Skip to content

Commit

Permalink
Enhance FBGEMM nightly CI (pytorch#1407)
Browse files Browse the repository at this point in the history
Summary:
This PR updates FBGEMM nightly CI to make it more usable and maintainable by using GitHub Actions more effectively.

## 1. Implement label-triggered wheel tests during a PR review

Previously, we needed to apply a hack to trigger wheel-related tests before merging the PR, which is neither comprehensive nor efficient.

To address this issue, this PR adds optional wheel-related tests.  Now a PR with a label such as `"test_wheel_nightly"`, `"test_wheel_prerelease"`, and `"test_wheel_release"` enables the corresponding wheel-related tests.  We don't need to modify YML files again, so we can easily run these tests.  Please trigger these tests for suspicious PRs that touch the installation logic.  Note that binaries won't be pushed to PYPI if this method is used.

## 2. Remove duplicated logic across YML files

Nightly/release/cpu/gpu wheel scripts shared most in common, but the logic were scattered across different files, which significantly lowered maintainability.

To address this issue, this PR collects all the wheel-related test logic into a single file (`build_wheel.yml`) and makes the callers (e.g., scheduled nightly trigger or per-PR tests using labels) pass flags to control the flow (e.g., per-PR tests do not push the wheel binary to PYPI).  No duplication improves maintainability.

<img width="1269" alt="Screen Shot 2022-10-21 at 9 58 56 AM" src="https://user-images.githubusercontent.com/15073003/197249555-bb00f3b1-9bc9-46f1-8fea-7874460723f6.png">

## 3. Support handy wheel-related tests on a local machine

Previously, the core wheel-related build/test logics were embedded into GitHub workflow files (`*.yml`) mixed with AWS-specific commands, so it was very tedious to test the nightly logic on a local machine, lowering the developer efficiency.

To address this issue, this PR extracts core scripts from GitHub workflow files and makes them standalone so that the developer can try wheel-build locally (i.e., without access to AWS machine). It uses conda to create a virtual software environment, so it should be handy enough in many cases though this is not as robust as a container-based solution.

```sh
# For example, check prerelease PyTorch (pytorch-test package) locally.
git clone https://github.com/pytorch/FBGEMM.git
cd FBGEMM
git submodule update --init --recursive
bash .github/scripts/build_wheel.bash -p 3.10 -c 11.7 -v -P pytorch-test -o fbgemm_gpu_test
bash .github/scripts/test_wheel.bash -p 3.10 -c 11.7 -v -P pytorch-test -w fbgemm_gpu/dist/fbgemm_gpu_test-2022.10.20-cp310-cp310-manylinux1_x86_64.whl
git clone https://github.com/pytorch/torchrec.git
cd torchrec
git submodule update --init --recursive
bash ../.github/scripts/test_torchrec.bash -o torchrec_nightly -p 3.10 -c 11.7 -v -P pytorch-test -w ../fbgemm_gpu/dist/fbgemm_gpu_test-2022.10.20-cp310-cp310-manylinux1_x86_64.whl
```

## 4. [Temporary] Add TorchRec integration test

TorchRec is one of our most important users, but we didn't test TorchRec integration in the nightly CI, so the changes in FBGEMM sometimes surprised the TorchRec developers.

To address this issue, though it is temporary, but this PR adds a TorchRec integration test before pushing a binary to PYPI so that we can make sure that FBGEMM works with TorchRec.  However, now broken TorchRec can break FBGEMM-nightly; we will investigate a more sustainable solution that makes everyone happy.

## 5. Better interface to manually trigger the CI

It looks nice. I love to operate the release management in an automated way as much as possible to reduce human errors at the very last minute. In the future, we can add more options if needed.

<img width="844" alt="Screen Shot 2022-10-21 at 5 53 03 AM" src="https://user-images.githubusercontent.com/15073003/197240190-e808df23-795a-451d-a7d9-ec1c43c68e92.png">

Pull Request resolved: pytorch#1407

Differential Revision: D40597700

Pulled By: shintaro-iwasaki

fbshipit-source-id: 1b49b09a5f724ccfeb2e86fc0848e3f1dca0ef3c
  • Loading branch information
shintaro-iwasaki authored and facebook-github-bot committed Oct 28, 2022
1 parent d8de717 commit e011384
Show file tree
Hide file tree
Showing 11 changed files with 659 additions and 891 deletions.
126 changes: 126 additions & 0 deletions .github/scripts/build_wheel.bash
Original file line number Diff line number Diff line change
@@ -0,0 +1,126 @@
#!/bin/bash

# Exit on failure
set -e

# shellcheck source=/dev/null
. "$(dirname "$(realpath -s "$0")")/setup_env.bash"

verbose=0
package_name=""
python_version=""
pytorch_channel_name=""
pytorch_cuda_version="x"
miniconda_prefix="${HOME}/miniconda"

usage () {
echo "Usage: bash build_wheel.bash -o PACKAGE_NAME -p PYTHON_VERSION -P PYTORCH_CHANNEL_NAME -c PYTORCH_CUDA_VERSION [-m MINICONDA_PREFIX] [-v] [-h]"
echo "-v : verbose"
echo "-h : help"
echo "PACKAGE_NAME : output package name (e.g., fbgemm_gpu_nightly)"
echo "PYTHON_VERSION : Python version (e.g., 3.7, 3.8, 3.10)"
echo "PYTORCH_CHANNEL_NAME: PyTorch's channel name (e.g., pytorch-nightly, pytorch-test (=pre-release), pytorch (=stable release))"
echo "PYTORCH_CUDA_VERSION: PyTorch's CUDA version (e.g., 11.6, 11.7)"
echo "MINICONDA_PREFIX : path to install Miniconda (default: \$HOME/miniconda)"
echo "Example 1: Python 3.10 + PyTorch nightly (CUDA 11.7), install miniconda at /home/user/tmp/miniconda"
echo " bash build_wheel.bash -v -P pytorch-nightly -p 3.10 -c 11.7 -m /home/user/tmp/miniconda"
echo "Example 2: Python 3.10 + PyTorch stable (CPU), install miniconda at \$HOME/miniconda"
echo " bash build_wheel.bash -v -P pytorch -p 3.10 -c \"\""
}

while getopts vfho:p:P:c:m: flag
do
case "$flag" in
v) verbose="1";;
o) package_name="${OPTARG}";;
p) python_version="${OPTARG}";;
P) pytorch_channel_name="${OPTARG}";;
c) pytorch_cuda_version="${OPTARG}";;
m) miniconda_prefix="${OPTARG}";;
h) usage
exit 0;;
*) usage
exit 1;;
esac
done

if [ "$python_version" == "" ] || [ "$pytorch_cuda_version" == "x" ] || [ "$miniconda_prefix" == "" ] || [ "$pytorch_channel_name" == "" ] || [ "$package_name" == "" ]; then
usage
exit 1
fi
python_tag="${python_version//\./}"

if [ "$verbose" == "1" ]; then
# Print each line verbosely
set -x -e
fi

################################################################################
echo "## 0. Minimal check"
################################################################################

if [ ! -d "fbgemm_gpu" ]; then
echo "Error: this script must be executed in FBGEMM/"
exit 1
elif [ "$(which gcc 2>/dev/null)" == "" ]; then
echo "Error: GCC is needed to compile FBGEMM"
exit 1
fi

################################################################################
echo "## 1. Set up Miniconda"
################################################################################

setup_miniconda "$miniconda_prefix"

################################################################################
echo "## 2. Create build_binary environment"
################################################################################

create_conda_environment build_binary "$python_version" "$pytorch_channel_name" "$pytorch_cuda_version"

cd fbgemm_gpu

# cuDNN is needed to "build" FBGEMM
install_cudnn "$miniconda_prefix/build_only/cudnn"
export CUDNN_INCLUDE_DIR="$miniconda_prefix/build_only/cudnn/include"
export CUDNN_LIBRARY="$miniconda_prefix/build_only/cudnn/lib"

conda run -n build_binary python -m pip install -r requirements.txt

# TODO: Do we need these checks?
ldd --version
conda info
conda run -n build_binary python --version
gcc --version
conda run -n build_binary python -c "import torch.distributed"
conda run -n build_binary python -c "import skbuild"
conda run -n build_binary python -c "import numpy"
cd ../

################################################################################
echo "## 3. Build FBGEMM_GPU"
################################################################################

cd fbgemm_gpu
rm -rf dist _skbuild
if [ "$pytorch_cuda_version" == "" ]; then
# CPU version
build_arg="--cpu_only"
package_name="${package_name}_cpu"
else
# GPU version
# We build only CUDA 7.0 and 8.0 (i.e., for v100 and a100) because of 100 MB binary size limit from PYPI website.
build_arg="-DTORCH_CUDA_ARCH_LIST=7.0;8.0"
fi

# manylinux1_x86_64 is specified for pypi upload: distribute python extensions as wheels on Linux
conda run -n build_binary python setup.py bdist_wheel --package_name="${package_name}" --python-tag="py${python_tag}" "${build_arg}" --plat-name=manylinux1_x86_64
cd ../

# Usage:
# pip install $(ls fbgemm_gpu/dist/${package_name}-*.whl)
# python -c "import fbgemm_gpu"

wheel_name="$(ls fbgemm_gpu/dist/"${package_name}"-*.whl)"
echo "Successfully built $wheel_name"
68 changes: 68 additions & 0 deletions .github/scripts/setup_env.bash
Original file line number Diff line number Diff line change
@@ -0,0 +1,68 @@
#!/bin/bash

setup_miniconda () {
miniconda_prefix="$1"
if [ "$miniconda_prefix" == "" ]; then
echo "Usage: setup_miniconda MINICONDA_PREFIX_PATH"
echo "Example:"
echo " setup_miniconda /home/user/tmp/miniconda"
exit 1
fi
if [ ! -f "${miniconda_prefix}/bin/conda" ]; then
# Download miniconda if not exists
mkdir -p "$miniconda_prefix"
wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh -O miniconda.sh
bash miniconda.sh -b -p "$miniconda_prefix" -u
fi
# These variables will be exported outside
export PATH="${miniconda_prefix}/bin:${PATH}"
export CONDA="${miniconda_prefix}"
}

create_conda_environment () {
env_name="$1"
python_version="$2"
pytorch_channel_name="$3"
pytorch_cuda_version="$4"
if [ "$python_version" == "" ]; then
echo "Usage: create_conda_environment ENV_NAME PYTHON_VERSION PYTORCH_CHANNEL_NAME PYTORCH_CUDA_VERSION"
echo "Example:"
echo " create_conda_environment build_binary 3.10 pytorch-nightly 11.7"
exit 1
fi
# -y removes existing environment
conda create -y --name "$env_name" python="$python_version"
if [ "$pytorch_cuda_version" == "" ]; then
# CPU version
conda install -n "$env_name" -y pytorch cpuonly -c "$pytorch_channel_name"
else
# GPU version
conda install -n "$env_name" -y pytorch pytorch-cuda="$pytorch_cuda_version" -c "$pytorch_channel_name" -c nvidia
fi
}

install_cudnn () {
install_path="$1"
if [ "$install_path" == "" ]; then
echo "Usage: install_cudnn INSTALL_PATH"
echo "Example:"
echo " install_cudnn \$(pwd)/cudnn_install"
exit 1
fi

rm -rf "$install_path"
mkdir -p "$install_path"

# Install cuDNN manually
# See https://github.com/pytorch/builder/blob/main/common/install_cuda.sh
mkdir -p tmp_cudnn
cd tmp_cudnn || exit
wget -q https://ossci-linux.s3.amazonaws.com/cudnn-linux-x86_64-8.5.0.96_cuda11-archive.tar.xz -O cudnn-linux-x86_64-8.5.0.96_cuda11-archive.tar.xz
tar xf cudnn-linux-x86_64-8.5.0.96_cuda11-archive.tar.xz
rm -rf "${install_path:?}/include"
rm -rf "${install_path:?}/lib"
mv cudnn-linux-x86_64-8.5.0.96_cuda11-archive/include "$install_path"
mv cudnn-linux-x86_64-8.5.0.96_cuda11-archive/lib "$install_path"
cd ../
rm -rf tmp_cudnn
}
102 changes: 102 additions & 0 deletions .github/scripts/test_torchrec.bash
Original file line number Diff line number Diff line change
@@ -0,0 +1,102 @@
#!/bin/bash

# Exit on failure
set -e

# shellcheck source=/dev/null
. "$(dirname "$(realpath -s "$0")")/setup_env.bash"

verbose=0
torchrec_package_name=""
python_version=""
pytorch_cuda_version="x"
fbgemm_wheel_path="x"
miniconda_prefix="${HOME}/miniconda"

usage () {
echo "Usage: bash test_torchrec.bash -o PACKAGE_NAME -p PYTHON_VERSION -P PYTORCH_CHANNEL_NAME -c PYTORCH_CUDA_VERSION -w FBGEMM_WHEEL_PATH [-m MINICONDA_PREFIX] [-v] [-h]"
echo "-v : verbose"
echo "-h : help"
echo "PACKAGE_NAME : output package name of TorchRec (e.g., torchrec_nightly)"
echo " Note: TorchRec is sensitive to its package name"
echo " e.g., torchrec needs fbgemm-gpu while torchrec_nightly needs fbgemm-gpu-nightly"
echo "PYTHON_VERSION : Python version (e.g., 3.7, 3.8, 3.10)"
echo "PYTORCH_CHANNEL_NAME: PyTorch's channel name (e.g., pytorch-nightly, pytorch-test (=pre-release), pytorch (=stable release))"
echo "PYTORCH_CUDA_VERSION: PyTorch's CUDA version (e.g., 11.6, 11.7)"
echo "FBGEMM_WHEEL_PATH : path to FBGEMM_GPU's wheel file"
echo "MINICONDA_PREFIX : path to install Miniconda (default: \$HOME/miniconda)"
echo "Example: Python 3.10 + PyTorch nightly (CUDA 11.7), install miniconda at \$HOME/miniconda, using dist/fbgemm_gpu_nightly.whl"
echo " bash test_torchrec.bash -v -o torchrec_nightly -p 3.10 -P pytorch-nightly -c 11.7 -w dist/fbgemm_gpu_nightly.whl"
}

while getopts vho:p:P:c:m:w: flag
do
case "$flag" in
v) verbose="1";;
o) torchrec_package_name="${OPTARG}";;
p) python_version="${OPTARG}";;
P) pytorch_channel_name="${OPTARG}";;
c) pytorch_cuda_version="${OPTARG}";;
m) miniconda_prefix="${OPTARG}";;
w) fbgemm_wheel_path="${OPTARG}";;
h) usage
exit 0;;
*) usage
exit 1;;
esac
done

if [ "$torchrec_package_name" == "" ] || [ "$python_version" == "" ] || [ "$pytorch_cuda_version" == "x" ] || [ "$miniconda_prefix" == "" ] || [ "$pytorch_channel_name" == "" ] || [ "$fbgemm_wheel_path" == "" ]; then
usage
exit 1
fi
python_tag="${python_version//\./}"

if [ "$verbose" == "1" ]; then
# Print each line verbosely
set -x -e
fi

################################################################################
echo "## 0. Minimal check"
################################################################################

if [ ! -d "torchrec" ]; then
echo "Error: this script must be executed in torchrec/"
exit 1
fi

################################################################################
echo "## 1. Set up Miniconda"
################################################################################

setup_miniconda "$miniconda_prefix"

################################################################################
echo "## 2. Create test_binary environment"
################################################################################

create_conda_environment test_binary "$python_version" "$pytorch_channel_name" "$pytorch_cuda_version"

# Comment out FBGEMM_GPU since we will install it from "$fbgemm_wheel_path"
sed -i 's/fbgemm-gpu/#fbgemm-gpu/g' requirements.txt
conda run -n test_binary python -m pip install -r requirements.txt
# Install FBGEMM_GPU from a local wheel file.
conda run -n test_binary python -m pip install "$fbgemm_wheel_path"
conda run -n test_binary python -c "import fbgemm_gpu"

################################################################################
echo "## 3. Build TorchRec"
################################################################################

rm -rf dist
conda run -n test_binary python setup.py bdist_wheel --package_name "${torchrec_package_name}" --python-tag="py${python_tag}"

################################################################################
echo "## 4. Import TorchRec"
################################################################################

conda run -n test_binary python -m pip install dist/"${torchrec_package_name}"*.whl
conda run -n test_binary python -c "import torchrec"

echo "Test succeeded"
Loading

0 comments on commit e011384

Please sign in to comment.