This project (https://onnx.ai/onnx-mlir/) provides compiler technology to transform a valid Open Neural Network Exchange (ONNX) graph into code that implements the graph with minimum runtime support. It implements the ONNX standard and is based on the underlying LLVM/MLIR compiler technology.
System | Build Status | Model Zoo Status |
---|---|---|
s390x-Linux | ||
ppc64le-Linux | ||
amd64-Linux | ||
amd64-Windows | ||
amd64-macOS | ||
This project contributes:
- an ONNX Dialect that can be integrated in other projects,
- a compiler interfaces that lower ONNX graphs into MLIR files/LLVM bytecodes/C & Java libraries,
- an
onnx-mlir
driver to perform these lowering, - and a python/C/C++/Java runtime environment.
Current levels of support for the code generation of ONNX operations are listed here for a generic CPU and IBM's Telum integrated AI accelerator.
For ongoing discussions, we use an #onnx-mlir-discussion
slack channel established under the Linux Foundation AI and Data Workspace.
Join this workspace using this link.
We use GitHub Issues for request for comments, questions, or bug reports. Security-related issues are reported using the channels listed in the SECURITY page.
We hold informal weekly meetings on Tuesdays where we discuss current issues and progress. Meeting agenda, notes, and links (to participate) are found here. Please email [email protected] to request a 15-30 min time slot to discuss a specific topic of interest.
The preferred approach to using and developing ONNX-MLIR is to use Docker Images and Containers, as getting the proper code dependences may be tricky on some systems. Our instructions on using ONNX-MLIR with Dockers are here.
If you intend to develop code, you should look at our workflow document which help you setup your Docker environment in a way that let you contribute code easily.
ONNX-MLIR runs natively on Linux, OSX, and Windows. Detailed instructions are provided below.
python >= 3.8
gcc >= 6.4
protobuf >= 4.21.12
cmake >= 3.13.4
make >= 4.2.1 or ninja >= 1.10.2
java >= 1.11 (optional)
All the PyPi
package dependencies and their appropriate versions are captured in requirements.txt.
Look here for help to set up the prerequisite software.
At any point in time, ONNX-MLIR depends on a specific commit of the LLVM project that has been shown to work with the project. Periodically the maintainers need to move to a more recent LLVM level. Among other things, this requires to update the LLVM commit string in clone-mlir.sh. When updating ONNX-MLIR, it is good practice to check that the commit string of the MLIR/LLVM is the same as the one listed in that file. See instructions here when third-party ONNX also need to be updated.
Directions to install MLIR and ONNX-MLIR are dependent on your OS.
After installation, an onnx-mlir
executable should appear in the build/Debug/bin
or build/Release/bin
directory.
If you have difficulties building, rebuilding, or testing onnx-mlir
, check this page for helpful hints.
The usage of onnx-mlir
is as such:
OVERVIEW: ONNX-MLIR modular optimizer driver
USAGE: onnx-mlir [options] <input file>
OPTIONS:
Generic Options:
--help - Display available options (--help-hidden for more)
--help-list - Display list of available options (--help-list-hidden for more)
--version - Display the version of this program
ONNX-MLIR Options:
These are frontend options.
Choose target to emit:
--EmitONNXBasic - Ingest ONNX and emit the basic ONNX operations without inferred shapes.
--EmitONNXIR - Ingest ONNX and emit corresponding ONNX dialect.
--EmitMLIR - Lower the input to MLIR built-in transformation dialect.
--EmitLLVMIR - Lower the input to LLVM IR (LLVM MLIR dialect).
--EmitObj - Compile the input to an object file.
--EmitLib - Compile and link the input into a shared library (default).
--EmitJNI - Compile the input to a jar file.
Optimization levels:
--O0 - Optimization level 0 (default).
--O1 - Optimization level 1.
--O2 - Optimization level 2.
--O3 - Optimization level 3.
The full list of options is given by the -help
option.
The -
and the --
prefix for flags can be used interchangeably.
Note that just as most compilers, the default optimization level is -O0
.
We recommend using -O3
for most applications.
Options are also read from the ONNX_MLIR_FLAGS
environment variable. For example, ONNX_MLIR_FLAGS="-O3"
will ensure -O3
for all compilations.
For example, use the following command to lower an ONNX model (e.g., add.onnx) to ONNX dialect:
./onnx-mlir --EmitONNXIR add.onnx
The output should look like:
module {
func.func @main_graph(%arg0: tensor<10x10x10xf32>, %arg1: tensor<10x10x10xf32>) -> tensor<10x10x10xf32> {
%0 = "onnx.Add"(%arg0, %arg1) : (tensor<10x10x10xf32>, tensor<10x10x10xf32>) -> tensor<10x10x10xf32>
return %0 : tensor<10x10x10xf32>
}
}
An example based on the add operation is found here, which build an ONNX model using a python script, and then provide a main program to load the model's value, compute, and print the models output.
An end to end example is provided here, which train, compile, and execute a simple MNIST example using our C/C++, Python, or Java interface.
Documentation is provided in the docs
sub-directory; the DocumentList page provides an organized list of documents. Information is also provided on our public facing
onnx.ai/onnx-mlir pages.
We are welcoming contributions from the community. Please consult the CONTRIBUTING page for help on how to proceed.
ONNX-MLIR requires committers to sign their code using the Developer Certificate of Origin (DCO).
Practically, each git commit
needs to be signed, see here for specific instructions.
The ONNX-MLIR code of conduct is described at https://onnx.ai/codeofconduct.html.
- The onnx-mlir-serving project implements a GRPC server written with C++ to serve onnx-mlir compiled models. Benefiting from C++ implementation, ONNX Serving has very low latency overhead and high throughput.