mirror of
https://github.com/saymrwulf/onnxruntime.git
synced 2026-07-28 20:11:22 +00:00
[EP Perf] Update EP Perf dockerfiles with cuda12/cudnn9 (#19781)
### Description * Update name of existing dockerfiles and add support to test latest TensorRT EA binary located in the image * Add cuda 12.3/cuDNN 9/TensorRT 8.6 dockerfile * Add detail to CI prompts and configs Instruction to test latest TRT via BIN: 1. Select `BIN` in TensorRT Version 2. In Variables, update related tarCudaVersion, **clear** tarCudnnVersion (not required in latest TRT tar binary) , and path to binary.
This commit is contained in:
Yifan Li
GitHub
parent
3170a48e60
commit
069d2d6f54
4 changed files with 183 additions and 60 deletions
|
|
@ -14,9 +14,10 @@ import sys
|
|||
from typing import List, Optional
|
||||
|
||||
TRT_DOCKER_FILES = {
|
||||
"8.4": "tools/ci_build/github/linux/docker/Dockerfile.ubuntu_cuda11_6_tensorrt8_4",
|
||||
"8.5": "tools/ci_build/github/linux/docker/Dockerfile.ubuntu_cuda11_8_tensorrt8_5",
|
||||
"8.6": "tools/ci_build/github/linux/docker/Dockerfile.ubuntu_cuda11_8_tensorrt8_6",
|
||||
"8.4.cuda_11_6_cudnn_8": "tools/ci_build/github/linux/docker/Dockerfile.ubuntu_cuda11_6_tensorrt8_4",
|
||||
"8.5.cuda_11_8_cudnn_8": "tools/ci_build/github/linux/docker/Dockerfile.ubuntu_cuda11_8_tensorrt8_5",
|
||||
"8.6.cuda_11_8_cudnn_8": "tools/ci_build/github/linux/docker/Dockerfile.ubuntu_cuda11_8_tensorrt8_6",
|
||||
"8.6.cuda_12_3_cudnn_9": "tools/ci_build/github/linux/docker/Dockerfile.ubuntu_cuda12_3_tensorrt8_6",
|
||||
"BIN": "tools/ci_build/github/linux/docker/Dockerfile.ubuntu_tensorrt_bin",
|
||||
}
|
||||
|
||||
|
|
@ -99,18 +100,11 @@ def docker_build_trt(args: argparse.Namespace):
|
|||
:param args: The arguments to this script.
|
||||
"""
|
||||
|
||||
if not is_valid_ver_str(args.trt_version, min_comps=2, max_comps=4):
|
||||
print(f"[ERROR]: Invalid TensorRT version '{args.trt_version}'", file=sys.stderr)
|
||||
sys.exit(1)
|
||||
|
||||
vers_comps = args.trt_version.split(".")
|
||||
trt_ver_key = f"{vers_comps[0]}.{vers_comps[1]}"
|
||||
|
||||
if trt_ver_key not in TRT_DOCKER_FILES:
|
||||
if args.trt_version not in TRT_DOCKER_FILES:
|
||||
print(f"[ERROR]: TensorRT version '{args.trt_version}' is currently unsupported", file=sys.stderr)
|
||||
sys.exit(1)
|
||||
|
||||
docker_file = TRT_DOCKER_FILES[trt_ver_key]
|
||||
docker_file = TRT_DOCKER_FILES[args.trt_version]
|
||||
docker_file_path = os.path.normpath(os.path.join(args.repo_path, docker_file))
|
||||
|
||||
if not os.path.isfile(docker_file_path):
|
||||
|
|
@ -144,11 +138,7 @@ def docker_build_trt_bin(args: argparse.Namespace):
|
|||
sys.exit(1)
|
||||
|
||||
if not is_valid_ver_str(args.tar_cuda_version, 2, 2):
|
||||
print("[ERROR]: Must specify a valid CUDA version for binary TensorRT installs (e.g., 11.x)", file=sys.stderr)
|
||||
sys.exit(1)
|
||||
|
||||
if not is_valid_ver_str(args.tar_cudnn_version, 2, 2):
|
||||
print("[ERROR]: Must specify a valid cuDNN version for binary TensorRT installs (e.g., 8.x)", file=sys.stderr)
|
||||
print("[ERROR]: Must specify a valid CUDA version for binary TensorRT installs (e.g., 12.4)", file=sys.stderr)
|
||||
sys.exit(1)
|
||||
|
||||
if not os.path.isfile(docker_file_path):
|
||||
|
|
@ -170,8 +160,6 @@ def docker_build_trt_bin(args: argparse.Namespace):
|
|||
"--build-arg",
|
||||
f"TAR_CUDA_VERSION={args.tar_cuda_version}",
|
||||
"--build-arg",
|
||||
f"TAR_CUDNN_VERSION={args.tar_cudnn_version}",
|
||||
"--build-arg",
|
||||
f"TRT_BINS_DIR={args.trt_bins_dir}",
|
||||
"-f",
|
||||
f"{docker_file_path}",
|
||||
|
|
@ -195,7 +183,9 @@ def parse_arguments() -> argparse.Namespace:
|
|||
parser.add_argument("-r", "--repo_path", required=True, help="Path to the onnxruntime repository")
|
||||
parser.add_argument("-i", "--image_name", required=True, help="The resulting Docker image name")
|
||||
parser.add_argument("-b", "--branch", default="main", help="Name of the onnxruntime git branch to checkout")
|
||||
parser.add_argument("-t", "--trt_version", default="8.6.1.6", help="TensorRT version (e.g., 8.6.1.6)")
|
||||
parser.add_argument(
|
||||
"-t", "--trt_version", default="8.6.cuda_11_8_cudnn_8", help="TensorRT version (e.g., 8.6.cuda_11_8_cudnn_8)"
|
||||
)
|
||||
parser.add_argument("-a", "--cuda_arch", default="75", help="CUDA architecture (e.g., 75)")
|
||||
|
||||
# Command-line options for installing TensorRT from binaries.
|
||||
|
|
@ -208,12 +198,7 @@ def parse_arguments() -> argparse.Namespace:
|
|||
parser.add_argument(
|
||||
"--tar_cuda_version",
|
||||
default="",
|
||||
help="CUDA version (e.g., 11.8) used to find TensorRT EA binary tar.gz package",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--tar_cudnn_version",
|
||||
default="",
|
||||
help="CUDA version (e.g., 8.6) used to find TensorRT EA binary tar.gz package",
|
||||
help="CUDA version (e.g., 12.4) used to find TensorRT EA binary tar.gz package",
|
||||
)
|
||||
parser.add_argument("--trt_bins_dir", default="", help="Directory containing TensorRT tar.gz package")
|
||||
parser.add_argument(
|
||||
|
|
|
|||
|
|
@ -8,15 +8,16 @@ parameters:
|
|||
- name: TrtVersion
|
||||
displayName: TensorRT Version
|
||||
type: string
|
||||
default: 8.6.1.6
|
||||
default: 8.6.cuda_11_8_cudnn_8
|
||||
values:
|
||||
- 8.4.1.5
|
||||
- 8.5.1.1
|
||||
- 8.6.1.6
|
||||
- 8.4.cuda_11_6_cudnn_8
|
||||
- 8.5.cuda_11_8_cudnn_8
|
||||
- 8.6.cuda_11_8_cudnn_8
|
||||
- 8.6.cuda_12_3_cudnn_9
|
||||
- BIN
|
||||
|
||||
- name: UseTensorrtOssParser
|
||||
displayName: Use TensorRT-OSS Parser
|
||||
displayName: Use TensorRT-OSS Parser (not compatible with BIN)
|
||||
type: boolean
|
||||
default: false
|
||||
|
||||
|
|
@ -86,11 +87,11 @@ jobs:
|
|||
- script: 'ls -al $(trtBinsDir)'
|
||||
displayName: 'Show available TensorRT .tar.gz packages'
|
||||
|
||||
- script: 'cp $(trtBinsDir)/TensorRT-$(trtVersion).Linux.x86_64-gnu.cuda-$(tarCudaVersion).cudnn$(tarCudnnVersion).tar.gz $(Build.SourcesDirectory)/onnxruntime/python/tools/tensorrt/perf/build/'
|
||||
- script: 'cp $(trtBinsDir)/TensorRT-$(trtVersion).Linux.x86_64-gnu.cuda-$(tarCudaVersion).tar.gz $(Build.SourcesDirectory)/onnxruntime/python/tools/tensorrt/perf/build/'
|
||||
displayName: 'Copy TensorRT .tar.gz package into Docker build directory'
|
||||
|
||||
- script: 'python3 $(Build.SourcesDirectory)/onnxruntime/python/tools/tensorrt/perf/build/build_image.py -r $(Build.SourcesDirectory) -i $(image) -b $(branchName) -t $(trtVersion) -a 75 --install_bin --tar_cuda_version=$(tarCudaVersion) --tar_cudnn_version=$(tarCudnnVersion) --trt_bins_dir=.'
|
||||
displayName: 'Install TensorRT from binaries and build latest ORT Image'
|
||||
- script: 'python3 $(Build.SourcesDirectory)/onnxruntime/python/tools/tensorrt/perf/build/build_image.py -r $(Build.SourcesDirectory) -i $(image) -b $(branchName) -t $(trtVersion) -a 75 --install_bin --tar_cuda_version=$(tarCudaVersion) --trt_bins_dir=.'
|
||||
displayName: 'Install TensorRT $(tarTrtVersion) from binaries and build latest ORT Image'
|
||||
workingDirectory: '$(Build.SourcesDirectory)/onnxruntime/python/tools/tensorrt/perf/build'
|
||||
|
||||
# Build ORT with TensorRT built-in parser
|
||||
|
|
|
|||
|
|
@ -0,0 +1,96 @@
|
|||
# --------------------------------------------------------------
|
||||
# Copyright (c) Microsoft Corporation. All rights reserved.
|
||||
# Licensed under the MIT License.
|
||||
# --------------------------------------------------------------
|
||||
# Dockerfile to run ONNXRuntime with TensorRT integration
|
||||
|
||||
# Build base image with required system packages
|
||||
FROM nvidia/cuda:12.3.1-devel-ubuntu20.04 AS base
|
||||
|
||||
# The local directory into which to build and install CMAKE
|
||||
ARG ONNXRUNTIME_LOCAL_CODE_DIR=/code
|
||||
|
||||
ENV PATH /usr/local/nvidia/bin:/usr/local/cuda/bin:/usr/src/tensorrt/bin:${ONNXRUNTIME_LOCAL_CODE_DIR}/cmake-3.27.3-linux-x86_64/bin:/opt/miniconda/bin:${PATH}
|
||||
ENV DEBIAN_FRONTEND=noninteractive
|
||||
|
||||
RUN apt-get update &&\
|
||||
apt-get install -y sudo git bash unattended-upgrades wget
|
||||
RUN unattended-upgrade
|
||||
|
||||
# Install python3
|
||||
RUN apt-get install -y --no-install-recommends \
|
||||
python3 \
|
||||
python3-pip \
|
||||
python3-dev \
|
||||
python3-wheel &&\
|
||||
cd /usr/local/bin &&\
|
||||
ln -s /usr/bin/python3 python &&\
|
||||
ln -s /usr/bin/pip3 pip;
|
||||
|
||||
RUN pip install --upgrade pip
|
||||
RUN pip install setuptools>=68.2.2
|
||||
|
||||
# Install cuDNN v9
|
||||
RUN apt-get -y install cudnn9-cuda-12
|
||||
|
||||
# Install TensorRT
|
||||
RUN v="8.6.1.6-1+cuda12.0" &&\
|
||||
apt-key adv --fetch-keys https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/x86_64/7fa2af80.pub &&\
|
||||
apt-get update &&\
|
||||
sudo apt-get install -y libnvinfer8=${v} libnvonnxparsers8=${v} libnvparsers8=${v} libnvinfer-plugin8=${v} libnvinfer-lean8=${v} libnvinfer-vc-plugin8=${v} libnvinfer-dispatch8=${v}\
|
||||
libnvinfer-headers-dev=${v} libnvinfer-headers-plugin-dev=${v} libnvinfer-dev=${v} libnvonnxparsers-dev=${v} libnvparsers-dev=${v} libnvinfer-plugin-dev=${v} libnvinfer-lean-dev=${v} libnvinfer-vc-plugin-dev=${v} libnvinfer-dispatch-dev=${v}\
|
||||
python3-libnvinfer=${v} libnvinfer-samples=${v} tensorrt-dev=${v} tensorrt-libs=${v}
|
||||
|
||||
# Compile trtexec
|
||||
RUN cd /usr/src/tensorrt/samples/trtexec && make
|
||||
|
||||
# Install Valgrind
|
||||
RUN apt-get install -y valgrind
|
||||
|
||||
# Build final image from base. Builds ORT.
|
||||
FROM base as final
|
||||
ARG BUILD_USER=onnxruntimedev
|
||||
ARG BUILD_UID=1000
|
||||
RUN adduser --gecos 'onnxruntime Build User' --disabled-password $BUILD_USER --uid $BUILD_UID
|
||||
USER $BUILD_USER
|
||||
|
||||
# ONNX Runtime arguments
|
||||
|
||||
# URL to the github repo from which to clone ORT.
|
||||
ARG ONNXRUNTIME_REPO=https://github.com/Microsoft/onnxruntime
|
||||
|
||||
# The local directory into which to clone ORT.
|
||||
ARG ONNXRUNTIME_LOCAL_CODE_DIR=/code
|
||||
|
||||
# The git branch of ORT to checkout and build.
|
||||
ARG ONNXRUNTIME_BRANCH=main
|
||||
|
||||
# Optional. The specific commit to pull and build from. If not set, the latest commit is used.
|
||||
ARG ONNXRUNTIME_COMMIT_ID
|
||||
|
||||
# The supported CUDA architecture
|
||||
ARG CMAKE_CUDA_ARCHITECTURES=75
|
||||
|
||||
WORKDIR ${ONNXRUNTIME_LOCAL_CODE_DIR}
|
||||
|
||||
# Clone ORT repository with branch
|
||||
RUN git clone --single-branch --branch ${ONNXRUNTIME_BRANCH} --recursive ${ONNXRUNTIME_REPO} onnxruntime &&\
|
||||
/bin/sh onnxruntime/dockerfiles/scripts/install_common_deps.sh
|
||||
|
||||
WORKDIR ${ONNXRUNTIME_LOCAL_CODE_DIR}/onnxruntime
|
||||
|
||||
# Reset to a specific commit if specified by build args.
|
||||
RUN if [ -z "$ONNXRUNTIME_COMMIT_ID" ] ; then echo "Building branch ${ONNXRUNTIME_BRANCH}" ;\
|
||||
else echo "Building branch ${ONNXRUNTIME_BRANCH} @ commit ${ONNXRUNTIME_COMMIT_ID}" &&\
|
||||
git reset --hard ${ONNXRUNTIME_COMMIT_ID} && git submodule update --recursive ; fi
|
||||
|
||||
# Build ORT
|
||||
ENV CUDA_MODULE_LOADING "LAZY"
|
||||
ARG PARSER_CONFIG=""
|
||||
RUN /bin/sh build.sh ${PARSER_CONFIG} --parallel --build_shared_lib --cuda_home /usr/local/cuda --cudnn_home /usr/lib/x86_64-linux-gnu/ --use_tensorrt --tensorrt_home /usr/lib/x86_64-linux-gnu/ --config Release --build_wheel --skip_tests --skip_submodule_sync --cmake_extra_defines '"CMAKE_CUDA_ARCHITECTURES='${CMAKE_CUDA_ARCHITECTURES}'"'
|
||||
|
||||
# Switch to root to continue following steps of CI
|
||||
USER root
|
||||
|
||||
# Intall ORT wheel
|
||||
RUN pip install ${ONNXRUNTIME_LOCAL_CODE_DIR}/onnxruntime/build/Linux/Release/dist/*.whl
|
||||
|
|
@ -4,29 +4,15 @@
|
|||
# --------------------------------------------------------------
|
||||
# Dockerfile to run ONNXRuntime with TensorRT installed from provided binaries
|
||||
|
||||
FROM nvidia/cuda:11.7.1-cudnn8-devel-ubuntu20.04
|
||||
# Build base image with required system packages
|
||||
FROM nvidia/cuda:12.3.1-devel-ubuntu20.04 AS base
|
||||
|
||||
# The local directory into which to build and install CMAKE
|
||||
ARG ONNXRUNTIME_LOCAL_CODE_DIR=/code
|
||||
|
||||
# ONNX Runtime Variables
|
||||
ARG ONNXRUNTIME_REPO=https://github.com/Microsoft/onnxruntime
|
||||
ARG ONNXRUNTIME_BRANCH=main
|
||||
ARG CMAKE_CUDA_ARCHITECTURES=37;50;52;60;61;70;75;80
|
||||
|
||||
# Must provide version numbers used to build the name of the tar file containing TensorRT binaries.
|
||||
# See: https://docs.nvidia.com/deeplearning/tensorrt/install-guide/index.html#installing-tar
|
||||
ARG TAR_TRT_VERSION
|
||||
ARG TAR_CUDA_VERSION
|
||||
ARG TAR_CUDNN_VERSION
|
||||
|
||||
# Directory containing TensorRT tar.gz installation package
|
||||
ARG TRT_BINS_DIR=.
|
||||
|
||||
ENV PATH /usr/local/nvidia/bin:/usr/local/cuda/bin:/code/cmake-3.27.3-linux-x86_64/bin:/opt/miniconda/bin:${PATH}
|
||||
|
||||
ENV PATH /usr/local/nvidia/bin:/usr/local/cuda/bin:${ONNXRUNTIME_LOCAL_CODE_DIR}/cmake-3.27.3-linux-x86_64/bin:/opt/miniconda/bin:${PATH}
|
||||
ENV DEBIAN_FRONTEND=noninteractive
|
||||
|
||||
COPY ${TRT_BINS_DIR}/TensorRT-${TAR_TRT_VERSION}.Linux.x86_64-gnu.cuda-${TAR_CUDA_VERSION}.cudnn${TAR_CUDNN_VERSION}.tar.gz /TensorRT-${TAR_TRT_VERSION}.tar.gz
|
||||
|
||||
RUN apt-get update &&\
|
||||
apt-get install -y sudo git bash unattended-upgrades wget
|
||||
RUN unattended-upgrade
|
||||
|
|
@ -44,22 +30,77 @@ RUN apt-get install -y --no-install-recommends \
|
|||
RUN pip install --upgrade pip
|
||||
RUN pip install setuptools>=68.2.2
|
||||
|
||||
# Install cuDNN v9
|
||||
RUN apt-get -y install cudnn9-cuda-12
|
||||
|
||||
# Install TensorRT
|
||||
# Must provide version numbers used to build the name of the tar file containing TensorRT binaries.
|
||||
# See: https://docs.nvidia.com/deeplearning/tensorrt/install-guide/index.html#installing-tar
|
||||
ARG TAR_TRT_VERSION
|
||||
ARG TAR_CUDA_VERSION
|
||||
|
||||
# Directory containing TensorRT tar.gz installation package
|
||||
ARG TRT_BINS_DIR=.
|
||||
COPY ${TRT_BINS_DIR}/TensorRT-${TAR_TRT_VERSION}.Linux.x86_64-gnu.cuda-${TAR_CUDA_VERSION}.tar.gz /TensorRT-${TAR_TRT_VERSION}.tar.gz
|
||||
|
||||
# Install TensorRT from tar.gz
|
||||
RUN tar -xzvf /TensorRT-${TAR_TRT_VERSION}.tar.gz
|
||||
|
||||
RUN cd /TensorRT-${TAR_TRT_VERSION}/python &&\
|
||||
python3 -m pip install tensorrt-${TAR_TRT_VERSION}-cp38-none-linux_x86_64.whl
|
||||
python3 -m pip install tensorrt*cp38*.whl
|
||||
|
||||
RUN cp -r /TensorRT-${TAR_TRT_VERSION}/lib/* /usr/lib/x86_64-linux-gnu/
|
||||
RUN cp /TensorRT-${TAR_TRT_VERSION}/include/* /usr/local/include/
|
||||
RUN cp /TensorRT-${TAR_TRT_VERSION}/bin/* /usr/local/bin/
|
||||
|
||||
WORKDIR /code
|
||||
# Install Valgrind
|
||||
RUN apt-get install -y valgrind
|
||||
|
||||
# Build final image from base. Builds ORT.
|
||||
FROM base as final
|
||||
ARG BUILD_USER=onnxruntimedev
|
||||
ARG BUILD_UID=1000
|
||||
RUN adduser --gecos 'onnxruntime Build User' --disabled-password $BUILD_USER --uid $BUILD_UID
|
||||
USER $BUILD_USER
|
||||
|
||||
# ONNX Runtime arguments
|
||||
|
||||
# URL to the github repo from which to clone ORT.
|
||||
ARG ONNXRUNTIME_REPO=https://github.com/Microsoft/onnxruntime
|
||||
|
||||
# The local directory into which to clone ORT.
|
||||
ARG ONNXRUNTIME_LOCAL_CODE_DIR=/code
|
||||
|
||||
# The git branch of ORT to checkout and build.
|
||||
ARG ONNXRUNTIME_BRANCH=main
|
||||
|
||||
# Optional. The specific commit to pull and build from. If not set, the latest commit is used.
|
||||
ARG ONNXRUNTIME_COMMIT_ID
|
||||
|
||||
# The supported CUDA architecture
|
||||
ARG CMAKE_CUDA_ARCHITECTURES=75
|
||||
|
||||
# Prepare onnxruntime repository & build onnxruntime with TensorRT
|
||||
WORKDIR ${ONNXRUNTIME_LOCAL_CODE_DIR}
|
||||
|
||||
# Clone ORT repository with branch
|
||||
RUN git clone --single-branch --branch ${ONNXRUNTIME_BRANCH} --recursive ${ONNXRUNTIME_REPO} onnxruntime &&\
|
||||
/bin/sh onnxruntime/dockerfiles/scripts/install_common_deps.sh &&\
|
||||
cd onnxruntime &&\
|
||||
/bin/sh build.sh --parallel --build_shared_lib --cuda_home /usr/local/cuda --cudnn_home /usr/lib/x86_64-linux-gnu/ --use_tensorrt --tensorrt_home /usr/lib/x86_64-linux-gnu/ --config Release --build_wheel --skip_tests --skip_submodule_sync --cmake_extra_defines '"CMAKE_CUDA_ARCHITECTURES='${CMAKE_CUDA_ARCHITECTURES}'"' &&\
|
||||
pip install /code/onnxruntime/build/Linux/Release/dist/*.whl &&\
|
||||
cd ..
|
||||
/bin/sh onnxruntime/dockerfiles/scripts/install_common_deps.sh
|
||||
|
||||
WORKDIR ${ONNXRUNTIME_LOCAL_CODE_DIR}/onnxruntime
|
||||
|
||||
# Reset to a specific commit if specified by build args.
|
||||
RUN if [ -z "$ONNXRUNTIME_COMMIT_ID" ] ; then echo "Building branch ${ONNXRUNTIME_BRANCH}" ;\
|
||||
else echo "Building branch ${ONNXRUNTIME_BRANCH} @ commit ${ONNXRUNTIME_COMMIT_ID}" &&\
|
||||
git reset --hard ${ONNXRUNTIME_COMMIT_ID} && git submodule update --recursive ; fi
|
||||
|
||||
# Build ORT
|
||||
ENV CUDA_MODULE_LOADING "LAZY"
|
||||
ARG PARSER_CONFIG=""
|
||||
RUN /bin/sh build.sh ${PARSER_CONFIG} --parallel --build_shared_lib --cuda_home /usr/local/cuda --cudnn_home /usr/lib/x86_64-linux-gnu/ --use_tensorrt --tensorrt_home /usr/lib/x86_64-linux-gnu/ --config Release --build_wheel --skip_tests --skip_submodule_sync --cmake_extra_defines '"CMAKE_CUDA_ARCHITECTURES='${CMAKE_CUDA_ARCHITECTURES}'"'
|
||||
|
||||
# Switch to root to continue following steps of CI
|
||||
USER root
|
||||
|
||||
# Intall ORT wheel
|
||||
RUN pip install ${ONNXRUNTIME_LOCAL_CODE_DIR}/onnxruntime/build/Linux/Release/dist/*.whl
|
||||
Loading…
Reference in a new issue