Skip to content

A backend which allows the usage of ONNX MLIR compiled models (model.so) with the Triton Inference Server.

License

Notifications You must be signed in to change notification settings

ghas-results/onnxmlir-triton-backend

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

58 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

onnxmlir-triton-backend

A triton backend which allows the usage of onnx-mlir or zDLC compiled models (model.so) with the triton inference server.

At the moment there is no GPU support.

Model Repository

The backend expects the compiled model to be named model.so The complete structure looks

  <model-repository-path>/
    <model-name>/
      config.pbtxt
      <version>/
        model.so
      <version>/
        model.so
      ...

Model Configuration

Specify the backend name onnxmlir in the config.pbtxt:

backend: "onnxmlir"

For more options see Model Configuration.

Build and Install

You can either build the backend and copy the shared library manually to your triton installation or let make install directly install it tor your triton installation.

Manual Install

mkdir build
cd build
cmake -DCMAKE_INSTALL_PREFIX:PATH=`pwd`/install ..
make install

produces build/install/backends/onnxmlir/libtriton_onnxmlir.so

The libtriton_onnxmlir.so needs to be copied to backends/onnxmlir directory in the triton installation directory (usually /opt/tritonserver).

Direct Install

You can specify the triton install directory on the cmake command so make install will install it directly in your trition installation. If triton is installed to /opt/tritonserver you can use

mkdir build
cd build
cmake -DCMAKE_INSTALL_PREFIX:PATH=/opt/tritonserver ..
make install

About

A backend which allows the usage of ONNX MLIR compiled models (model.so) with the Triton Inference Server.

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Languages

  • C++ 82.9%
  • CMake 17.1%