NOTE: dali_backend is available in tritonserver-20.11 and later

:exclamation: IMPORTANT :exclamation:

dali_backend is new and rapidly growing. Official tritonserver releases might be behind on some features and bug fixes. We encourage you to use the latest version of dali_backend. Docker build section explains, how to build a tritonserver docker image with main branch of dali_backend and DALI nightly release. This is a way to get daily updates!

DALI TRITON Backend

This repository contains code for DALI Backend for Triton Inference Server.

NVIDIA DALI (R), the Data Loading Library, is a collection of highly optimized building blocks, and an execution engine, to accelerate the pre-processing of the input data for deep learning applications. DALI provides both the performance and the flexibility to accelerate different data pipelines as one library. This library can then be easily integrated into different deep learning training and inference applications, regardless of used deep learning framework.

To find out more about DALI please refer to our main page. Getting started and Tutorials will guide you through your first steps and Supported operations will help you put together GPU-powered data processing pipelines.

See any bugs?

Feel free to post an issue here or in DALI's github repository.

How to use?

1. DALI data pipeline is expressed within Triton as a Model. To create such Model, you have to put together a DALI Pipeline in Python. Then, you have to serialize it (by calling the Pipeline.serialize method) or use the Autoserialization to generate a Model file. As an example, we'll use simple resizing pipeline:

    import nvidia.dali as dali
    from nvidia.dali.plugin.triton import autoserialize
   
    @autoserialize 
    @dali.pipeline_def(batch_size=256, num_threads=4, device_id=0)
    def pipe():
        images = dali.fn.external_source(device="cpu", name="DALI_INPUT_0")
        images = dali.fn.image_decoder(images, device="mixed")
        images = dali.fn.resize(images, resize_x=224, resize_y=224)
        return images
   

2. Model file shall be incorporated in Triton's Model Repository. Here's the example:

    model_repository
    └── dali
        ├── 1
        │   └── model.dali
        └── config.pbtxt
   

3. As it's typical in Triton, your DALI Model file shall be named model.dali. You can override this name in the model configuration, by setting default_model_filename option. Here's the whole config.pbtxt we use for the ResizePipeline example:

    name: "dali"
    backend: "dali"
    max_batch_size: 256
    input [
    {
        name: "DALI_INPUT_0"
        data_type: TYPE_UINT8
        dims: [ -1 ]
    }
    ]
   
    output [
    {
        name: "DALI_OUTPUT_0"
        data_type: TYPE_UINT8
        dims: [ 224, 224, 3 ]
    }
    ]
   

You can omit writing most of the configuration file if you specify information about the inputs, outputs and max batch size in the pipeline definition. Refer to Configuration auto-complete for the details about this feature.

Configuration auto-complete

To simplify the model deployment, Triton Server can infer parts of the configuration file from the model file itself. In case of DALI backend, the information about the inputs, outputs and the max batch size can be specified in the pipeline definition and does not need to be repeated in the configuration file. Below you can see how to include the configuration info in the Python pipeline definition:

import nvidia.dali as dali
from nvidia.dali.plugin.triton import autoserialize
import nvidia.dali.types as types

@autoserialize
@dali.pipeline_def(batch_size=256, num_threads=4, device_id=0, output_dtype=[types.UINT8], output_ndim=[3])
def pipe():
    images = dali.fn.external_source(device="cpu", name="DALI_INPUT_0", dtype=types.UINT8, ndim=1)
    images = dali.fn.image_decoder(images, device="mixed")
    images = dali.fn.resize(images, resize_x=224, resize_y=224)
    return images

As you can see, we added dtype and ndim (number of dimensions) arguments to the external source operator. They provide the information needed to fill the inputs field in the configuration file. To fill the outputs field, we added the output_dtype and output_ndim arguments to the pipeline definition. Those are expected to be lists with a value for each output.

This way we can limit the configuration file to just naming the model and specifying the DALI backend:

name: "dali"
backend: "dali"

Partial configuration

You can still provide some of the information if it is not present in the pipeline definition or to override some of the values. For example, you can use the configuration file to give new names to the model outputs which might be useful when using them later in an ensemble model. You can also overwrite the max batch size. The configuration file for the pipeline defined above could look like this:

name: "dali"
backend: "dali"
max_batch_size: 128

output [
{
    name: "DALI_OUTPUT_0"
    dims: [ 224, 224, 3 ]
}
]

Such configuration file overwrites the max batch size value to 128. It also renames the pipeline output to "DALI_OUTPUT_0" and specifies its shape to be (224, 224, 3).

Refer DALI model configuration file documentation for details on model parameters that can specified in the configuation file.

Autoserialization

When using DALI Backend in Triton, user has to provide a DALI model in the Model Repository. A canonical way of expressing a model is to include a serialized DALI model file there and naming the file properly (model.dali by default). The issue that arises from storing model in a serialized file is that, after serialization, the model is obscure and almost impossible to read anymore. Autoserialization feature allows user to express the model in Python code in the model repository. With the Python-defined model, DALI Backend uses internal serialization mechanism and exempts user from manual serialization.

To use the autoserialization feature, user needs to put a Python-definition of the DALI pipeline inside the model file (model.dali by default, but the default file name can be configured in the config.pbtxt). Such pipeline definition has to be decorated with @autoserialize, e.g.:

import nvidia.dali as dali

@dali.plugin.triton.autoserialize
@dali.pipeline_def(batch_size=3, num_threads=1, device_id=0)
def pipe():
    '''
    An identity pipeline with autoserialization enabled
    '''
    data = dali.fn.external_source(device="cpu", name="DALI_INPUT_0")
    return data

Proper DALI pipeline definition in Python, together with autoserialization, shall meet the following conditions:

1. Only a pipeline_def can be decorated with autoserialize.
2. Only one pipeline definition may be decorated with autoserialize in a given model version.

While loading a model file, DALI Backend follows the precedence:

1. First, DALI Backend tries to load a serialized model from the user-specified model location in default_model_filename property (model.dali if not specified explicitly);
2. If the previous fails, DALI Backend tries to load and autoserialize a Python pipeline definition from the user-specified model location. Important: In this case we require, that the file name with the model definition ends with .py, e.g. mymodel.py;
3. If the previous fails, DALI Backend tries to load and autoserialize a Python pipeline definition from the dali.py file in a given model version.

If you did not tweak a model path definition in the config.pbtxt file, you should follow the rule of thumb:

1. If you have a serialized pipeline, call the file model.dali and put it into the model repository,
2. If you have a python definition of a pipeline, which shall be autoserialized, call it dali.py.

Tips & Tricks:

1. Currently, the only way to pass an input to the DALI pipeline from Triton is to use the fn.external_source operator. Therefore, there's a high chance, that you'll want to use it to feed the encoded images (or any other data) into DALI.
2. Give your fn.external_source operator the same name you give to the Input in config.pbtxt.

Known limitations:

1. DALI's ImageDecoder accepts data only from the CPU - keep this in mind when putting together your DALI pipeline.
2. Triton accepts only homogeneous batch shape. Feel free to pad your batch of encoded images with zeros
3. Due to DALI limitations, you might observe unnaturally increased memory consumption when defining instance group for DALI model with higher count than 1. We suggest using default instance group for DALI model.

How to build?

Docker build

Building DALI Backend with docker is as simple as:

git clone --recursive https://github.com/triton-inference-server/dali_backend.git
cd dali_backend
docker build -f docker/Dockerfile.release -t tritonserver:dali-latest .

And tritonserver:dali-latest becomes your new tritonserver docker image

Bare metal

Prerequisites

To build dali_backend you'll need CMake 3.17+

Using fresh DALI release

On the event you'd need to use newer DALI version than it's provided in tritonserver image, you can use DALI's nightly builds. Just install whatever DALI version you like using pip (refer to the link for more info how to do it). In this case, while building dali_backend, you'd need to pass -D TRITON_SKIP_DALI_DOWNLOAD=ON option to your CMake build. dali_backend will find the latest DALI installed in your system and use this particular version.

Building

Building DALI Backend is really straightforward. One thing to remember is to clone dali_backend repository with all the submodules:

git clone --recursive https://github.com/triton-inference-server/dali_backend.git
cd dali_backend
mkdir build
cd build
cmake ..
make

The building process will generate unittest executable. You can use it to run unit tests for DALI Backend