git clone https://github.com/llvm/torch-mlir
cd torch-mlir
git submodule update --init --progress
Optionally, use --depth=1
to make a shallow clone of the submodules.
While this is running, you can already setup the Python venv and dependencies in the next step.
Also, ensure that you have the appropriate python-dev
package installed
to access the Python development libraries / headers.
python -m venv mlir_venv
source mlir_venv/bin/activate
# Some older pip installs may not be able to handle the recent PyTorch deps
python -m pip install --upgrade pip
# Install latest PyTorch nightlies and build requirements.
python -m pip install -r requirements.txt
python -m pip install -r torchvision-requirements.txt
Two setups are possible to build: in-tree and out-of-tree. The in-tree setup is the most straightforward, as it will build LLVM dependencies as well.
The following command generates configuration files to build the project in-tree, that is, using llvm/llvm-project as the main build. This will build LLVM as well as torch-mlir and its subprojects. On Windows, use the "Developer PowerShell for Visual Studio" to ensure that the compiler and linker binaries are in the PATH
variable.
cmake -GNinja -Bbuild \
-DCMAKE_BUILD_TYPE=Release \
-DPython3_FIND_VIRTUALENV=ONLY \
-DLLVM_ENABLE_PROJECTS=mlir \
-DLLVM_EXTERNAL_PROJECTS="torch-mlir" \
-DLLVM_EXTERNAL_TORCH_MLIR_SOURCE_DIR="$PWD" \
-DMLIR_ENABLE_BINDINGS_PYTHON=ON \
-DLLVM_TARGETS_TO_BUILD=host \
externals/llvm-project/llvm
- Enabling clang on Linux
-DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++
- Enabling ccache
-DCMAKE_C_COMPILER_LAUNCHER=ccache -DCMAKE_CXX_COMPILER_LAUNCHER=ccache
- Enabling LLD (links in seconds compared to minutes)
-DCMAKE_EXE_LINKER_FLAGS_INIT="-fuse-ld=lld" -DCMAKE_MODULE_LINKER_FLAGS_INIT="-fuse-ld=lld" -DCMAKE_SHARED_LINKER_FLAGS_INIT="-fuse-ld=lld"
# Use --ld-path= instead of -fuse-ld=lld for clang > 13
- Enabling libtorch binary cache
By default we download the latest version of libtorch everytime you build so we are always on the latest version. Set
-DLIBTORCH_CACHE=ON
to not download the latest version everytime. If libtorch gets out of date and you test against a newer PyTorch you may notice failures.
-DLIBTORCH_CACHE=ON
- Enabling building libtorch as part of your build By default we download the latest version of libtorch. We have an experimental path to build libtorch (and PyTorch wheels) from source.
-DLIBTORCH_SRC_BUILD=ON # Build Libtorch from source
-DLIBTORCH_VARIANT=shared # Set the variant of libtorch to build / link against. (`shared`|`static` and optionally `cxxabi11`)
- Enabling
--debug
and--debug-only
flags (see MLIR docs) for thetorch-mlir-opt
tool
-DCMAKE_BUILD_TYPE=RelWithDebInfo \ # or =Debug
-DLLVM_ENABLE_ASSERTIONS=ON \
If you have built llvm-project separately in the directory $LLVM_INSTALL_DIR
, you can also build the project out-of-tree using the following command as template:
cmake -GNinja -Bbuild \
-DCMAKE_BUILD_TYPE=Release \
-DPython3_FIND_VIRTUALENV=ONLY \
-DMLIR_DIR="$LLVM_INSTALL_DIR/lib/cmake/mlir/" \
-DLLVM_DIR="$LLVM_INSTALL_DIR/lib/cmake/llvm/" \
-DMLIR_ENABLE_BINDINGS_PYTHON=ON \
-DLLVM_TARGETS_TO_BUILD=host \
.
The same QoL CMake flags can be used to enable clang, ccache, and lld. Be sure to have built LLVM with -DLLVM_ENABLE_PROJECTS=mlir
.
Be aware that the installed version of LLVM needs in general to match the committed version in externals/llvm-project
. Using a different version may or may not work.
After either cmake run (in-tree/out-of-tree), use one of the following commands to build the project:
# Build just torch-mlir (not all of LLVM)
cmake --build build --target tools/torch-mlir/all
# Run unit tests.
cmake --build build --target check-torch-mlir
# Run Python regression tests.
cmake --build build --target check-torch-mlir-python
# Build everything (including LLVM if in-tree)
cmake --build build
export PYTHONPATH=`pwd`/build/tools/torch-mlir/python_packages/torch_mlir:`pwd`/test/python/fx_importer
$env:PYTHONPATH = "$PWD/build/tools/torch-mlir/python_packages/torch_mlir;$PWD/test/python/fx_importer"
To test the MLIR output to torch dialect, you can use test/python/fx_importer/basic_test.py
.
Make sure you have activated the virtualenv and set the PYTHONPATH
above
(if running on Windows, modify the environment variable as shown above):
source mlir_venv/bin/activate
export PYTHONPATH=`pwd`/build/tools/torch-mlir/python_packages/torch_mlir:`pwd`/test/python/fx_importer
python test/python/fx_importer/basic_test.py
This will display the basic example in TORCH dialect.
To test the compiler's output to the different MLIR dialects, you can also use the deprecated path
using torchscript with the example projects/pt1/examples/torchscript_resnet18_all_output_types.py
.
This path doesn't give access to the current generation work that is being driven via the fx_importer
and may lead to errors.
Same as above, but with different python path and example:
export PYTHONPATH=`pwd`/build/tools/torch-mlir/python_packages/torch_mlir:`pwd`/projects/pt1/examples
python projects/pt1/examples/torchscript_resnet18_all_output_types.py
This will display the Resnet18 network example in three dialects: TORCH, LINALG on TENSORS and TOSA.
The main functionality is on torch_mlir.torchscript.compile()
's output_type
.
Ex:
module = torch_mlir.torchscript.compile(resnet18, torch.ones(1, 3, 224, 224), output_type="torch")
output_type
can be: TORCH
, TOSA
, LINALG_ON_TENSORS
, RAW
and STABLEHLO
.
Jupyter notebook:
python -m ipykernel install --user --name=torch-mlir --env PYTHONPATH "$PYTHONPATH"
# Open in jupyter, and then navigate to
# `examples/resnet_inference.ipynb` and use the `torch-mlir` kernel to run.
jupyter notebook
Example IR for a simple 1 layer MLP to show the compilation steps from TorchScript.
The build_tools/write_env_file.sh
script will output a .env
file in the workspace folder with the correct PYTHONPATH set. This allows
tools like VSCode to work by default for debugging. This file can also be
manually source
'd in a shell.
NOTE Our Bazel build follows LLVM's Bazel build policy: only the subcommunity interested in Bazel is responsible for fixing it. Average Torch-MLIR developers should not be notified of any Bazel build issues and are not responsible for fixing any breakages (though any help is, of course, welcome). For more info, see LLVM's Peripheral Support Tier definition.
Torch-MLIR can also be built using Bazel (apart from the official CMake build) for users that depend on Bazel in their workflows. To build torch-mlir-opt
using Bazel, follow these steps:
- Launch an interactive docker container with the required deps installed:
./utils/bazel/docker/run_docker.sh
- Build torch-mlir:
bazel build @torch-mlir//:torch-mlir-opt
The built binary should be at bazel-bin/external/torch-mlir/torch-mlir-opt
.
- Test torch-mlir (lit test only):
bazel test @torch-mlir//test/...
We welcome patches to torch-mlir's Bazel build. If you do contribute, please complete your PR with an invocation of buildifier to ensure the BUILD files are formatted consistently:
bazel run @torch-mlir//:buildifier
We have preliminary support for building with Docker images. Currently this is not very convenient for day-to-day interactive development and debugging flows but is very useful for reproducing failures from the CI. This is a new flow and we would like your feedback on how it works for you and please feel free to file any feedback or issues.
Install Docker Engine. You don't need Docker Desktop.
You have three types of builds selectable with the Environment Variable TM_PACKAGES
:torch-mlir
the
Release build, out-of-tree
where torch-mlir is build with a pre-built MLIR and in-tree
where torch-mlir
is built as part of the LLVM project along with MLIR.
We mount a ccache and pip cache inside the docker container to speed up iterative builds. Iterative builds should be as fast as running without docker.
Build MLIR and Torch-MLIR together as part of the LLVM repo.
TM_PACKAGES="in-tree" ./build_tools/python_deploy/build_linux_packages.sh
Build LLVM/MLIR first and then build Torch-MLIR referencing that build
TM_PACKAGES="out-of-tree" ./build_tools/python_deploy/build_linux_packages.sh
Build in a manylinux Docker image so we can upload artifacts to PyPI.
TM_PACKAGES="torch-mlir" ./build_tools/python_deploy/build_linux_packages.sh
If you wanted to build all the CIs locally
TM_PACKAGES="out-of-tree in-tree" ./build_tools/python_deploy/build_linux_packages.sh
If you wanted to build all the CIs and the Release builds (just with Python 3.10 since most other Python builds are redundant)
TM_PACKAGES="torch-mlir out-of-tree in-tree" TM_PYTHON_VERSIONS="cp310-cp310" ./build_tools/python_deploy/build_linux_packages.sh
Note: The Release docker still runs as root so it may generate some files owned by root:root. We hope to move it to run as a user in the future.
Docker builds tend to leave a wide variety of files around. Luckily most are owned by the user but there are still some that need to be removed as superuser.
rm -rf build build_oot llvm-build docker_venv externals/pytorch/build .ccache
If you would like to build your own docker image (usually not necessary). You can run:
cd ./build_tools/docker
docker build -t your-name/torch-mlir-ci --no-cache .
The following additional environmental variables can be used to customize your docker build:
- Custom Release Docker image:
Defaults to
stellaraccident/manylinux2014_x86_64-bazel-5.1.0:latest
TM_RELEASE_DOCKER_IMAGE="stellaraccident/manylinux2014_x86_64-bazel-5.1.0:latest"
- Custom CI Docker image:
Defaults to
powderluv/torch-mlir-ci:latest
. This assumes an Ubuntu LTS like image. You can build your own with./build_tools/docker/Dockerfile
TM_CI_DOCKER_IMAGE="powderluv/torch-mlir-ci:latest"
- Custom Python Versions for Release builds:
Version of Python to use in Release builds. Ignored in CIs. Defaults to
cp38-cp38 cp39-cp39 cp310-cp310
TM_PYTHON_VERSIONS="cp38-cp38 cp39-cp39 cp310-cp310"
- Location to store Release build wheels
TM_OUTPUT_DIR="./build_tools/python_deploy/wheelhouse"
- What "packages" to build:
Defaults to torch-mlir. Options are
torch-mlir out-of-tree in-tree
TM_PACKAGES="torch-mlir out-of-tree in-tree"
- Use pre-built Pytorch:
Defaults to using pre-built Pytorch. Setting it to
OFF
builds from source
TM_USE_PYTORCH_BINARY="OFF"
- Skip running tests
Skip running tests if you want quick build only iteration. Default set to
OFF
TM_SKIP_TESTS="OFF"
We have preliminary support for building Python packages. This can be done with the following commands:
python -m pip install --upgrade pip
python -m pip install -r requirements.txt
CMAKE_GENERATOR=Ninja python setup.py bdist_wheel
Torch-MLIR has two types of tests:
-
End-to-end execution tests. These compile and run a program and check the result against the expected output from execution on native Torch. These use a homegrown testing framework (see
projects/pt1/python/torch_mlir_e2e_test/framework.py
) and the test suite lives atprojects/pt1/python/torch_mlir_e2e_test/test_suite/__init__.py
. -
Compiler and Python API unit tests. These use LLVM's
lit
testing framework. For example, these might involve usingtorch-mlir-opt
to run a pass and check the output withFileCheck
. These tests usually live in thetest/
directory with a parallel file naming scheme to thelib/*
structure. More details about this kind of test is available in the upstream LLVM Testing Guide.
Note An
.env
file must be generated viabuild_tools/write_env_file.sh
before these commands can be run.
The following assumes you are in the projects/pt1
directory:
# Run all tests on the reference backend
./tools/e2e_test.sh
# Run tests that match the regex `Conv2d`, with verbose errors.
./tools/e2e_test.sh --filter Conv2d --verbose
# Run tests on the TOSA backend.
./tools/e2e_test.sh --config tosa
Alternatively, you can run the tests via Python directly:
cd projects/pt1
python -m e2e_testing.main -f 'AtenEmbeddingBag'
To run all of the unit tests, run:
ninja check-torch-mlir-all
This can be broken down into
ninja check-torch-mlir check-torch-mlir-python
To run more fine-grained tests, you can do, for check-torch-mlir
:
cd $TORCH_MLIR_BUILD_DIR/tools/torch-mlir/test
$TORCH_MLIR_BUILD_DIR/bin/llvm-lit $TORCH_MLIR_SRC_ROOT/test -v --filter=canonicalize
See the lit
documentation for details on the available lit args.
For example, if you wanted to test just test/Dialect/Torch/canonicalize.mlir
,
then you might do
cd $TORCH_MLIR_BUILD_DIR/tools/torch-mlir/test
$TORCH_MLIR_BUILD_DIR/bin/llvm-lit $TORCH_MLIR_SRC_ROOT/test -v --filter=canonicalize.mlir
Most of the unit tests use the FileCheck
tool to verify expected outputs.
Torch-MLIR by default builds with the latest nightly PyTorch version. This can be toggled to build from latest PyTorch source with
-DTORCH_MLIR_USE_INSTALLED_PYTORCH=OFF
-DTORCH_MLIR_SRC_PYTORCH_REPO=vivekkhandelwal1/pytorch # Optional. Github path. Defaults to pytorch/pytorch
-DTORCH_MLIR_SRC_PYTORCH_BRANCH=master # Optional. Defaults to PyTorch's main branch
Torch-MLIR depends on llvm-project
(which contains, among other things,
upstream MLIR) and stablehlo
, both of which are submodules in the externals/
directory. We aim to update these at least weekly to bring in the latest
features and spread out over time the effort of updating our code for MLIR API
breakages.
NOTE: This section is in flux. Specifically, the mlir-hlo
dep has been
dropped and the project is running off of a stablehlo
fork which can be
patched for certain OS combinations. As of 2023-09-12, stellaraccident@
is massaging this situation. Please reach out for advice updating.
Since downstream projects may want to build Torch-MLIR (and thus LLVM and MLIR-HLO) in various configurations (Release versus Debug builds; on Linux, Windows, or macOS; possibly with Clang, LLD, and LLDB enabled), it is crucial to pick LLVM commits that pass tests for all combinations of these configurations.
So every week, we track the so-called green commit (i.e. the LLVM commit which works with all of the above configurations) in Issue llvm#1178. In addition to increasing our confidence that the resulting update will not break downstream projects, basing our submodule updates on these green commits also helps us stay in sync with LLVM updates in other projects like ONNX-MLIR and MLIR-HLO. The person responsible for the update each week is listed here.
- Lookup green commit hashes: From the Github issue llvm#1178, find the LLVM and MLIR-HLO green commits for the week when Torch-MLIR is being updated.
- Update the
llvm-project
submodule: In theexternals/llvm-project
directory, rungit fetch
followed bygit checkout <llvm-commit-hash>
(where<llvm-commit-hash>
is the green commit hash for the LLVM project from Step 1). - Update the
mlir-hlo
submodule: In theexternals/mlir-hlo
directory, rungit fetch
followed bygit checkout <mlir-hlo-commit-hash>
(where<mlir-hlo-commit-hash>
is the green commit hash for the MLIR-HLO project from Step 1). - Rebuild and test Torch-MLIR: See the section "CMake Build" above for instructions, fixing any issues that arise. This might involve fixing various API breakages introduced upstream (they are likely unrelated to what you are working on). If these fixes are too complex, please file a work-in-progress PR explaining the issues you are running into asking for help so that someone from the community can help.
- Update Abstract Interpretation Library: Run
build_tools/update_abstract_interp_lib.sh
. This is sometimes needed because upstream changes can affect canonicalization and other minor details of the IR in the abstract interpretation library. See docs/abstract_interp_lib.md for more details on the abstract interpretation library.
Here are some examples of PRs updating the LLVM and MLIR-HLO submodules:
To enable ASAN, pass -DLLVM_USE_SANITIZER=Address
to CMake. This should "just
work" with all C++ tools like torch-mlir-opt
. When running a Python script
such as through ./projects/pt1/tools/e2e_test.sh
, you will need to do:
LD_PRELOAD="$(clang -print-file-name=libclang_rt.asan-x86_64.so)" ./projects/pt1/tools/e2e_test.sh -s
# See instructions here for how to get the libasan path for GCC:
# https://stackoverflow.com/questions/48833176/get-location-of-libasan-from-gcc-clang
TODO: Add ASan docs for LTC.
- GitHub wiki: https://github.com/llvm/torch-mlir/wiki
- Of particular interest in the How to add end-to-end support for new Torch ops doc.