Note: OpenVINO backend is beta quality. As a result you may encounter performance and functional issues that will be resolved in future releases.

The Triton backend for the OpenVINO. You can learn more about Triton backends in the backend repo. Ask questions or report problems in the main Triton issues page. The backend is designed to run models in Intermediate Representation (IR). See here for instruction to convert a model to IR format. The backend is implemented using openVINO C++ API. Auto completion of the model config is not supported in the backend and complete config.pbtxt must be provided with the model.

OpenVINO backend currently supports inference only on Intel CPU devices using OpenVINO CPU plugin. Note the CPU plugin does not support iGPU.

Cmake 3.17 or higher is required. First install the required dependencies.

$ apt-get install patchelf rapidjson-dev python3-dev

Follow the steps below to build the backend shared library.

$ mkdir build
$ cd build
$ cmake -DCMAKE_INSTALL_PREFIX:PATH=`pwd`/install -DTRITON_BUILD_OPENVINO_VERSION=2021.2.200 -DTRITON_BUILD_CONTAINER_VERSION=20.12 ..
$ make install

The following required Triton repositories will be pulled and used in the build. By default the "main" branch/tag will be used for each repo but the listed CMake argument can be used to override.

• triton-inference-server/backend: -DTRITON_BACKEND_REPO_TAG=[tag]
• triton-inference-server/core: -DTRITON_CORE_REPO_TAG=[tag]
• triton-inference-server/common: -DTRITON_COMMON_REPO_TAG=[tag]

Configuration of OpenVINO for a model is done through the Parameters section of the model's 'config.pbtxt' file. The parameters and their description are as follows.

• PERFORMANCE_HINT: Presetting performance tuning options. Accepted values LATENCY for low concurrency use case and THROUGHPUT for high concurrency scenarios.
• CPU_EXTENSION_PATH: Required for CPU custom layers. Absolute path to a shared library with the kernels implementations.
• INFERENCE_NUM_THREADS: Maximum number of threads that can be used for inference tasks. Should be a non-negative number. Default is equal to number of cores.
• COMPILATION_NUM_THREADS: Maximum number of threads that can be used for compilation tasks. Should be a non-negative number.
• HINT_BF16: Hint for device to use bfloat16 precision for inference. Possible value is YES.
• NUM_STREAMS: The number of executor logical partitions. Set the value to AUTO to creates bare minimum of streams to improve the performance, or set the value to NUMA to creates as many streams as needed to accommodate NUMA and avoid associated penalties. Set a numerical value to set explicit number of streams.
• SKIP_OV_DYNAMIC_BATCHSIZE: The topology of some models do not support openVINO dynamic batch sizes. Set the value of this parameter to YES, in order to skip the dynamic batch sizes in backend.
• ENABLE_BATCH_PADDING: By default an error will be generated if backend receives a request with batch size less than max_batch_size specified in the configuration. This error can be avoided at a cost of performance by specifying ENABLE_BATCH_PADDING parameter as YES.
• RESHAPE_IO_LAYERS: By setting this parameter as YES, the IO layers are reshaped to the dimensions provided in model configuration. By default, the dimensions in the model is used.

Assuming Triton was not started with --disable-auto-complete-config command line option, the OpenVINO Backend makes use of the model configuration available in OpenVINO models to populate the required fields in the model's "config.pbtxt". You can learn more about Triton's support for auto-completing model configuration from here.

However, not all OpenVINO models carry sufficient configuration information to auto-complete the model's "config.pbtxt". As a result, a partial "config.pbtxt" could still be required for some models.

OpenVINO backend can complete the following fields in model configuration:

Auto-completing max_batch_size follows the following rules:

• The model has included sufficient layout information.
• Autocomplete has determined the model is capable of batching requests.
• max_batch_size is 0 in the model configuration or max_batch_size is omitted from the model configuration.

If the above two rules are met, max_batch_size is set to default-max-batch-size. Otherwise max_batch_size is set as 0.

The OpenVINO Backend is able to fill in the name, data_type, and dims provided this information is available in the model.

Autocompleting inputs/outputs follows the following rules:

• If inputs and/or outputs is empty or undefined in the model configuration, the missing inputs and/or outputs will be autocompleted.
• Auto-complete will skip over any defined/filled-in inputs and outputs.

If max_batch_size > 1, after auto-completing max_batch_size, and no dynamic_batching and sequence_batching is provided, then dynamic_batching will be enabled with default settings.

Latency mode with low concurrency on the client side. Recommended for performance optimization with low number of parallel clients.

parameters: [
{
   key: "NUM_STREAMS"
   value: {
     string_value: "1"
   }
},
{
   key: "PERFORMANCE_HINT"
   value: {
     string_value: "LATENCY"
   }
}
]

Throughput mode with high concurrency on the client side. Recommended for throughput optimization with high number of parallel clients. Number of streams should be lower or equal to number of parallel clients and lower of equal to the number of CPU cores. For example, with ~20 clients on the host with 12 CPU cores, the config could be like:

instance_group [
    {
      count: 12
      kind: KIND_CPU
    }
  ]
parameters: [
{
   key: "NUM_STREAMS"
   value: {
     string_value: "12"
   }
}
]

When loading model with the non default format of Intermediate Representation and the name model.xml, use and extra parameter "default_model_filename". For example, using TensorFlow saved_model format use:

default_model_filename: "model.saved_model"

and copy the model to the subfolder called "model.saved_model"

model_repository/
└── model
    ├── 1
    │   └── model.saved_model
    │       ├── saved_model.pb
    │       └── variables
    └── config.pbtxt

Other allowed values are model.pdmodel or model.onnx.

Following section shows how to use OpenVINO dynamic shapes. -1 denotes dimension accepting any value on input. In this case while model originally accepted input with layout NCHW and shape (1,3,224,224), now it accepts any batch size and resolution.

Note: If the model is originally exported with dynamic shapes, there is no need to manually specify dynamic shapes in config.

Note: Some models might not support a shape with an arbitrary dimension size. An error will be raised during the model initialization when the shaped in the config.pbtxt is not possible. Clients will also receive an error if the request includes unsupported input shapes.

input [
  {
    name: "input"
    data_type: TYPE_FP32
    dims: [ -1, 3, -1, -1]
  }
]
output [
  {
    name: "output"
    data_type: TYPE_FP32
    dims: [ -1, 1001]
  }
]
parameters: {
key: "RESHAPE_IO_LAYERS"
value: {
string_value:"yes"
}
}

• Models with the scalar on the input (shape without any dimension are not supported)
• Reshaping using dimension ranges is not supported.