Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Remove pytorch custom build and move to pytorch repository #2405

Merged
merged 2 commits into from
Feb 3, 2021
Merged

Remove pytorch custom build and move to pytorch repository #2405

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
f99dc21
Remove custom build of PyTorch
CoderHam Jan 8, 2021
33f471b
Additional pytorch build related cleanup
CoderHam Jan 8, 2021
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
113 changes: 11 additions & 102 deletions build.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -286,7 +286,7 @@ def backend_cmake_args(images, components, be, install_dir):
elif be == 'dali':
args = dali_cmake_args()
elif be == 'pytorch':
args = pytorch_cmake_args()
args = pytorch_cmake_args(images)
elif be in EXAMPLE_BACKENDS:
args = []
else:
Expand All @@ -313,10 +313,14 @@ def backend_cmake_args(images, components, be, install_dir):
return cargs


def pytorch_cmake_args():
def pytorch_cmake_args(images):
if "pytorch" in images:
image = images["pytorch"]
else:
image = 'nvcr.io/nvidia/pytorch:{}-py3'.format(
FLAGS.upstream_container_version)
return [
'-DTRITON_PYTORCH_INCLUDE_PATHS=/opt/tritonserver/include/torch;/opt/tritonserver/include/torch/torch/csrc/api/include;/opt/tritonserver/include/torchvision',
'-DTRITON_PYTORCH_LIB_PATHS=/opt/tritonserver/backends/pytorch'
'-DTRITON_PYTORCH_DOCKER_IMAGE={}'.format(image),
]


Expand Down Expand Up @@ -376,10 +380,6 @@ def create_dockerfile_buildbase(ddir, dockerfile_name, argmap, backends):
ARG BASE_IMAGE={}
'''.format(argmap['TRITON_VERSION'], argmap['TRITON_CONTAINER_VERSION'],
argmap['BASE_IMAGE'])
if 'pytorch' in backends:
df += '''
ARG PYTORCH_IMAGE={}
'''.format(argmap['PYTORCH_IMAGE'])
if 'onnxruntime' in backends:
df += '''
ARG ONNX_RUNTIME_VERSION={}
Expand All @@ -389,34 +389,6 @@ def create_dockerfile_buildbase(ddir, dockerfile_name, argmap, backends):
ARG ONNX_RUNTIME_OPENVINO_VERSION={}
'''.format(argmap['ONNX_RUNTIME_OPENVINO_VERSION'])

if 'pytorch' in backends:
df += '''
############################################################################
## PyTorch stage: Use PyTorch container for libtorch and torchvision
############################################################################
FROM ${PYTORCH_IMAGE} AS tritonserver_pytorch

# Must rebuild in the pytorch container to disable some features that
# are not relevant for inferencing and so that OpenCV libraries are
# not included in the server (which will likely conflict with custom
# backends using opencv). The uninstalls seem excessive but is the
# recommendation from pytorch CONTRIBUTING.md.
WORKDIR /opt/pytorch
RUN (conda uninstall -y pytorch || true) && \
(conda uninstall -y ninja || true) && \
pip uninstall -y torch && \
pip uninstall -y torch
RUN cd pytorch && \
python setup.py clean && \
TORCH_CUDA_ARCH_LIST="5.2 6.0 6.1 7.0 7.5 8.0 8.6+PTX" \
CUDA_HOME="/usr/local/cuda" \
CMAKE_PREFIX_PATH="$(dirname $(which conda))/../" \
USE_DISTRIBUTED=OFF USE_OPENMP=OFF USE_NCCL=OFF USE_SYSTEM_NCCL=OFF \
USE_OPENCV=OFF USE_LEVELDB=OFF USE_LMDB=OFF USE_REDIS=OFF \
BUILD_TEST=OFF \
pip install --no-cache-dir -v .
'''

if 'onnxruntime' in backends:
df += '''
############################################################################
Expand Down Expand Up @@ -587,43 +559,6 @@ def create_dockerfile_buildbase(ddir, dockerfile_name, argmap, backends):
apt-get update && \
apt-get install -y --no-install-recommends \
cmake-data=3.18.4-0kitware1ubuntu20.04.1 cmake=3.18.4-0kitware1ubuntu20.04.1
'''
if 'pytorch' in backends:
df += '''
# LibTorch and Torchvision headers and libraries
COPY --from=tritonserver_pytorch \
/opt/conda/lib/python3.8/site-packages/torch/lib/libc10.so \
/opt/tritonserver/backends/pytorch/
COPY --from=tritonserver_pytorch \
/opt/conda/lib/python3.8/site-packages/torch/lib/libc10_cuda.so \
/opt/tritonserver/backends/pytorch/
COPY --from=tritonserver_pytorch /opt/conda/lib/libmkl_core.so /opt/tritonserver/backends/pytorch/
COPY --from=tritonserver_pytorch /opt/conda/lib/libmkl_gnu_thread.so /opt/tritonserver/backends/pytorch/
COPY --from=tritonserver_pytorch /opt/conda/lib/libmkl_intel_lp64.so /opt/tritonserver/backends/pytorch/
COPY --from=tritonserver_pytorch /opt/conda/lib/libmkl_intel_thread.so /opt/tritonserver/backends/pytorch/
COPY --from=tritonserver_pytorch /opt/conda/lib/libmkl_def.so /opt/tritonserver/backends/pytorch/
COPY --from=tritonserver_pytorch /opt/conda/lib/libmkl_avx2.so /opt/tritonserver/backends/pytorch/
COPY --from=tritonserver_pytorch /opt/conda/lib/libmkl_vml_def.so /opt/tritonserver/backends/pytorch/
COPY --from=tritonserver_pytorch /opt/conda/lib/libmkl_rt.so /opt/tritonserver/backends/pytorch/
COPY --from=tritonserver_pytorch /opt/conda/lib/libiomp5.so /opt/tritonserver/backends/pytorch/
COPY --from=tritonserver_pytorch /opt/conda/lib/python3.8/site-packages/torch/include \
/opt/tritonserver/include/torch
COPY --from=tritonserver_pytorch /opt/conda/lib/python3.8/site-packages/torch/lib/libtorch.so \
/opt/tritonserver/backends/pytorch/
COPY --from=tritonserver_pytorch /opt/conda/lib/python3.8/site-packages/torch/lib/libtorch_cpu.so \
/opt/tritonserver/backends/pytorch/
COPY --from=tritonserver_pytorch /opt/conda/lib/python3.8/site-packages/torch/lib/libtorch_cuda.so \
/opt/tritonserver/backends/pytorch/
COPY --from=tritonserver_pytorch /opt/pytorch/vision/torchvision/csrc \
/opt/tritonserver/include/torchvision/torchvision/
COPY --from=tritonserver_pytorch /opt/pytorch/vision/build/libtorchvision.so \
/opt/tritonserver/backends/pytorch/
COPY --from=tritonserver_pytorch /opt/pytorch/pytorch/LICENSE \
/opt/tritonserver/backends/pytorch/
RUN cd /opt/tritonserver/backends/pytorch && \
for i in `find . -mindepth 1 -maxdepth 1 -type f -name '*\.so*'`; do \
patchelf --set-rpath '$ORIGIN' $i; \
done
'''
if 'onnxruntime' in backends:
df += '''
Expand Down Expand Up @@ -843,10 +778,6 @@ def create_dockerfile(ddir, dockerfile_name, argmap, backends):
COPY --chown=1000:1000 TRITON_VERSION .
COPY --chown=1000:1000 --from=tritonserver_build /tmp/tritonbuild/install/bin/tritonserver bin/
COPY --chown=1000:1000 --from=tritonserver_build /tmp/tritonbuild/install/lib/libtritonserver.so lib/
'''
if 'pytorch' in backends:
df += '''
COPY --chown=1000:1000 --from=tritonserver_build /opt/tritonserver/backends/pytorch backends/pytorch
'''
if 'onnxruntime' in backends:
df += '''
Expand Down Expand Up @@ -917,16 +848,6 @@ def container_build(backends, images):
base_image,
}

# If building the pytorch backend then need to include pytorch in
# buildbase image.
if 'pytorch' in backends:
if 'pytorch' in images:
pytorch_image = images['pytorch']
else:
pytorch_image = 'nvcr.io/nvidia/pytorch:{}-py3'.format(
FLAGS.upstream_container_version)
dockerfileargmap['PYTORCH_IMAGE'] = pytorch_image

# If building the ORT backend then need to include ORT in
# buildbase image.
if 'onnxruntime' in backends:
Expand All @@ -940,11 +861,9 @@ def container_build(backends, images):
FLAGS.version][3]

cachefrommap = [
'tritonserver_pytorch', 'tritonserver_pytorch_cache0',
'tritonserver_pytorch_cache1', 'tritonserver_onnx',
'tritonserver_onnx_cache0', 'tritonserver_onnx_cache1',
'tritonserver_buildbase', 'tritonserver_buildbase_cache0',
'tritonserver_buildbase_cache1'
'tritonserver_onnx', 'tritonserver_onnx_cache0',
'tritonserver_onnx_cache1', 'tritonserver_buildbase',
'tritonserver_buildbase_cache0', 'tritonserver_buildbase_cache1'
]

cachefromargs = ['--cache-from={}'.format(k) for k in cachefrommap]
Expand All @@ -965,16 +884,6 @@ def container_build(backends, images):
# does caching with multi-stage images, we must build each
# stage separately to make sure it is cached.

# PyTorch
if 'pytorch' in backends:
p = subprocess.Popen(commonargs + cachefromargs + [
'-t', 'tritonserver_pytorch', '--target',
'tritonserver_pytorch', '.'
])
p.wait()
fail_if(p.returncode != 0,
'docker build tritonserver_pytorch failed')

# ONNX Runtime
if 'onnxruntime' in backends:
p = subprocess.Popen(commonargs + cachefromargs + [
Expand Down