- Contributing to PyTorch
- Developing PyTorch
- Codebase structure
- Unit testing
- Writing documentation
- Profiling with
py-spy
- Managing multiple build trees
- C++ development tips
- CUDA development tips
- Windows development tips
- Caffe2 notes
If you are interested in contributing to PyTorch, your contributions will fall into two categories:
- You want to propose a new feature and implement it.
- Post about your intended feature, and we shall discuss the design and implementation. Once we agree that the plan looks good, go ahead and implement it.
- You want to implement a feature or bug-fix for an outstanding issue.
- Search for your issue here: https://github.com/pytorch/pytorch/issues
- Pick an issue and comment on the task that you want to work on this feature.
- If you need more context on a particular issue, please ask and we shall provide.
Once you finish implementing a feature or bug-fix, please send a Pull Request to https://github.com/pytorch/pytorch
This document covers some of the more technical aspects of contributing to PyTorch. For more non-technical guidance about how to contribute to PyTorch, see the Contributing Guide.
To develop PyTorch on your machine, here are some tips:
- Uninstall all existing PyTorch installs:
conda uninstall pytorch
pip uninstall torch
pip uninstall torch # run this command twice
- Clone a copy of PyTorch from source:
git clone https://github.com/pytorch/pytorch
cd pytorch
2.1. If you already have PyTorch from source, update it:
git pull --rebase
git submodule sync --recursive
git submodule update --init --recursive
If you want to have no-op incremental rebuilds (which are fast), see the section below titled "Make no-op build fast."
- Install PyTorch in
develop
mode:
A full set of instructions on installing PyTorch from source is here: https://github.com/pytorch/pytorch#from-source
The change you have to make is to replace
python setup.py install
with
python setup.py develop
This mode will symlink the Python files from the current local source tree into the Python install. Hence, if you modify a Python file, you do not need to reinstall PyTorch again and again. This is especially useful if you are only changing Python files.
For example:
- Install local PyTorch in
develop
mode - modify your Python file
torch/__init__.py
(for example) - test functionality
- modify your Python file
torch/__init__.py
- test functionality
- modify your Python file
torch/__init__.py
- test functionality
You do not need to repeatedly install after modifying Python files.
In case you want to reinstall, make sure that you uninstall PyTorch first by running pip uninstall torch
and python setup.py clean
. Then you can install in develop
mode again.
- c10 - Core library files that work everywhere, both server and mobile. We are slowly moving pieces from ATen/core here. This library is intended only to contain essential functionality, and appropriate to use in settings where binary size matters. (But you'll have a lot of missing functionality if you try to use it directly.)
- aten - C++ tensor library for PyTorch (no autograd support)
- src
- TH
THC
THNN
THCUNN - Legacy library code from the original
Torch. Try not to add things here; we're slowly porting these to
native.
- generic - Contains actual implementations of operators,
parametrized over
scalar_t
. Files here get compiled N times per supported scalar type in PyTorch.
- generic - Contains actual implementations of operators,
parametrized over
- ATen
- core - Core functionality of ATen. This is migrating to top-level c10 folder.
- native - Modern implementations of
operators. If you want to write a new operator, here is where
it should go. Most CPU operators go in the top level directory,
except for operators which need to be compiled specially; see
cpu below.
- cpu - Not actually CPU implementations of operators, but specifically implementations which are compiled with processor-specific instructions, like AVX. See the README for more details.
- cuda - CUDA implementations of operators.
- sparse - CPU and CUDA implementations of COO sparse tensor operations
- mkl mkldnn
miopen cudnn
- implementations of operators which simply bind to some backend library.
- TH
THC
THNN
THCUNN - Legacy library code from the original
Torch. Try not to add things here; we're slowly porting these to
native.
- src
- torch - The actual PyTorch library. Everything that is not
in csrc is a Python module, following the PyTorch Python
frontend module structure.
- csrc - C++ files composing the PyTorch library. Files
in this directory tree are a mix of Python binding code, and C++
heavy lifting. Consult
setup.py
for the canonical list of Python binding files; conventionally, they are often prefixed withpython_
.- jit - Compiler and frontend for TorchScript JIT frontend.
- autograd - Implementation of reverse-mode automatic differentiation.
- api - The PyTorch C++ frontend.
- distributed - Distributed training support for PyTorch.
- csrc - C++ files composing the PyTorch library. Files
in this directory tree are a mix of Python binding code, and C++
heavy lifting. Consult
- tools - Code generation scripts for the PyTorch library. See README of this directory for more details.
- test - Python unit tests for PyTorch Python frontend.
- test_torch.py - Basic tests for PyTorch functionality.
- test_autograd.py - Tests for non-NN automatic differentiation support.
- test_nn.py - Tests for NN operators and their automatic differentiation.
- test_jit.py - Tests for the JIT compiler and TorchScript.
- ...
- cpp - C++ unit tests for PyTorch C++ frontend.
- expect - Automatically generated "expect" files which are used to compare against expected output.
- onnx - Tests for ONNX export functionality, using both PyTorch and Caffe2.
- caffe2 - The Caffe2 library.
PyTorch's testing is located under test/
. Run the entire test suite with
python test/run_test.py
or run individual test files, like python test/test_nn.py
, for individual test suites.
We don't officially support pytest
, but it works well with our unittest
tests and offers
a number of useful features for local developing. Install it via pip install pytest
.
If you want to just run tests that contain a specific substring, you can use the -k
flag:
pytest test/test_nn.py -k Loss -v
The above is an example of testing a change to Loss functions: this command runs tests such as
TestNN.test_BCELoss
and TestNN.test_MSELoss
and can be useful to save keystrokes.
PyTorch uses Google style for formatting docstrings. Length of line inside docstrings block must be limited to 80 characters to fit into Jupyter documentation popups.
To build the documentation:
-
Build and install PyTorch
-
Install the prequesities
cd docs
pip install -r requirements.txt
# `katex` must also be available in your PATH.
# You can either install katex globally if you have properly configured npm:
# npm install -g katex
# Or if you prefer an uncontaminated global executable environment or do not want to go through the node configuration:
# npm install katex && export PATH="$PATH:$(pwd)/node_modules/.bin"
- Generate the documentation HTML files. The generated files will be in
docs/build/html
.
cd docs
make html
The .rst
source files live in docs/source. Some of the .rst
files pull in docstrings from PyTorch Python code (for example, via
the autofunction
or autoclass
directives). To vastly shorten doc build times,
it is helpful to remove the files you are not working on, only keeping the base
index.rst
file and the files you are editing. The Sphinx build will produce
missing file warnings but will still complete. For example, to work on jit.rst
:
cd docs/source
ls | grep rst | grep -v index | grep -v jit | xargs rm
# Make your changes, build the docs, etc.
# Don't commit the deletions!
git add index.rst jit.rst
...
For C++ documentation (https://pytorch.org/cppdocs), we use Doxygen and then convert it to Sphinx via Breathe and Exhale. Check the Doxygen reference for more information on the documentation syntax.
We run Doxygen in CI (Travis) to verify that you do not use invalid Doxygen
commands. To run this check locally, run ./check-doxygen.sh
from inside
docs/cpp
.
To build the documentation, follow the same steps as above, but run them from
docs/cpp
instead of docs
.
To view HTML files locally, you can open the files in your web browser. For example,
navigate to file:///your_pytorch_folder/docs/build/html/index.html
in a web
browser.
If you are developing on a remote machine, you can set up an SSH tunnel so that you can access the HTTP server on the remote machine from your local machine. To map remote port 8000 to local port 8000, use either of the following commands.
# For SSH
ssh my_machine -L 8000:my_machine:8000
# For Eternal Terminal
et my_machine -t="8000:8000"
Then navigate to localhost:8000
in your web browser.
It is helpful when submitting a PR that changes the docs to provide a rendered version of the result. If your change is small, you can add a screenshot of the changed docs to your PR.
If your change to the docs is large and affects multiple pages, you can host the docs yourself with the following steps, then add a link to the output in your PR. These instructions use GitHub pages to host the docs you have built. To do so, follow these steps to make a repo to host your changed documentation.
GitHub pages expects to be hosting a Jekyll generated website which does not work
well with the static resource paths used in the PyTorch documentation. To get around
this, you must add an empty file called .nojekyll
to your repo.
cd your_github_pages_repo
touch .nojekyll
git add .
git commit
git push
Then, copy built documentation and push the changes:
cd your_github_pages_repo
cp -r ~/my_pytorch_path/docs/build/html/* .
git add .
git commit
git push
Then you should be able to see the changes at your_github_username.github.com/your_github_pages_repo.
It is easy for code snippets in docstrings and .rst
files to get out of date. The docs
build includes the Sphinx Doctest Extension,
which can run code in documentation as a unit test. To use the extension, use
the .. testcode::
directive in your .rst
and docstrings.
To manually run these tests, follow steps 1 and 2 above, then run:
cd docs
make doctest
Evaluating the performance impact of code changes in PyTorch can be complicated,
particularly if code changes happen in compiled code. One simple way to profile
both Python and C++ code in PyTorch is to use
py-spy
, a sampling profiler for Python
that has the ability to profile native code and Python code in the same session.
py-spy
can be installed via pip
:
$ pip install py-spy
To use py-spy
, first write a Python test script that exercises the
functionality you would like to profile. For example, this script profiles
torch.add
:
import torch
t1 = torch.tensor([[1, 1], [1, 1.]])
t2 = torch.tensor([[0, 0], [0, 0.]])
for _ in range(1000000):
torch.add(t1, t2)
Since the torch.add
operation happens in microseconds, we repeat it a large
number of times to get good statistics. The most straightforward way to use
py-spy
with such a script is to generate a flame
graph:
$ py-spy record -o profile.svg --native -- python test_tensor_tensor_add.py
This will output a file named profile.svg
containing a flame graph you can
view in a web browser or SVG viewer. Individual stack frame entries in the graph
can be selected interactively with your mouse to zoom in on a particular part of
the program execution timeline. The --native
command-line option tells
py-spy
to record stack frame entries for PyTorch C++ code. To get line numbers
for C++ code it may be necessary to compile PyTorch in debug mode by prepending
your setup.py develop
call to compile PyTorch with DEBUG=1
. Depending on
your operating system it may also be necessary to run py-spy
with root
privileges.
py-spy
can also work in an htop
-like "live profiling" mode and can be
tweaked to adjust the stack sampling rate, see the py-spy
readme for more
details.
One downside to using python setup.py develop
is that your development
version of PyTorch will be installed globally on your account (e.g., if
you run import torch
anywhere else, the development version will be
used.
If you want to manage multiple builds of PyTorch, you can make use of conda environments to maintain separate Python package environments, each of which can be tied to a specific build of PyTorch. To set one up:
conda create -n pytorch-myfeature
source activate pytorch-myfeature
# if you run python now, torch will NOT be installed
python setup.py develop
If you are working on the C++ code, there are a few important things that you will want to keep in mind:
- How to rebuild only the code you are working on.
- How to make rebuilds in the absence of changes go faster.
python setup.py build
will build everything by default, but sometimes you are
only interested in a specific component.
- Working on a test binary? Run
(cd build && ninja bin/test_binary_name)
to rebuild only that test binary (without rerunning cmake). (Replaceninja
withmake
if you don't have ninja installed). - Don't need Caffe2? Pass
BUILD_CAFFE2_OPS=0
to disable build of Caffe2 operators.
On the initial build, you can also speed things up with the environment
variables DEBUG
, USE_DISTRIBUTED
, USE_MKLDNN
, USE_CUDA
, BUILD_TEST
, USE_FBGEMM
, USE_NNPACK
and USE_QNNPACK
.
DEBUG=1
will enable debug builds (-g -O0)REL_WITH_DEB_INFO=1
will enable debug symbols with optimizations (-g -O3)USE_DISTRIBUTED=0
will disable distributed (c10d, gloo, mpi, etc.) build.USE_MKLDNN=0
will disable using MKL-DNN.USE_CUDA=0
will disable compiling CUDA (in case you are developing on something not CUDA related), to save compile time.BUILD_TEST=0
will disable building C++ test binaries.USE_FBGEMM=0
will disable using FBGEMM (quantized 8-bit server operators).USE_NNPACK=0
will disable compiling with NNPACK.USE_QNNPACK=0
will disable QNNPACK build (quantized 8-bit operators).
For example:
DEBUG=1 USE_DISTRIBUTED=0 USE_MKLDNN=0 USE_CUDA=0 BUILD_TEST=0 USE_FBGEMM=0 USE_NNPACK=0 USE_QNNPACK=0 python setup.py develop
For subsequent builds (i.e., when build/CMakeCache.txt
exists), the build
options passed for the first time will persist; please run ccmake build/
, run
cmake-gui build/
, or directly edit build/CMakeCache.txt
to adapt build
options.
When using python setup.py develop
, PyTorch will generate
a compile_commands.json
file that can be used by many editors
to provide command completion and error highlighting for PyTorch's
C++ code. You need to pip install ninja
to generate accurate
information for the code in torch/csrc
. More information at:
By default, cmake will use its Makefile generator to generate your build
system. You can get faster builds if you install the ninja build system
with pip install ninja
. If PyTorch was already built, you will need
to run python setup.py clean
once after installing ninja for builds to
succeed.
Even when dependencies are tracked with file modification, there are many situations where files get rebuilt when a previous compilation was exactly the same.
Using ccache in a situation like this is a real time-saver. The ccache manual describes two ways to use ccache. In the PyTorch project, currently only the latter method of masquerading as the compiler via symlinks works for CUDA compilation.
Here are the instructions for installing ccache from source (tested at commit
7abac8f
of the ccache
repo):
# install and export ccache
if ! ls ~/ccache/bin/ccache
then
sudo apt-get update
sudo apt-get install -y automake autoconf
sudo apt-get install -y asciidoc
mkdir -p ~/ccache
pushd /tmp
rm -rf ccache
git clone https://github.com/ccache/ccache.git
pushd ccache
./autogen.sh
./configure
make install prefix=~/ccache
popd
popd
mkdir -p ~/ccache/lib
mkdir -p ~/ccache/cuda
ln -s ~/ccache/bin/ccache ~/ccache/lib/cc
ln -s ~/ccache/bin/ccache ~/ccache/lib/c++
ln -s ~/ccache/bin/ccache ~/ccache/lib/gcc
ln -s ~/ccache/bin/ccache ~/ccache/lib/g++
ln -s ~/ccache/bin/ccache ~/ccache/cuda/nvcc
~/ccache/bin/ccache -M 25Gi
fi
export PATH=~/ccache/lib:$PATH
export CUDA_NVCC_EXECUTABLE=~/ccache/cuda/nvcc
Alternatively, ccache
provided by newer Linux distributions (e.g. Debian/sid)
also works, but the nvcc
symlink to ccache
as described above is still required.
Note that the original nvcc
binary (typically at /usr/local/cuda/bin
) must
be on your PATH
, otherwise ccache
will emit the following error:
ccache: error: Could not find compiler "nvcc" in PATH
For example, here is how to install/configure ccache
on Ubuntu:
# install ccache
sudo apt install ccache
# update symlinks and create/re-create nvcc link
sudo /usr/sbin/update-ccache-symlinks
sudo ln -s /usr/bin/ccache /usr/lib/ccache/nvcc
# config: cache dir is ~/.ccache, conf file ~/.ccache/ccache.conf
# max size of cache
ccache -M 25Gi # -M 0 for unlimited
# unlimited number of files
ccache -F 0
# deploy (and add to ~/.bashrc for later)
export PATH="/usr/lib/ccache:$PATH"
It is also possible to install ccache
via conda
by installing it from the
community-maintained conda-forge
channel. Here is how to set up ccache
this
way:
# install ccache
conda install -c conda-forge ccache
# set up ccache compiler symlinks
mkdir ~/ccache
mkdir ~/ccache/lib
mkdir ~/ccache/cuda
ln -s $CONDA_PREFIX/bin/ccache ~/ccache/lib/cc
ln -s $CONDA_PREFIX/bin/ccache ~/ccache/lib/c++
ln -s $CONDA_PREFIX/bin/ccache ~/ccache/lib/gcc
ln -s $CONDA_PREFIX/bin/ccache ~/ccache/lib/g++
ln -s $CONDA_PREFIX/bin/ccache ~/ccache/cuda/nvcc
# update PATH to reflect symlink locations, consider
# adding this to your .bashrc
export PATH=~/ccache/lib:$PATH
export CUDA_NVCC_EXECUTABLE=~/ccache/cuda/nvcc
# increase ccache cache size to 25 GiB
ccache -M 25Gi
To check this is working, do two clean builds of pytorch in a row. The second build should be substantially and noticeably faster than the first build.
If you are editing a single file and rebuilding in a tight loop, the time spent
linking will dominate. The system linker available in most Linux distributions
(GNU ld
) is quite slow. Use a faster linker, like lld.
The easiest way to use lld
this is download the
latest LLVM binaries and run:
ln -s /path/to/downloaded/ld.lld /usr/local/bin/ld
We have very extensive tests in the test/cpp/api folder. The
tests are a great way to see how certain components are intended to be used.
When compiling PyTorch from source, the test runner binary will be written to
build/bin/test_api
. The tests use the GoogleTest
framework, which you can read up about to learn how to configure the test runner. When
submitting a new feature, we care very much that you write appropriate tests.
Please follow the lead of the other tests to see how to write a new test case.
If you are working on the CUDA code, here are some useful CUDA debugging tips:
CUDA_DEVICE_DEBUG=1
will enable CUDA device function debug symbols (-g -G
). This will be particularly helpful in debugging device code. However, it will slow down the build process for about 50% (compared to onlyDEBUG=1
), so use wisely.cuda-gdb
andcuda-memcheck
are your best CUDA debugging friends. Unlikegdb
,cuda-gdb
can display actual values in a CUDA tensor (rather than all zeros).- CUDA supports a lot of C++11/14 features such as,
std::numeric_limits
,std::nextafter
,std::tuple
etc. in device code. Many of such features are possible because of the --expt-relaxed-constexpr nvcc flag. There is a known issue that ROCm errors out on device code, which uses such stl functions. - A good performance metric for a CUDA kernel is the
Effective Memory Bandwidth.
It is useful for you to measure this metric whenever you are writing/optimizing a CUDA
kernel. Following script shows how we can measure the effective bandwidth of CUDA
uniform_
kernel.import torch import time size = 128*512 nrep = 100 nbytes_read_write = 4 # this is number of bytes read + written by a kernel. Change this to fit your kernel. for i in range(10): a=torch.Tensor(size).cuda().uniform_() torch.cuda.synchronize() start = time.time() # dry run to alloc out = a.uniform_() torch.cuda.synchronize() start = time.time() for i in range(nrep): out = a.uniform_() torch.cuda.synchronize() end = time.time() timec = (end-start)/nrep print("uniform, size, elements", size, "forward", timec, "bandwidth (GB/s)", size*(nbytes_read_write)*1e-9/timec) size *=2
Hope this helps, and thanks for considering to contribute.
For building from source on Windows, consult our documentation on it.
Occasionally, you will write a patch which works on Linux, but fails CI on Windows. There are a few aspects in which MSVC (the Windows compiler toolchain we use) is stricter than Linux, which are worth keeping in mind when fixing these problems.
-
Symbols are NOT exported by default on Windows; instead, you have to explicitly mark a symbol as exported/imported in a header file with
__declspec(dllexport)
/__declspec(dllimport)
. We have codified this pattern into a set of macros which follow the convention*_API
, e.g.,CAFFE2_API
inside Caffe2 and ATen. (Every separate shared library needs a unique macro name, because symbol visibility is on a per shared library basis. See c10/macros/Macros.h for more details.)The upshot is if you see an "unresolved external" error in your Windows build, this is probably because you forgot to mark a function with
*_API
. However, there is one important counterexample to this principle: if you want a templated function to be instantiated at the call site, do NOT mark it with*_API
(if you do mark it, you'll have to explicitly instantiate all of the specializations used by the call sites.) -
If you link against a library, this does not make its dependencies transitively visible. You must explicitly specify a link dependency against every library whose symbols you use. (This is different from Linux where in most environments, transitive dependencies can be used to fulfill unresolved symbols.)
-
If you have a Windows box (we have a few on EC2 which you can request access to) and you want to run the build, the easiest way is to just run
.jenkins/pytorch/win-build.sh
. If you need to rebuild, runREBUILD=1 .jenkins/pytorch/win-build.sh
(this will avoid blowing away your Conda environment.)
Even if you don't know anything about MSVC, you can use cmake to build simple programs on Windows; this can be helpful if you want to learn more about some peculiar linking behavior by reproducing it on a small example. Here's a simple example cmake file that defines two dynamic libraries, one linking with the other:
project(myproject CXX)
set(CMAKE_CXX_STANDARD 14)
add_library(foo SHARED foo.cpp)
add_library(bar SHARED bar.cpp)
# NB: don't forget to __declspec(dllexport) at least one symbol from foo,
# otherwise foo.lib will not be created.
target_link_libraries(bar PUBLIC foo)
You can build it with:
mkdir build
cd build
cmake ..
cmake --build .
The PyTorch codebase sometimes likes to use exciting C++ features, and these exciting features lead to exciting bugs in Windows compilers. To add insult to injury, the error messages will often not tell you which line of code actually induced the erroring template instantiation.
We've found the most effective way to debug these problems is to carefully read over diffs, keeping in mind known bugs in MSVC/NVCC. Here are a few well known pitfalls and workarounds:
-
This is not actually a bug per se, but in general, code generated by MSVC is more sensitive to memory errors; you may have written some code that does a use-after-free or stack overflows; on Linux the code might work, but on Windows your program will crash. ASAN may not catch all of these problems: stay vigilant to the possibility that your crash is due to a real memory problem.
-
(NVCC)
c10::optional
does not work when used from device code. Don't use it from kernels. Upstream issue: akrzemi1/Optional#58 and our local issue #10329. -
constexpr
generally works less well on MSVC.- The idiom
static_assert(f() == f())
to test iff
is constexpr does not work; you'll get "error C2131: expression did not evaluate to a constant". Don't use these asserts on Windows. (Example:c10/util/intrusive_ptr.h
)
- The idiom
-
(NVCC) Code you access inside a
static_assert
will eagerly be evaluated as if it were device code, and so you might get an error that the code is "not accessible".
class A {
static A singleton_;
static constexpr inline A* singleton() {
return &singleton_;
}
};
static_assert(std::is_same(A*, decltype(A::singleton()))::value, "hmm");
-
The compiler will run out of heap space if you attempt to compile files that are too large. Splitting such files into separate files helps. (Example:
THTensorMath
,THTensorMoreMath
,THTensorEvenMoreMath
.) -
MSVC's preprocessor (but not the standard compiler) has a bug where it incorrectly tokenizes raw string literals, ending when it sees a
"
. This causes preprocessor tokens inside the literal like an#endif
to be incorrectly treated as preprocessor directives. See https://godbolt.org/z/eVTIJq as an example. -
Either MSVC or the Windows headers have a PURE macro defined and will replace any occurrences of the PURE token in code with an empty string. This is why we have AliasAnalysisKind::PURE_FUNCTION and not AliasAnalysisKind::PURE. The same is likely true for other identifiers that we just didn't try to use yet.
Clang-Tidy is a C++ linter and static analysis tool based on the clang compiler. We run clang-tidy in our CI to make sure that new C++ code is safe, sane and efficient. See our .travis.yml file for the simple commands we use for this.
To run clang-tidy locally, follow these steps:
-
Install clang-tidy. First, check if you already have clang-tidy by simply writing
clang-tidy
in your terminal. If you don't yet have clang-tidy, you should be able to install it easily with your package manager, e.g. by writingapt-get install clang-tidy
on Ubuntu. See https://apt.llvm.org for details on how to install the latest version. Note that newer versions of clang-tidy will have more checks than older versions. In our CI, we run clang-tidy-6.0. -
Use our driver script to run clang-tidy over any changes relative to some git revision (you may want to replace
HEAD~1
withHEAD
to pick up uncommitted changes). Changes are picked up based on agit diff
with the given revision:
python tools/clang_tidy.py -d build -p torch/csrc --diff 'HEAD~1'
Above, it is assumed you are in the PyTorch root folder. path/to/build
should
be the path to where you built PyTorch from source, e.g. build
in the PyTorch
root folder if you used setup.py build
. You can use -c <clang-tidy-binary>
to change the clang-tidy this script uses. Make sure you have PyYaml installed,
which is in PyTorch's requirements.txt
.
We use clang-tidy and flake8 (installed with flake8-bugbear, flake8-comprehensions, flake8-mypy, and flake8-pyi) to perform additional formatting and semantic checking of code. We provide a pre-commit git hook for performing these checks, before a commit is created:
ln -s ../../tools/git-pre-commit .git/hooks/pre-commit
You'll need to install an appropriately configured flake8; see Lint as you type for documentation on how to do this.
ASAN is very useful for debugging memory errors in C++. We run it in CI, but here's how to get the same thing to run on your local machine.
First, install LLVM 8. The easiest way is to get prebuilt
binaries and extract them to
folder (later called $LLVM_ROOT
).
Then set up the appropriate scripts. You can put this in your .bashrc
:
LLVM_ROOT=<wherever your llvm install is>
PYTORCH_ROOT=<wherever your pytorch checkout is>
LIBASAN_RT="$LLVM_ROOT/lib/clang/8.0.0/lib/linux/libclang_rt.asan-x86_64.so"
build_with_asan()
{
LD_PRELOAD=${LIBASAN_RT} \
CC="$LLVM_ROOT/bin/clang" \
CXX="$LLVM_ROOT/bin/clang++" \
LDSHARED="clang --shared" \
LDFLAGS="-stdlib=libstdc++" \
CFLAGS="-fsanitize=address -fno-sanitize-recover=all -shared-libasan -pthread" \
CXX_FLAGS="-pthread" \
NO_CUDA=1 USE_OPENMP=0 BUILD_CAFFE2_OPS=0 NO_DISTRIBUTED=1 DEBUG=1 \
python setup.py develop
}
run_with_asan()
{
LD_PRELOAD=${LIBASAN_RT} $@
}
# you can look at build-asan.sh to find the latest options the CI uses
export ASAN_OPTIONS=detect_leaks=0:symbolize=1:strict_init_order=true
export UBSAN_OPTIONS=print_stacktrace=1:suppressions=$PYTORCH_ROOT/ubsan.supp
export ASAN_SYMBOLIZER_PATH=$LLVM_ROOT/bin/llvm-symbolizer
Then you can use the scripts like:
suo-devfair ~/pytorch ❯ build_with_asan
suo-devfair ~/pytorch ❯ run_with_asan python test/test_jit.py
The scripts above specify the clang
and clang++
binaries directly, which
bypasses ccache
. Here's how to get ccache
to work:
- Make sure the ccache symlinks for
clang
andclang++
are set up (see CONTRIBUTING.md) - Make sure
$LLVM_ROOT/bin
is available on your$PATH
. - Change the
CC
andCXX
variables inbuild_with_asan()
to point directly toclang
andclang++
.
The “standard” workflow for ASAN assumes you have a standalone binary:
- Recompile your binary with
-fsanitize=address
. - Run the binary, and ASAN will report whatever errors it find.
Unfortunately, PyTorch is a distributed as a shared library that is loaded by a third-party executable (Python). It’s too much of a hassle to recompile all of Python every time we want to use ASAN. Luckily, the ASAN folks have a workaround for cases like this:
- Recompile your library with
-fsanitize=address -shared-libasan
. The extra-shared-libasan
tells the compiler to ask for the shared ASAN runtime library. - Use
LD_PRELOAD
to tell the dynamic linker to load the ASAN runtime library before anything else.
More information can be found here.
We need LD_PRELOAD
because there is a cmake check that ensures that a
simple program builds and runs. If we are building with ASAN as a shared
library, we need to LD_PRELOAD
the runtime library, otherwise there will
dynamic linker errors and the check will fail.
We don’t actually need either of these if we fix the cmake checks.
Python leaks a lot of memory. Possibly we could configure a suppression file, but we haven’t gotten around to it.
In 2018, we merged Caffe2 into the PyTorch source repository. While the steady state aspiration is that Caffe2 and PyTorch share code freely, in the meantime there will be some separation.
If you submit a PR to only PyTorch or only Caffe2 code, CI will only
run for the project you edited. The logic for this is implemented
in .jenkins/pytorch/dirty.sh
and .jenkins/caffe2/dirty.sh
; you
can look at this to see what path prefixes constitute changes.
This also means if you ADD a new top-level path, or you start
sharing code between projects, you need to modify these files.
There are a few "unusual" directories which, for historical reasons, are Caffe2/PyTorch specific. Here they are:
-
CMakeLists.txt
,Makefile
,binaries
,cmake
,conda
,modules
,scripts
are Caffe2-specific. Don't put PyTorch code in them without extra coordination. -
mypy*
,requirements.txt
,setup.py
,test
,tools
are PyTorch-specific. Don't put Caffe2 code in them without extra coordination.