Skip to content

Commit

Permalink
Re-organize documentation around docs (pytorch#2258)
Browse files Browse the repository at this point in the history 
Summary:
- Clean up the scripts for handling platform name when building packages
- Clarify and re-organize the docs around documenting the codebase

Pull Request resolved: pytorch#2258

Test Plan: https://deploy-preview-2258--pytorch-fbgemm-docs.netlify.app/

Reviewed By: jianyuh

Differential Revision: D52614741

Pulled By: q10

fbshipit-source-id: 2e3bb4d2d54b2233d639971421cfa637150d5805
  • Loading branch information
@q10 @facebook-github-bot
q10 authored and facebook-github-bot committed Jan 9, 2024
1 parent 0a23158 commit 127474c
Show file tree
Hide file tree
Showing 11 changed files with 602 additions and 458 deletions.
134 changes: 92 additions & 42 deletions .github/scripts/fbgemm_gpu_build.bash
View file
Original file line number Diff line number Diff line change
Expand Up @@ -194,6 +194,76 @@ __configure_fbgemm_gpu_build () {
echo "[BUILD] FBGEMM_GPU build arguments have been set: ${build_args[@]}"
}

__build_fbgemm_gpu_set_package_name () {
# Determine the package name based on release type and variant
export package_name="fbgemm_gpu"

# Append qualifiers for the non-release version
if [ "$fbgemm_release_type" != "release" ]; then
export package_name="${package_name}_${fbgemm_release_type}"
fi

# Append cpu or rocm for the non-CUDA case
if [ "$fbgemm_variant" == "cpu" ]; then
export package_name="${package_name}-cpu"
elif [ "$fbgemm_variant" == "rocm" ]; then
export package_name="${package_name}-rocm"
fi

echo "[BUILD] Determined and set Python package name to use: ${package_name}"
}

__build_fbgemm_gpu_set_python_tag () {
# shellcheck disable=SC2207,SC2086
local python_version=($(conda run --no-capture-output ${env_prefix} python --version))

# shellcheck disable=SC2206
local python_version_arr=(${python_version[1]//./ })

# Set the python tag (e.g. Python 3.12 -> py312)
export python_tag="py${python_version_arr[0]}${python_version_arr[1]}"
echo "[BUILD] Extracted and set Python tag: ${python_tag}"
}

__build_fbgemm_gpu_set_python_plat_name () {
if [[ $KERN_NAME == 'Darwin' ]]; then
# This follows PyTorch package naming conventions
# See https://pypi.org/project/torch/#files
if [[ $MACHINE_NAME == 'arm64' ]]; then
export python_plat_name="macosx_11_0_${MACHINE_NAME}"
else
export python_plat_name="macosx_10_9_${MACHINE_NAME}"
fi

elif [[ $KERN_NAME == 'Linux' ]]; then
# manylinux2014 is specified, bc manylinux1 does not support aarch64
# See https://github.com/pypa/manylinux
export python_plat_name="manylinux2014_${MACHINE_NAME}"

else
echo "[BUILD] Unsupported OS platform: ${KERN_NAME}"
return 1
fi

echo "[BUILD] Extracted and set Python platform name: ${python_plat_name}"
}

__build_fbgemm_gpu_set_run_multicore () {
# shellcheck disable=SC2155
local core=$(lscpu | grep "Core(s)" | awk '{print $NF}') && echo "core = ${core}" || echo "core not found"
# shellcheck disable=SC2155
local sockets=$(lscpu | grep "Socket(s)" | awk '{print $NF}') && echo "sockets = ${sockets}" || echo "sockets not found"
local re='^[0-9]+$'

export run_multicore=""
if [[ $core =~ $re && $sockets =~ $re ]] ; then
local n_core=$((core * sockets))
export run_multicore=" -j ${n_core}"
fi

echo "[BUILD] Set multicore run option for setup.py: ${run_multicore}"
}

__build_fbgemm_gpu_common_pre_steps () {
# Private function that uses variables instantiated by its caller

Expand All @@ -203,28 +273,23 @@ __build_fbgemm_gpu_common_pre_steps () {
(test_binpath "${env_name}" c++) || return 1
(test_binpath "${env_name}" g++) || return 1

# Determine the package name based on release type and variant
package_name="fbgemm_gpu"
if [ "$fbgemm_release_type" != "release" ]; then
package_name="${package_name}_${fbgemm_release_type}"
# Set the default the FBGEMM_GPU variant to be CUDA
if [ "$fbgemm_variant" != "cpu" ] && [ "$fbgemm_variant" != "rocm" ]; then
export fbgemm_variant="cuda"
fi
if [ "$fbgemm_variant" == "cpu" ]; then
package_name="${package_name}-cpu"
elif [ "$fbgemm_variant" == "rocm" ]; then
package_name="${package_name}-rocm"
else
# Set to the default variant
fbgemm_variant="cuda"
fi
echo "[BUILD] Determined Python package name to use: ${package_name}"

# Extract the Python tag
# shellcheck disable=SC2207,SC2086
python_version=($(conda run --no-capture-output ${env_prefix} python --version))
# shellcheck disable=SC2206
python_version_arr=(${python_version[1]//./ })
python_tag="py${python_version_arr[0]}${python_version_arr[1]}"
echo "[BUILD] Extracted Python tag: ${python_tag}"
# Extract and set the package name given the FBGEMM_GPU variant
__build_fbgemm_gpu_set_package_name

# Extract and set the Python tag
__build_fbgemm_gpu_set_python_tag

# Extract and set the platform name
__build_fbgemm_gpu_set_python_plat_name

# Set multicore run option for setup.py if the number of cores on the machine
# permit for this
__build_fbgemm_gpu_set_run_multicore

echo "[BUILD] Running pre-build cleanups ..."
print_exec rm -rf dist
Expand Down Expand Up @@ -332,33 +397,14 @@ build_fbgemm_gpu_package () {
echo "################################################################################"
echo ""

# manylinux2014 is specified, bc manylinux1 does not support aarch64
# See https://github.com/pypa/manylinux
local plat_name="manylinux2014_${MACHINE_NAME}"

echo "[BUILD] Checking arch_list = ${arch_list}"
echo "[BUILD] Checking build_args:"
echo "${build_args[@]}"

# shellcheck disable=SC2155
local core=$(lscpu | grep "Core(s)" | awk '{print $NF}') && echo "core = ${core}" || echo "core not found"
# shellcheck disable=SC2155
local sockets=$(lscpu | grep "Socket(s)" | awk '{print $NF}') && echo "sockets = ${sockets}" || echo "sockets not found"
local re='^[0-9]+$'
local run_multicore=""
if [[ $core =~ $re && $sockets =~ $re ]] ; then
local n_core=$((core * sockets))
local run_multicore=" -j ${n_core}"
fi

# Distribute Python extensions as wheels on Linux
echo "[BUILD] Building FBGEMM-GPU wheel (VARIANT=${fbgemm_variant}) ..."
# shellcheck disable=SC2086
print_exec conda run --no-capture-output ${env_prefix} \
python setup.py "${run_multicore}" bdist_wheel \
--package_name="${package_name}" \
--python-tag="${python_tag}" \
--plat-name="${plat_name}" \
--plat-name="${python_plat_name}" \
--verbose \
"${build_args[@]}"

Expand Down Expand Up @@ -410,7 +456,9 @@ build_fbgemm_gpu_install () {
echo "[BUILD] Building + installing FBGEMM-GPU (VARIANT=${fbgemm_variant}) ..."
# shellcheck disable=SC2086
print_exec conda run --no-capture-output ${env_prefix} \
python setup.py install "${build_args[@]}"
python setup.py "${run_multicore}" install \
--verbose \
"${build_args[@]}"

# Run checks on the built libraries
(run_fbgemm_gpu_postbuild_checks "${fbgemm_variant}") || return 1
Expand Down Expand Up @@ -460,7 +508,9 @@ build_fbgemm_gpu_develop () {
echo "[BUILD] Building (develop) FBGEMM-GPU (VARIANT=${fbgemm_variant}) ..."
# shellcheck disable=SC2086
print_exec conda run --no-capture-output ${env_prefix} \
python setup.py build develop "${build_args[@]}"
python setup.py "${run_multicore}" build develop \
--verbose \
"${build_args[@]}"

# Run checks on the built libraries
(run_fbgemm_gpu_postbuild_checks "${fbgemm_variant}") || return 1
Expand Down
58 changes: 35 additions & 23 deletions fbgemm_gpu/docs/src/fbgemm_gpu-development/BuildInstructions.rst
View file
Original file line number Diff line number Diff line change
Expand Up @@ -380,6 +380,33 @@ build cache:
python setup.py clean
Set Wheel Build Variables
~~~~~~~~~~~~~~~~~~~~~~~~~

When building out the Python wheel, the package name, Python version tag, and
Python platform name must first be properly set:

.. code:: sh
# Set the package name depending on the build variant
export package_name=fbgemm_gpu_{cpu, cuda, rocm}
# Set the Python version tag. It should follow the convention `py<major><minor>`,
# e.g. Python 3.12 -> py312
export python_tag=py312
# Determine the processor architecture
export ARCH=$(uname -m)
# Set the Python platform name for the Linux case
export python_plat_name="manylinux2014_${ARCH}"
# For the macOS (x86_64) case
export python_plat_name="macosx_10_9_${ARCH}"
# For the macOS (arm64) case
export python_plat_name="macosx_11_0_${ARCH}"
# For the Windows case
export python_plat_name="win_${ARCH}"
.. _fbgemm-gpu.build.process.cpu:

CPU-Only Build
Expand All @@ -391,20 +418,16 @@ For CPU-only builds, the ``--cpu_only`` flag needs to be specified.
# !! Run in fbgemm_gpu/ directory inside the Conda environment !!
export ARCH=$(uname -m)
python_tag=py310
package_name=fbgemm_gpu_cpu
# Build the wheel artifact only
python setup.py bdist_wheel \
--package_name="${package_name}" \
--package_variant=cpu \
--package_name="${package_name}" \
--python-tag="${python_tag}" \
--plat-name="manylinux1_${ARCH}"
--plat-name="${python_plat_name}"
# Build and install the library into the Conda environment
python setup.py install --package_variant=cpu
python setup.py install \
--package_variant=cpu
.. _fbgemm-gpu.build.process.cuda:

Expand All @@ -419,9 +442,6 @@ CUDA device, however, is not required for building the package.
# !! Run in fbgemm_gpu/ directory inside the Conda environment !!
# Determine the processor architecture
export ARCH=$(uname -m)
# [OPTIONAL] Specify the CUDA installation paths
# This may be required if CMake is unable to find nvcc
export CUDACXX=/path/to/nvcc
Expand All @@ -437,10 +457,6 @@ CUDA device, however, is not required for building the package.
# Specify NVML path
export NVML_LIB_PATH=/path/to/libnvidia-ml.so
# Update to reflect the version of Python in the Conda environment
python_tag=py310
package_name=fbgemm_gpu
# Build for SM70/80 (V100/A100 GPU); update as needed
# If not specified, only the CUDA architecture supported by current system will be targeted
# If not specified and no CUDA device is present either, all CUDA architectures will be targeted
Expand All @@ -452,10 +468,10 @@ CUDA device, however, is not required for building the package.
# Build the wheel artifact only
python setup.py bdist_wheel \
--package_name="${package_name}" \
--package_variant=cuda \
--package_name="${package_name}" \
--python-tag="${python_tag}" \
--plat-name="manylinux1_${ARCH}" \
--plat-name="${python_plat_name}" \
--nvml_lib_path=${NVML_LIB_PATH} \
-DTORCH_CUDA_ARCH_LIST="${cuda_arch_list}"
Expand All @@ -478,22 +494,18 @@ the package.
# !! Run in fbgemm_gpu/ directory inside the Conda environment !!
export ARCH=$(uname -m)
export ROCM_PATH=/path/to/rocm
# Build for the target architecture of the ROCm device installed on the machine (e.g. 'gfx906;gfx908;gfx90a')
# See https://wiki.gentoo.org/wiki/ROCm for list
export PYTORCH_ROCM_ARCH=$(${ROCM_PATH}/bin/rocminfo | grep -o -m 1 'gfx.*')
python_tag=py310
package_name=fbgemm_gpu_rocm
# Build the wheel artifact only
python setup.py bdist_wheel \
--package_name="${package_name}" \
--package_variant=rocm \
--package_name="${package_name}" \
--python-tag="${python_tag}" \
--plat-name="manylinux1_${ARCH}" \
--plat-name="${python_plat_name}" \
-DHIP_ROOT_DIR="${ROCM_PATH}" \
-DCMAKE_C_FLAGS="-DTORCH_USE_HIP_DSA" \
-DCMAKE_CXX_FLAGS="-DTORCH_USE_HIP_DSA"
Expand Down
Loading

0 comments on commit 127474c

Please sign in to comment.