Installation ¶

Precompiled Binaries ¶

Dependencies ¶

You need to have either PyTorch or LibTorch installed based on if you are using Python or C++ and you must have CUDA, cuDNN and TensorRT installed.

https://www.pytorch.org

https://developer.nvidia.com/cuda

https://developer.nvidia.com/cudnn

https://developer.nvidia.com/tensorrt

Python Package ¶

You can install the python package using

         pip3 install torch-tensorrt -f https://github.com/NVIDIA/Torch-TensorRT/releases

        

C++ Binary Distribution ¶

Precompiled tarballs for releases are provided here: https://github.com/NVIDIA/Torch-TensorRT/releases

Compiling From Source ¶

Dependencies for Compilation ¶

Torch-TensorRT is built with Bazel, so begin by installing it.

The easiest way is to install bazelisk using the method of your choosing https://github.com/bazelbuild/bazelisk

Otherwise you can use the following instructions to install binaries https://docs.bazel.build/versions/master/install.html

Finally if you need to compile from source (e.g. aarch64 until bazel distributes binaries for the architecture) you can use these instructions
export BAZEL_VERSION=$(cat <PATH_TO_TORCHTRT_ROOT>/.bazelversion)
mkdir bazel
cd bazel
curl -fSsL -O https://github.com/bazelbuild/bazel/releases/download/$BAZEL_VERSION/bazel-$BAZEL_VERSION-dist.zip
unzip bazel-$BAZEL_VERSION-dist.zip
bash ./compile.sh
cp output/bazel /usr/local/bin/

You will also need to have CUDA installed on the system (or if running in a container, the system must have the CUDA driver installed and the container must have CUDA)

The correct LibTorch version will be pulled down for you by bazel.

NOTE: For best compatability with official PyTorch, use torch==1.10.0+cuda113, TensorRT 8.0 and cuDNN 8.2 for CUDA 11.3 however Torch-TensorRT itself supports TensorRT and cuDNN for other CUDA versions for usecases such as using NVIDIA compiled distributions of PyTorch that use other versions of CUDA e.g. aarch64 or custom compiled version of PyTorch.

Choosing the Right ABI ¶

Likely the most complicated thing about compiling Torch-TensorRT is selecting the correct ABI. There are two options which are incompatible with each other, pre-cxx11-abi and the cxx11-abi. The complexity comes from the fact that while the most popular distribution of PyTorch (wheels downloaded from pytorch.org/pypi directly) use the pre-cxx11-abi, most other distributions you might encounter (e.g. ones from NVIDIA - NGC containers, and builds for Jetson as well as certain libtorch builds and likely if you build PyTorch from source) use the cxx11-abi. It is important you compile Torch-TensorRT using the correct ABI to function properly. Below is a table with general pairings of PyTorch distribution sources and the recommended commands:

PyTorch Source	Recommended C++ Compilation Command	Recommended Python Compilation Command
PyTorch whl file from PyTorch.org	bazel build //:libtorchtrt -c opt –config pre_cxx11_abi	python3 setup.py bdist_wheel
libtorch-shared-with-deps- * .zip from PyTorch.org	bazel build //:libtorchtrt -c opt –config pre_cxx11_abi	python3 setup.py bdist_wheel
libtorch-cxx11-abi-shared-with-deps- * .zip from PyTorch.org	bazel build //:libtorchtrt -c opt	python3 setup.py bdist_wheel –use-cxx11-abi
PyTorch preinstalled in an NGC container	bazel build //:libtorchtrt -c opt	python3 setup.py bdist_wheel –use-cxx11-abi
PyTorch from the NVIDIA Forums for Jetson	bazel build //:libtorchtrt -c opt	python3 setup.py bdist_wheel –jetpack-version 4.6 –use-cxx11-abi
PyTorch built from Source	bazel build //:libtorchtrt -c opt	python3 setup.py bdist_wheel –use-cxx11-abi

NOTE: For all of the above cases you must correctly declare the source of PyTorch you intend to use in your WORKSPACE file for both Python and C++ builds. See below for more information

You then have two compilation options:

Building using cuDNN & TensorRT tarball distributions ¶

This is recommended so as to build Torch-TensorRT hermetically and insures any compilation errors are not caused by version issues

Make sure when running Torch-TensorRT that these versions of the libraries are prioritized in your $LD_LIBRARY_PATH

You need to download the tarball distributions of TensorRT and cuDNN from the NVIDIA website.

Place these files in a directory (the directories thrid_party/distdir/[x86_64-linux-gnu | aarch64-linux-gnu] exist for this purpose)

Then compile referencing the directory with the tarballs

If you get errors regarding the packages, check their sha256 hashes and make sure they match the ones listed in WORKSPACE

Release Build ¶

         bazel build //:libtorchtrt -c opt --distdir thrid_party/distdir/[x86_64-linux-gnu | aarch64-linux-gnu]

        

A tarball with the include files and library can then be found in bazel-bin

Debug Build ¶

To build with debug symbols use the following command

         bazel build //:libtorchtrt -c dbg --distdir thrid_party/distdir/[x86_64-linux-gnu | aarch64-linux-gnu]

        

A tarball with the include files and library can then be found in bazel-bin

Pre CXX11 ABI Build ¶

To build using the pre-CXX11 ABI use the pre_cxx11_abi config

         bazel build //:libtorchtrt --config pre_cxx11_abi -c [dbg/opt] --distdir thrid_party/distdir/[x86_64-linux-gnu | aarch64-linux-gnu]

        

A tarball with the include files and library can then be found in bazel-bin

Building using locally installed cuDNN & TensorRT ¶

If you encounter bugs and you compiled using this method please disclose that you used local sources in the issue (an ldd dump would be nice too)

Install TensorRT, CUDA and cuDNN on the system before starting to compile.

In WORKSPACE comment out:

         # Downloaded distributions to use with --distdir
http_archive(
    name = "cudnn",
    urls = ["<URL>",],

    build_file = "@//third_party/cudnn/archive:BUILD",
    sha256 = "<TAR SHA256>",
    strip_prefix = "cuda"
)

http_archive(
    name = "tensorrt",
    urls = ["<URL>",],

    build_file = "@//third_party/tensorrt/archive:BUILD",
    sha256 = "<TAR SHA256>",
    strip_prefix = "TensorRT-<VERSION>"
)

        

and uncomment

         # Locally installed dependencies
new_local_repository(
    name = "cudnn",
    path = "/usr/",
    build_file = "@//third_party/cudnn/local:BUILD"
)

new_local_repository(
name = "tensorrt",
path = "/usr/",
build_file = "@//third_party/tensorrt/local:BUILD"
)

        

Release Build ¶

Compile using:

bazel build //:libtorchtrt -c opt

A tarball with the include files and library can then be found in bazel-bin

Debug Build ¶

To build with debug symbols use the following command

bazel build //:libtorchtrt -c dbg

A tarball with the include files and library can then be found in bazel-bin

Pre CXX11 ABI Build ¶

To build using the pre-CXX11 ABI use the pre_cxx11_abi config

bazel build //:libtorchtrt --config pre_cxx11_abi -c [dbg/opt]

Building the Python package ¶

Begin by installing ninja

You can build the Python package using setup.py (this will also build the correct version of libtorchtrt.so )

python3 setup.py [install/bdist_wheel]

Debug Build ¶

python3 setup.py develop [--user]

This also compiles a debug build of libtorchtrt.so

Building Natively on aarch64 (Jetson) ¶

Prerequisites ¶

Install or compile a build of PyTorch/LibTorch for aarch64

NVIDIA hosts builds the latest release branch for Jetson here:

https://forums.developer.nvidia.com/t/pytorch-for-jetson-version-1-10-now-available/72048

Enviorment Setup ¶

To build natively on aarch64-linux-gnu platform, configure the WORKSPACE with local available dependencies.

Disable the rules with http_archive for x86_64 by commenting the following rules:

         #http_archive(
#    name = "libtorch",
#    build_file = "@//third_party/libtorch:BUILD",
#    strip_prefix = "libtorch",
#    urls = ["https://download.pytorch.org/libtorch/cu102/libtorch-cxx11-abi-shared-with-deps-1.5.1.zip"],
#    sha256 = "cf0691493d05062fe3239cf76773bae4c5124f4b039050dbdd291c652af3ab2a"
#)

#http_archive(
#    name = "libtorch_pre_cxx11_abi",
#    build_file = "@//third_party/libtorch:BUILD",
#    strip_prefix = "libtorch",
#    sha256 = "818977576572eadaf62c80434a25afe44dbaa32ebda3a0919e389dcbe74f8656",
#    urls = ["https://download.pytorch.org/libtorch/cu102/libtorch-shared-with-deps-1.5.1.zip"],
#)

# Download these tarballs manually from the NVIDIA website
# Either place them in the distdir directory in third_party and use the --distdir flag
# or modify the urls to "file:///<PATH TO TARBALL>/<TARBALL NAME>.tar.gz

#http_archive(
#    name = "cudnn",
#    urls = ["https://developer.nvidia.com/compute/machine-learning/cudnn/secure/8.0.1.13/10.2_20200626/cudnn-10.2-linux-x64-v8.0.1.13.tgz"],
#    build_file = "@//third_party/cudnn/archive:BUILD",
#    sha256 = "0c106ec84f199a0fbcf1199010166986da732f9b0907768c9ac5ea5b120772db",
#    strip_prefix = "cuda"
#)

#http_archive(
#    name = "tensorrt",
#    urls = ["https://developer.nvidia.com/compute/machine-learning/tensorrt/secure/7.1/tars/TensorRT-7.1.3.4.Ubuntu-18.04.x86_64-gnu.cuda-10.2.cudnn8.0.tar.gz"],
#    build_file = "@//third_party/tensorrt/archive:BUILD",
#    sha256 = "9205bed204e2ae7aafd2e01cce0f21309e281e18d5bfd7172ef8541771539d41",
#    strip_prefix = "TensorRT-7.1.3.4"
#)

NOTE: You may also need to configure the CUDA version to 10.2 by setting the path for the cuda new_local_repository

        

Configure the correct paths to directory roots containing local dependencies in the new_local_repository rules:

NOTE: If you installed PyTorch using a pip package, the correct path is the path to the root of the python torch package. In the case that you installed with sudo pip install this will be /usr/local/lib/python3.6/dist-packages/torch . In the case you installed with pip install --user this will be $HOME/.local/lib/python3.6/site-packages/torch .

In the case you are using NVIDIA compiled pip packages, set the path for both libtorch sources to the same path. This is because unlike PyTorch on x86_64, NVIDIA aarch64 PyTorch uses the CXX11-ABI. If you compiled for source using the pre_cxx11_abi and only would like to use that library, set the paths to the same path but when you compile make sure to add the flag --config=pre_cxx11_abi

         new_local_repository(
    name = "libtorch",
    path = "/usr/local/lib/python3.6/dist-packages/torch",
    build_file = "third_party/libtorch/BUILD"
)

new_local_repository(
    name = "libtorch_pre_cxx11_abi",
    path = "/usr/local/lib/python3.6/dist-packages/torch",
    build_file = "third_party/libtorch/BUILD"
)

new_local_repository(
    name = "cudnn",
    path = "/usr/",
    build_file = "@//third_party/cudnn/local:BUILD"
)

new_local_repository(
    name = "tensorrt",
    path = "/usr/",
    build_file = "@//third_party/tensorrt/local:BUILD"
)

        

Compile C++ Library and Compiler CLI ¶

NOTE: Due to shifting dependency locations between Jetpack 4.5 and 4.6 there is a now a flag to inform bazel of the Jetpack version
--platforms //toolchains:jetpack_4.x

Compile Torch-TensorRT library using bazel command:

bazel build //:libtorchtrt --platforms //toolchains:jetpack_4.6

Compile Python API ¶

NOTE: Due to shifting dependencies locations between Jetpack 4.5 and Jetpack 4.6 there is now a flag for setup.py which sets the jetpack version (default: 4.6)

Compile the Python API using the following command from the //py directory:

python3 setup.py install --use-cxx11-abi

If you have a build of PyTorch that uses Pre-CXX11 ABI drop the --use-cxx11-abi flag

If you are building for Jetpack 4.5 add the --jetpack-version 4.5 flag