model repository, the folder structure should look like below:

models
|-- model_a
    |-- 1
    |   |-- model.py
    |-- config.pbtxt
    `-- triton_python_backend_stub

Note the location of triton_python_backend_stub in the directory structure above.

Creating Custom Execution Environments

If you want to create a tar file that contains all your Python dependencies or you want to use different Python environments for each Python model you need to create a Custom Execution Environment in Python backend. Currently, Python backend supports conda-pack for this purpose. conda-pack ensures that your conda environment is portable. You can create a tar file for your conda environment using conda-pack command:

conda-pack
Collecting packages...
Packing environment at '/home/iman/miniconda3/envs/python-3-6' to 'python-3-6.tar.gz'
[########################################] | 100% Completed |  4.5s

Important Note: Before installing the packages in your conda environment, make sure that you have exported PYTHONNOUSERSITE environment variable:

export PYTHONNOUSERSITE=True

If this variable is not exported and similar packages are installed outside your conda environment, your tar file may not contain all the dependencies required for an isolated Python environment.

Alternatively, Python backend also supports unpacked conda execution environments, given it points to an activation script to setup the conda environment. To do this, the execution environment can be first packed using conda-pack and then unpacked, or created using conda create -p. In this case, the conda activation script is located in: $path_to_conda_pack/lib/python<your.python.version>/site-packages/conda_pack/scripts/posix/activate This speeds up the server loading time for models.

After creating the packed file from the conda environment or creating a conda environment with a custom activation script, you need to tell Python backend to use that environment for your model. You can do this by adding the lines below to the config.pbtxt file:

name: "model_a"
backend: "python"

...

parameters: {
  key: "EXECUTION_ENV_PATH",
  value: {string_value: "/home/iman/miniconda3/envs/python-3-6/python3.6.tar.gz"}
}

It is also possible to provide the execution environment path relative to the model folder in model repository:

name: "model_a"
backend: "python"

...

parameters: {
  key: "EXECUTION_ENV_PATH",
  value: {string_value: "$$TRITON_MODEL_DIRECTORY/python3.6.tar.gz"}
}

In this case, python3.tar.gz should be placed in the model folder and the model repository should look like below:

models
|-- model_a
|   |-- 1
|   |   `-- model.py
|   |-- config.pbtxt
|   |-- python3.6.tar.gz
|   `-- triton_python_backend_stub

In the example above, $$TRITON_MODEL_DIRECTORY is resolved to $pwd/models/model_a.

This is useful if you want to use S3, GCS, or Azure and you do not have access to the absolute path of the execution env that is stored in the cloud object storage service.

Important Notes

The version of the Python interpreter in the execution environment must match the version of triton_python_backend_stub.
If you don't want to use a different Python interpreter, you can skip Building Custom Python Backend Stub. In this case you only need to pack your environment using conda-pack and provide the path to tar file in the model config. However, the previous note still applies here and the version of the Python interpreter inside the conda environment must match the Python version of stub used by Python backend. The default version of the stub is Python 3.10.
You can share a single execution environment across multiple models. You need to provide the path to the tar file in the EXECUTION_ENV_PATH in the config.pbtxt of all the models that want to use the execution environment.
If $$TRITON_MODEL_DIRECTORY is used in the EXECUTION_ENV_PATH, the final EXECUTION_ENV_PATH must not escape from the $$TRITON_MODEL_DIRECTORY, as the behavior of accessing anywhere outside the $$TRITON_MODEL_DIRECTORY is undefined.
If a non-$$TRITON_MODEL_DIRECTORY EXECUTION_ENV_PATH is used, only local file system paths are currently supported. The behavior of using cloud paths is undefined.
If you need to compile the Python backend stub, it is recommended that you compile it in the official Triton NGC containers. Otherwise, your compiled stub may use dependencies that are not available in the Triton container that you are using for deployment. For example, compiling the Python backend stub on an OS other than Ubuntu 22.04 can lead to unexpected errors.

Error Handling

If there is an error that affects the initialize, execute, or finalize function of the Python model you can use TritonInferenceException. Example below shows how you can do error handling in finalize:

import triton_python_backend_utils as pb_utils


class TritonPythonModel:
    ...

    def finalize(self):
      if error_during_finalize:
        raise pb_utils.TritonModelException(
          "An error occurred during finalize.")

Managing Shared Memory

Starting from 21.04 release, Python backend uses shared memory to connect user's code to Triton. Note that this change is completely transparent and does not require any change to the existing user's model code.

Python backend, by default, allocates 64 MBs for each model instance. Then, it will grow the shared memory region by 64 MBs whenever an increase is required. You can configure the default shared memory used by each model instance using the shm-default-byte-size flag. The amount of shared memory growth can be configured using the shm-growth-byte-size.

You can also configure the timeout used for connecting Triton main process to the Python backend stubs using the stub-timeout-seconds. The default value is 30 seconds.

The config values described above can be passed to Triton using --backend-config flag:

/opt/tritonserver/bin/tritonserver --model-repository=`pwd`/models --backend-config=python,<config-key>=<config-value>

Also, if you are running Triton inside a Docker container you need to properly set the --shm-size flag depending on the size of your inputs and outputs. The default value for docker run command is 64MB which is very small.

Multiple Model Instance Support

Python interpreter uses a global lock known as GIL. Because of GIL, it is not possible have multiple threads running in the same Python interpreter simultaneously as each thread requires to acquire the GIL when accessing Python objects which will serialize all the operations. In order to work around this issue, Python backend spawns a separate process for each model instance. This is in contrast with how other Triton backends such as ONNXRuntime, TensorFlow, and PyTorch handle multiple instances. Increasing the instance count for these backends will create additional threads instead of spawning separate processes.

Running Multiple Instances of Triton Server

Python backend uses shared memory to transfer requests to the stub process. When running multiple instances of Triton Server on the same machine that use Python models, there would be shared memory region name conflicts that can result in segmentation faults or hangs. In order to avoid this issue, you need to specify different shm-region-prefix-name using the --backend-config flag.

# Triton instance 1
tritonserver --model-repository=/models --backend-config=python,shm-region-prefix-name=prefix1

# Triton instance 2
tritonserver --model-repository=/models --backend-config=python,shm-region-prefix-name=prefix2

Note that the hangs would only occur if the /dev/shm is shared between the two instances of the server. If you run the servers in different containers that don't share this location, you don't need to specify shm-region-prefix-name.

Business Logic Scripting

Triton's ensemble feature supports many use cases where multiple models are composed into a pipeline (or more generally a DAG, directed acyclic graph). However, there are many other use cases that are not supported because as part of the model pipeline they require loops, conditionals (if-then-else), data-dependent control-flow and other custom logic to be intermixed with model execution. We call this combination of custom logic and model executions Business Logic Scripting (BLS).

Starting from 21.08, you can implement BLS in your Python model. A new set of utility functions allows you to execute inference requests on other models being served by Triton as a part of executing your Python model. Note that BLS should only be used inside the execute function and is not supported in the initialize or finalize methods. Example below shows how to use this feature:

import triton_python_backend_utils as pb_utils


class TritonPythonModel:
  ...
    def execute(self, requests):
      ...
      # Create an InferenceRequest object. `model_name`,
      # `requested_output_names`, and `inputs` are the required arguments and
      # must be provided when constructing an InferenceRequest object. Make
      # sure to replace `inputs` argument with a list of `pb_utils.Tensor`
      # objects.
      inference_request = pb_utils.InferenceRequest(
          model_name='model_name',
          requested_output_names=['REQUESTED_OUTPUT_1', 'REQUESTED_OUTPUT_2'],
          inputs=[<pb_utils.Tensor object>])

      # `pb_utils.InferenceRequest` supports request_id, correlation_id,
      # model version, timeout and preferred_memory in addition to the
      # arguments described above.
      # These arguments are optional. An example containing all the arguments:
      # inference_request = pb_utils.InferenceRequest(model_name='model_name',
      #   requested_output_names=['REQUESTED_OUTPUT_1', 'REQUESTED_OUTPUT_2'],
      #   inputs=[<list of pb_utils.Tensor objects>],
      #   request_id="1", correlation_id=4, model_version=1, flags=0, timeout=5,
      #   preferred_memory=pb_utils.PreferredMemory(
      #     pb_utils.TRITONSERVER_MEMORY_GPU, # or pb_utils.TRITONSERVER_MEMORY_CPU
      #     0))

      # Execute the inference_request and wait for the response
      inference_response = inference_request.exec()

      # Check if the inference response has an error
      if inference_response.has_error():
          raise pb_utils.TritonModelException(
            inference_response.error().message())
      else:
          # Extract the output tensors from the inference response.
          output1 = pb_utils.get_output_tensor_by_name(
            inference_response, 'REQUESTED_OUTPUT_1')
          output2 = pb_utils.get_output_tensor_by_name(
            inference_response, 'REQUESTED_OUTPUT_2')

          # Decide the next steps for model execution based on the received
          # output tensors. It is possible to use the same output tensors
          # to for the final inference response too.

In addition to the inference_request.exec function that allows you to execute blocking inference requests, inference_request.async_exec allows you to perform async inference requests. This can be useful when you do not need the result of the inference immediately. Using async_exec function, it is possible to have multiple inflight inference requests and wait for the responses only when needed. Example below shows how to use async_exec:

import triton_python_backend_utils as pb_utils
import asyncio


class TritonPythonModel:
  ...

    # You must add the Python 'async' keyword to the beginning of `execute`
    # function if you want to use `async_exec` function.
    async def execute(self, requests):
      ...
      # Create an InferenceRequest object. `model_name`,
      # `requested_output_names`, and `inputs` are the required arguments and
      # must be provided when constructing an InferenceRequest object. Make
      # sure to replace `inputs` argument with a list of `pb_utils.Tensor`
      # objects.
      inference_request = pb_utils.InferenceRequest(
          model_name='model_name',
          requested_output_names=['REQUESTED_OUTPUT_1', 'REQUESTED_OUTPUT_2'],
          inputs=[<pb_utils.Tensor object>])

      infer_response_awaits = []
      for i in range(4):
        # async_exec function returns an
        # [Awaitable](https://docs.python.org/3/library/asyncio-task.html#awaitables)
        # object.
        infer_response_awaits.append(inference_request.async_exec())

      # Wait for all of the inference requests to complete.
      infer_responses = await asyncio.gather(*infer_response_awaits)

      for infer_response in infer_responses:
        # Check if the inference response has an error
        if inference_response.has_error():
            raise pb_utils.TritonModelException(
              inference_response.error().message())
        else:
            # Extract the output tensors from the inference response.
            output1 = pb_utils.get_output_tensor_by_name(
              inference_response, 'REQUESTED_OUTPUT_1')
            output2 = pb_utils.get_output_tensor_by_name(
              inference_response, 'REQUESTED_OUTPUT_2')

            # Decide the next steps for model execution based on the received
            # output tensors.

A complete example for sync and async BLS in Python backend is included in the Examples section.

Using BLS with Decoupled Models

Starting from 23.03 release, you can execute inference requests on decoupled models in both default mode and decoupled mode. By setting the decoupled parameter to True, the exec and async_exec function will return an iterator of inference responses returned by a decoupled model. If the decoupled parameter is set to False, the exec and async_exec function will return a single response as shown in the example above. Besides, you can set the timeout via the parameter 'timeout' in microseconds within the constructor of InferenceRequest. If the request times out, the request will respond with an error. The default of 'timeout' is 0 which indicates that the request has no timeout.

Additionally, starting from the 23.04 release, you have the flexibility to select a specific device to receive output tensors from BLS calls. This can be achieved by setting the optional preferred_memory parameter within the InferenceRequest constructor. To do this, you can create a PreferredMemory object and specify the preferred_memory_type as either TRITONSERVER_MEMORY_GPU or TRITONSERVER_MEMORY_CPU, as well as the preferred_device_id as an integer to indicate the memory type and device ID on which you wish to receive output tensors. If you do not specify the preferred_memory parameter, the output tensors will be allocated on the same device where the output tensors were received from the model to which the BLS call is made.

Example below shows how to use this feature:

import triton_python_backend_utils as pb_utils


class TritonPythonModel:
  ...
    def execute(self, requests):
      ...
      # Create an InferenceRequest object. `model_name`,
      # `requested_output_names`, and `inputs` are the required arguments and
      # must be provided when constructing an InferenceRequest object. Make
      # sure to replace `inputs` argument with a list of `pb_utils.Tensor`
      # objects.
      inference_request = pb_utils.InferenceRequest(
          model_name='model_name',
          requested_output_names=['REQUESTED_OUTPUT_1', 'REQUESTED_OUTPUT_2'],
          inputs=[<pb_utils.Tensor object>])

      # `pb_utils.InferenceRequest` supports request_id, correlation_id,
      # model version, timeout and preferred_memory in addition to the
      # arguments described above.
      # These arguments are optional. An example containing all the arguments:
      # inference_request = pb_utils.InferenceRequest(model_name='model_name',
      #   requested_output_names=['REQUESTED_OUTPUT_1', 'REQUESTED_OUTPUT_2'],
      #   inputs=[<list of pb_utils.Tensor objects>],
      #   request_id="1", correlation_id=4, model_version=1, flags=0, timeout=5,
      #   preferred_memory=pb_utils.PreferredMemory(
      #     pb_utils.TRITONSERVER_MEMORY_GPU, # or pb_utils.TRITONSERVER_MEMORY_CPU
      #     0))

      # Execute the inference_request and wait for the response. Here we are
      # running a BLS request on a decoupled model, hence setting the parameter
      # 'decoupled' to 'True'.
      inference_responses = inference_request.exec(decoupled=True)

      for inference_response in inference_responses:
        # Check if the inference response has an error
        if inference_response.has_error():
            raise pb_utils.TritonModelException(
              inference_response.error().message())

        # For some models, it is possible that the last response is empty
        if len(infer_response.output_tensors()) > 0:
          # Extract the output tensors from the inference response.
          output1 = pb_utils.get_output_tensor_by_name(
            inference_response, 'REQUESTED_OUTPUT_1')
          output2 = pb_utils.get_output_tensor_by_name(
            inference_response, 'REQUESTED_OUTPUT_2')

          # Decide the next steps for model execution based on the received
          # output tensors. It is possible to use the same output tensors to
          # for the final inference response too.

In addition to the inference_request.exec(decoupled=True) function that allows you to execute blocking inference requests on decoupled models, inference_request.async_exec(decoupled=True) allows you to perform async inference requests. This can be useful when you do not need the result of the inference immediately. Using async_exec function, it is possible to have multiple inflight inference requests and wait for the responses only when needed. Example below shows how to use async_exec:

import triton_python_backend_utils as pb_utils
import asyncio


class TritonPythonModel:
  ...

    # You must add the Python 'async' keyword to the beginning of `execute`
    # function if you want to use `async_exec` function.
    async def execute(self, requests):
      ...
      # Create an InferenceRequest object. `model_name`,
      # `requested_output_names`, and `inputs` are the required arguments and
      # must be provided when constructing an InferenceRequest object. Make
      # sure to replace `inputs` argument with a list of `pb_utils.Tensor`
      # objects.
      inference_request = pb_utils.InferenceRequest(
          model_name='model_name',
          requested_output_names=['REQUESTED_OUTPUT_1', 'REQUESTED_OUTPUT_2'],
          inputs=[<pb_utils.Tensor object>])

      infer_response_awaits = []
      for i in range(4):
        # async_exec function returns an
        # [Awaitable](https://docs.python.org/3/library/asyncio-task.html#awaitables)
        # object.
        infer_response_awaits.append(
          inference_request.async_exec(decoupled=True))

      # Wait for all of the inference requests to complete.
      async_responses = await asyncio.gather(*infer_response_awaits)

      for infer_responses in async_responses:
        for infer_response in infer_responses:
          # Check if the inference response has an error
          if inference_response.has_error():
              raise pb_utils.TritonModelException(
                inference_response.error().message())

          # For some models, it is possible that the last response is empty
          if len(infer_response.output_tensors()) > 0:
              # Extract the output tensors from the inference response.
              output1 = pb_utils.get_output_tensor_by_name(
                inference_response, 'REQUESTED_OUTPUT_1')
              output2 = pb_utils.get_output_tensor_by_name(
                inference_response, 'REQUESTED_OUTPUT_2')

              # Decide the next steps for model execution based on the received
              # output tensors.

A complete example for sync and async BLS for decoupled models is included in the Examples section.

Starting from the 22.04 release, the lifetime of the BLS output tensors have been improved such that if a tensor is no longer needed in your Python model it will be automatically deallocated. This can increase the number of BLS requests that you can execute in your model without running into the out of GPU or shared memory error.

Note: Async BLS is not supported on Python 3.6 or lower due to the async keyword and asyncio.run being introduced in Python 3.7.

Model Loading API

Starting from 23.07 release, you can use the model loading API to load models required by your BLS model. The model loading API is equivalent to the Triton C API for loading models which are documented in tritonserver.h. Below is an example of how to use the model loading API:

import triton_python_backend_utils as pb_utils

class TritonPythonModel:
    def initialize(self, args):
        self.model_name="onnx_model"
        # Check if the model is ready, and load the model if it is not ready.
        # You can specify the model version in string format. The version is
        # optional, and if not provided, the server will choose a version based
        # on the model and internal policy.
        if not pb_utils.is_model_ready(model_name=self.model_name,
                                       model_version="1"):
            # Load the model from the model repository
            pb_utils.load_model(model_name=self.model_name)

            # Load the model with an optional override model config in JSON
            # representation. If provided, this config will be used for
            # loading the model.
            config = "{\"backend\":\"onnxruntime\", \"version_policy\":{\"specific\":{\"versions\":[1]}}}"
            pb_utils.load_model(model_name=self.model_name, config=config)

            # Load the mode with optional override files. The override files are
            # specified as a dictionary where the key is the file path (with
            # "file:" prefix) and the value is the file content as bytes. The
            # files will form the model directory that the model will be loaded
            # from. If specified, 'config' must be provided to be the model
            # configuration of the override model directory.
            with open('models/onnx_int32_int32_int32/1/model.onnx', 'rb') as file:
                data = file.read()
            files = {"file:1/model.onnx": data}
            pb_utils.load_model(model_name=self.model_name,
                                config=config, files=files)

    def execute(self, requests):
        # Execute the model
        ...
        # If the model is no longer needed, you can unload it. You can also
        # specify whether the dependents of the model should also be unloaded by
        # setting the 'unload_dependents' parameter to True. The default value
        # is False. Need to be careful when unloading the model as it can affect
        # other model instances or other models that depend on it.
        pb_utils.unload_model(model_name=self.model_name,
                              unload_dependents=True)

Note that the model loading API is only supported if the server is running in explicit model control mode. Additionally, the model loading API should only be used after the server has been running, which means that the BLS model should not be loaded during server startup. You can use different client endpoints to load the model after the server has been started. The model loading API is currently not supported during the auto_complete_config and finalize functions.

Using BLS with Stateful Models

Stateful models require setting additional flags in the inference request to indicate the start and end of a sequence. The flags argument in the pb_utils.InferenceRequest object can be used to indicate whether the request is the first or last request in the sequence. An example indicating that the request is starting the sequence:

inference_request = pb_utils.InferenceRequest(model_name='model_name',
  requested_output_names=['REQUESTED_OUTPUT_1', 'REQUESTED_OUTPUT_2'],
  inputs=[<list of pb_utils.Tensor objects>],
  request_id="1", correlation_id=4,
  flags=pb_utils.TRITONSERVER_REQUEST_FLAG_SEQUENCE_START)

For indicating the ending of the sequence you can use the pb_utils.TRITONSERVER_REQUEST_FLAG_SEQUENCE_END flag. If the request is both starting and ending a sequence at the same time (i.e. the sequence has only a single request), you can use the bitwise OR operator to enable both of the flags:

flags = pb_utils.TRITONSERVER_REQUEST_FLAG_SEQUENCE_START | pb_utils.TRITONSERVER_REQUEST_FLAG_SEQUENCE_END

Limitation

You need to make sure that the inference requests performed as a part of your model do not create a circular dependency. For example, if model A performs an inference request on itself and there are no more model instances ready to execute the inference request, the model will block on the inference execution forever.
Async BLS is not supported when running a Python model in decoupled mode.

Interoperability and GPU Support

Starting from 21.09 release, Python backend supports DLPack for zero-copy transfer of Python backend tensors to other frameworks. The methods below are added to the pb_utils.Tensor object to facilitate the same:

`pb_utils.Tensor.to_dlpack() -> PyCapsule`

This method can be called on existing instantiated tensors to convert a Tensor to DLPack. The code snippet below shows how this works with PyTorch:

from torch.utils.dlpack import from_dlpack
import triton_python_backend_utils as pb_utils

class TritonPythonModel:

  def execute(self, requests):
    ...
    input0 = pb_utils.get_input_tensor_by_name(request, "INPUT0")

    # We have converted a Python backend tensor to a PyTorch tensor without
    # making any copies.
    pytorch_tensor = from_dlpack(input0.to_dlpack())

`pb_utils.Tensor.from_dlpack() -> Tensor`

This static method can be used for creating a Tensor object from the DLPack encoding of the tensor. For example:

from torch.utils.dlpack import to_dlpack
import torch
import triton_python_backend_utils as pb_utils

class TritonPythonModel:

  def execute(self, requests):
    ...
    pytorch_tensor = torch.tensor([1, 2, 3], device='cuda')

    # Create a Python backend tensor from the DLPack encoding of a PyTorch
    # tensor.
    input0 = pb_utils.Tensor.from_dlpack("INPUT0", to_dlpack(pytorch_tensor))

Python backend allows tensors implementing __dlpack__ and __dlpack_device__ interface to be converted to Python backend tensors. For instance:

input0 = pb_utils.Tensor.from_dlpack("INPUT0", pytorch_tensor)

This method only supports contiguous Tensors that are in C-order. If the tensor is not C-order contiguous an exception will be raised.

`pb_utils.Tensor.is_cpu() -> bool`

This function can be used to check whether a tensor is placed in CPU or not.

Input Tensor Device Placement

By default, the Python backend moves all input tensors to CPU before providing them to the Python model. Starting from 21.09, you can change this default behavior. By setting FORCE_CPU_ONLY_INPUT_TENSORS to "no", Triton will not move input tensors to CPU for the Python model. Instead, Triton will provide the input tensors to the Python model in either CPU or GPU memory, depending on how those tensors were last used. You cannot predict which memory will be used for each input tensor so your Python model must be able to handle tensors in both CPU and GPU memory. To enable this setting, you need to add this setting to the parameters section of model configuration:

parameters: { key: "FORCE_CPU_ONLY_INPUT_TENSORS" value: {string_value:"no"}}

Frameworks

Since Python Backend models can support most python packages, it is a common workflow for users to use Deep Learning Frameworks like PyTorch in their model.py implementation. This section will document some notes and FAQ about this workflow.

Note

Using a deep learning framework/package in a Python Backend model is not necessarily the same as using the corresponding Triton Backend implementation. For example, the PyTorch Backend is different from using a Python Backend model that uses import torch. If you are seeing significantly different results from a model executed by the framework (ex: PyTorch) compared to the Python Backend model running the same framework, some of the first things you should check is that the framework versions being used and the input/output preparation are the same.

PyTorch

For a simple example of using PyTorch in a Python Backend model, see the AddSubNet PyTorch example.

PyTorch Determinism

When running PyTorch code, you may notice slight differences in output values across runs or across servers depending on hardware, system load, driver, or even batch size. These differences are generally related to the selection of CUDA kernels used to execute the operations, based on the factors mentioned.

For most intents and purposes, these differences aren't large enough to affect a model's final prediction. However, to understand where these differences come from, see this doc.

On Ampere devices and later, there is an optimization related to FP32 operations called TensorFloat32 (TF32). Typically this optimization will improve overall performance at the cost of minor precision loss, but similarly this precision loss is acceptable for most model predictions. For more info on TF32 in PyTorch and how to enable/disable it as needed, see here.

TensorFlow

TensorFlow Determinism

Similar to the PyTorch determinism section above, TensorFlow can have slight differences in outputs based on various factors like hardware, system configurations, or batch sizes due to the library's internal CUDA kernel selection process. For more information on improving the determinism of outputs in TensorFlow, see here.

Custom Metrics

Starting from 23.05, you can utlize Custom Metrics API to register and collect custom metrics in the initialize, execute, and finalize functions of your Python model. The Custom Metrics API is the Python equivalent of the TRITON C API custom metrics support. You will need to take the ownership of the custom metrics created through the APIs and must manage their lifetime. Note that a MetricFamily object should be deleted only after all the Metric objects under it are deleted if you'd like to explicitly delete the custom metrics objects.

Example below shows how to use this feature:

import triton_python_backend_utils as pb_utils


class TritonPythonModel:
    def initialize(self, args):
      # Create a MetricFamily object to report the latency of the model
      # execution. The 'kind' parameter must be either 'COUNTER' or
      # 'GAUGE'.
      self.metric_family = pb_utils.MetricFamily(
          name="preprocess_latency_ns",
          description="Cumulative time spent pre-processing requests",
          kind=pb_utils.MetricFamily.COUNTER # or pb_utils.MetricFamily.GAUGE
      )

      # Create a Metric object under the MetricFamily object. The 'labels'
      # is a dictionary of key-value pairs.
      self.metric = self.metric_family.Metric(
        labels={"model" : "model_name", "version" : "1"}
      )

    def execute(self, requests):
      responses = []

      for request in requests:
        # Pre-processing - time it to capture latency
        start_ns = time.time_ns()
        self.preprocess(request)
        end_ns = time.time_ns()

        # Update metric to track cumulative pre-processing latency
        self.metric.increment(end_ns - start_ns)

      ...

        print("Cumulative pre-processing latency:", self.metric.value())

      return responses

You can look at the custom_metrics example which contains a complete example of demonstrating the Custom Metrics API for a Python model.

Examples

For using the Triton Python client in these examples you need to install the Triton Python Client Library. The Python client for each of the examples is in the client.py file.

AddSub in NumPy

There is no dependencies required for the AddSub NumPy example. Instructions on how to use this model is explained in the quick start section. You can find the files in examples/add_sub.

AddSubNet in PyTorch

In order to use this model, you need to install PyTorch. We recommend using pip method mentioned in the PyTorch website. Make sure that PyTorch is available in the same Python environment as other dependencies. Alternatively, you can create a Python Execution Environment. You can find the files for this example in examples/pytorch.

AddSub in JAX

The JAX example shows how to serve JAX in Triton using Python Backend. You can find the complete example instructions in examples/jax.

Business Logic Scripting

The BLS example needs the dependencies required for both of the above examples. You can find the complete example instructions in examples/bls and examples/bls_decoupled.

Preprocessing

The Preprocessing example shows how to use Python Backend to do model preprocessing. You can find the complete example instructions in examples/preprocessing.

Decoupled Models

The examples of decoupled models shows how to develop and serve decoupled models in Triton using Python backend. You can find the complete example instructions in examples/decoupled.

Model Instance Kind

Triton model configuration allows users to provide kind to instance group settings. A python backend model can be written to respect the kind setting to control the execution of a model instance either on CPU or GPU.

In the model instance kind example we demonstrate how this can be achieved for your python model.

Auto-complete config

The auto-complete config example demonstrates how to use the auto_complete_config function to define minimal model configuration when a configuration file is not available. You can find the complete example instructions in examples/auto_complete.

Custom Metrics

The example shows how to use custom metrics API in Python Backend. You can find the complete example instructions in examples/custom_metrics.

Running with Inferentia

Please see the README.md located in the python_backend/inferentia sub folder.

Logging

Starting from 22.09 release, your Python model can log information using the following methods:

import triton_python_backend_utils as pb_utils

class TritonPythonModel:

  def execute(self, requests):
    ...
    logger = pb_utils.Logger
    logger.log_info("Info Msg!")
    logger.log_warn("Warning Msg!")
    logger.log_error("Error Msg!")
    logger.log_verbose("Verbose Msg!")

Note: The logger can be defined and used in following class methods:

initialize
execute
finalize

Log messages can also be sent with their log-level explcitiy specified:

# log-level options: INFO, WARNING, ERROR, VERBOSE
logger.log("Specific Msg!", logger.INFO)

If no log-level is specified, this method will log INFO level messages.

Note that the Triton server's settings determine which log messages appear within the server log. For example, if a model attempts to log a verbose-level message, but Triton is not set to log verbose-level messages, it will not appear in the server log. For more information on Triton's log settings and how to adjust them dynamically, please see Triton's logging extension documentation.

Reporting problems, asking questions

We appreciate any feedback, questions or bug reporting regarding this project. When help with code is needed, follow the process outlined in the Stack Overflow (https://stackoverflow.com/help/mcve) document. Ensure posted examples are:

minimal – use as little code as possible that still produces the same problem
complete – provide all parts needed to reproduce the problem. Check if you can strip external dependency and still show the problem. The less time we spend on reproducing problems the more time we have to fix it
verifiable – test the code you're about to provide to make sure it reproduces the problem. Remove all other problems that are not related to your request/question.

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!