Expand support for CUDA 11, 12, MacOS, and ROCm (#3)

The ML Lib Builder now compiles libtensorflow, libtorch, and onnxruntime
for CUDA 11, CUDA 12, MacOS, and ROCm 5.7. These have been tested
on an internal HPE machine and intended for use with SmartSim

.gitignore

@@ -3,3 +3,4 @@ install
 *.tgz
 *.zip
 *.tar.gz
+slurm*.out

.gitmodules

@@ -1,3 +1,11 @@
 [submodule "pytorch"]
 	path = pytorch
 	url = https://github.com/pytorch/pytorch.git
+	shallow = true
+[submodule "tensorflow"]
+	path = tensorflow
+	url = https://github.com/tensorflow/tensorflow.git
+	shallow = true
+[submodule "onnxruntime"]
+	path = onnxruntime
+	url = https://github.com/microsoft/onnxruntime.git

Makefile

@@ -24,26 +24,32 @@
 # OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
 # OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
-PYTORCH_VERSION=2.0.1
-OSX_ARCHITECTURE=arm64
 
-TORCH_TARGET = libtorch-macos-$(OSX_ARCHITECTURE)-$(PYTORCH_VERSION).zip
-TORCH_BUILD = $(PWD)/build/libtorch
-TORCH_INSTALL = $(PWD)/install/libtorch
+ifeq ($(ARCH_FILE),)
+$(error Must specify ARCH_FILE)
+else
+include $(ARCH_FILE)
+endif
+
+INSTALL_DIR = $(PWD)/install
+BUILD_DIR = $(PWD)/build
+
+TORCH_ARCHIVE = $(INSTALL_DIR)/libtorch-$(PYTORCH_VERSION)-$(OS)-$(ARCHITECTURE)-$(STACK).tgz
+TORCH_BUILD_DIR = $(BUILD_DIR)/libtorch
+TORCH_INSTALL_DIR = $(INSTALL_DIR)/libtorch
 
-TORCH_CMAKE_OPTIONS =
-TORCH_CMAKE_OPTIONS += -DCMAKE_OSX_ARCHITECTURES=$(OSX_ARCHITECTURE)
-TORCH_CMAKE_OPTIONS += -DUSE_MKL=OFF -DUSE_MKLDNN=OFF -DUSE_ITT=OFF
-TORCH_CMAKE_OPTIONS += -DUSE_QNNPACK=OFF -DUSE_KINETO=OFF
+TF_ARCHIVE = $(INSTALL_DIR)/libtensorflow-$(TF_VERSION)-$(OS)-$(ARCHITECTURE)-$(STACK).tgz
+TF_INSTALL_DIR = $(INSTALL_DIR)/libtensorflow
+# Note: TF uses its own build system; cannot specify a build directory
+
+ONNXRT_ARCHIVE = $(INSTALL_DIR)/onnxruntime-$(ONNXRT_VERSION)-$(OS)-$(ARCHITECTURE)-$(STACK).tgz
+ONNXRT_BUILD_DIR = $(BUILD_DIR)/onnxruntime
+ONNXRT_INSTALL_DIR = $(INSTALL_DIR)/onnxruntime
 
 .PHONY: help
 help:
 	@grep "^# help\:" Makefile | grep -v grep | sed 's/\# help\: //' | sed 's/\# help\://'
 
-ifneq ($(shell uname), Darwin)
-	$(error This tool requires Mac OSX)
-endif
-
 # help:
 # help: ----Overview----
 # help: This makefile can be used to builds ML backends for use on arm64. Generally
@@ -55,35 +61,83 @@ endif
 # help: ----Meta targets----
 # help: clean						-- Cleans all build and install directories
 .PHONY: clean
-clean: clean_torch
+clean: clean_torch clean_tensorflow clean_onnxruntime
 
-# help:
 # help: ----Build Targets----
-# help: torch						-- Builds libtorch
-# help:
-.PHONY: torch
-torch: $(TORCH_TARGET)
 
-# Checkout a specific version of Torch and update all of the torch submodules
-.PHONY: checkout_torch
-checkout_torch:
-	cd pytorch && git checkout v$(PYTORCH_VERSION) && \
-		git submodule foreach --recursive git reset --hard && \
-		git submodule update --init --recursive
-
-$(TORCH_BUILD) $(TORCH_INSTALL):
+## Torch section
+$(TORCH_BUILD_DIR):
 	mkdir -p $@
 
-.PHONY: build_torch
-build_torch: $(TORCH_BUILD) $(TORCH_INSTALL) checkout_torch
-	cd $< && \
-		cmake -DCMAKE_INSTALL_PREFIX=$(TORCH_INSTALL) $(TORCH_CMAKE_OPTIONS) ../../pytorch && \
-		make install -j 6
+$(TORCH_ARCHIVE): $(TORCH_ARCHIVE_MODS) compile_torch
+	cd $(INSTALL_DIR) && tar -czf $@ libtorch/
 
-$(TORCH_TARGET): build_torch
-	cd install && zip -r ../$@ libtorch
+# help: build_torch					-- Builds libtorch
+.PHONY: build_torch
+build_torch: $(TORCH_ARCHIVE)
 
 .PHONY: clean_torch
 clean_torch:
-	rm -rf $(TORCH_BUILD) $(TORCH_TARGET) $(TORCH_INSTALL)
+	rm -rf $(TORCH_BUILD_DIR) $(TORCH_ARCHIVE) $(TORCH_INSTALL_DIR)
+	cd pytorch && git clean -fdx && git restore .
+	cd pytorch/third_party/kineto && git restore .
+
+.PHONY: compile_torch
+compile_torch: $(TORCH_BUILD_DIR) $(PYTORCH_PREBUILD_TARGETS)
+	cd $(TORCH_BUILD_DIR) && \
+		cmake -GNinja -DCMAKE_INSTALL_PREFIX=$(TORCH_INSTALL_DIR) -DPYTHON_EXECUTABLE=$$(which python) \
+		 	$(TORCH_CMAKE_OPTIONS) ../../pytorch && \
+		ninja install
+
+.PHONY: clean_tensorflow
+clean_tensorflow:
+	rm -rf $(TF_INSTALL_DIR)
+	cd tensorflow && \
+		bazel clean --expunge_async && \
+		git restore .
+
+.PHONY: clean_onnxruntime
+clean_onnxruntime:
+	rm -rf $(ONNXRT_INSTALL_DIR) $(ONNXRT_BUILD_DIR)
+	cd onnxruntime && \
+		git reset --hard && \
+		git clean -fdx && \
+		git restore .
+
+## Tensorflow section
+$(TF_INSTALL_DIR):
+	mkdir -p $@
+
+$(TF_ARCHIVE): $(TF_PREBUILD_TARGETS) $(TF_INSTALL_DIR)
+	cd tensorflow && \
+		bazel build $(TF_BAZEL_OPTS) //tensorflow/tools/lib_package:libtensorflow
+	cp tensorflow/bazel-bin/tensorflow/tools/lib_package/libtensorflow.tar.gz $(TF_INSTALL_DIR)
+	cd $(TF_INSTALL_DIR) && tar -xzf libtensorflow.tar.gz && rm -f libtensorflow.tar.gz
+	cd $(INSTALL_DIR) && tar -czf $@ libtensorflow
+
+# help: build_tensorflow			-- Builds Tensorflow
+.PHONY: build_tensorflow
+build_tensorflow: $(TF_ARCHIVE)
+
+## ONNX Runtime
+compile_onnxruntime: $(ONNXRT_PREBUILD_TARGETS)
+	cd onnxruntime && \
+		git apply ../patches/onnxruntime/build.install.patch
+	cd onnxruntime && python tools/ci_build/build.py \
+		--config Release \
+		--build_dir=$(ONNXRT_BUILD_DIR) \
+		--compile_no_warning_as_error \
+		--parallel \
+		--skip_tests \
+		--install_dir=$(ONNXRT_INSTALL_DIR) \
+		--build_shared_lib \
+		$(ONNXRT_OPTIONS)
+
+$(ONNXRT_ARCHIVE): compile_onnxruntime
+	cd $(ONNXRT_BUILD_DIR)/Release && make install
+	cd $(ONNXRT_INSTALL_DIR) && mv include/onnxruntime/* include && rm -rf include/onnxruntime && mv lib64 lib
+	cd $(INSTALL_DIR) && tar -czf $@ onnxruntime/
 
+# help: build_onnxruntime			-- Builds ONNX Runtime
+.PHONY: build_onnxruntime
+build_onnxruntime: $(ONNXRT_ARCHIVE)

architectures/linux-cuda-11.8.0

@@ -0,0 +1,89 @@
+
+# BSD 2-Clause License
+#
+# Copyright (c) 2024, Hewlett Packard Enterprise
+# All rights reserved.
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are met:
+#
+# 1. Redistributions of source code must retain the above copyright notice, this
+#    list of conditions and the following disclaimer.
+#
+# 2. Redistributions in binary form must reproduce the above copyright notice,
+#    this list of conditions and the following disclaimer in the documentation
+#    and/or other materials provided with the distribution.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+# DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+# FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+# SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+# OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+OS = linux
+ARCHITECTURE = x64
+CUDA_VERSION = 11.8.0
+STACK=cuda-$(CUDA_VERSION)
+
+# pyTorch options
+PYTORCH_VERSION = 2.4.0
+TORCH_CMAKE_OPTIONS = -DBUILD_PYTHON=OFF
+TORCH_CMAKE_OPTIONS += -DUSE_NCCL=ON -DUSE_CUDA=ON -DUSE_STATIC_MKL=ON -DUSE_ROCM=OFF -DUSE_CUDNN=ON
+TORCH_CMAKE_OPTIONS += -DTORCH_CUDA_ARCH_LIST="All" -DCUDNN_LIBRARY_PATH=${CUDNN_LIBRARY} -DCUDNN_INCLUDE_PATH=${CUDNN_INCLUDE_DIR}
+PYTORCH_PREBUILD_TARGETS = pytorch_checkout
+
+# Tensorflow options
+TF_VERSION = 2.14.1
+TF_TAG = v$(TF_VERSION)
+TF_REMOTE = https://github.com/tensorflow/tensorflow.git
+TF_PREBUILD_TARGETS = tf_checkout tf_prebuild
+TF_BAZEL_OPTS = --jobs 192
+
+
+ONNXRT_VERSION = 1.17.3
+ONNXRT_OPTIONS = --use_cuda --cudnn_home=${CUDNN_INSTALL_PATH}
+ONNXRT_PREBUILD_TARGETS = onnxrt_checkout onnxrt_patch
+
+pytorch_checkout:
+	cd pytorch && \
+		git checkout v${PYTORCH_VERSION} && \
+		git submodule update --init --recursive
+
+pytorch_rocm_prebuild:
+	cd pytorch; python tools/amd_build/build_amd.py
+	sed -i 's/attr.memoryType/attr.type/g' pytorch/aten/src/ATen/hip/detail/HIPHooks.cpp
+	cd pytorch; git apply ../patches/pytorch/caffe2_rocm_path.patch
+
+tf_prebuild:
+	cd tensorflow; \
+		USE_DEFAULT_PYTHON_LIB_PATH=1 \
+		PYTHON_BIN_PATH=$$(which python) \
+		TF_NEED_CLANG=1 \
+		TF_NEED_ROCM=0 \
+		TF_NEED_CUDA=1 \
+		TF_NEED_TENSORRT=0 \
+		TF_CUDA_CLANG=0 \
+		CC_OPT_FLAGS="-Wno-sign-compare" \
+		TF_SET_ANDROID_WORKSPACE=0 \
+		python configure.py
+
+tf_checkout:
+	cd tensorflow; \
+		git fetch $(TF_REMOTE) $(TF_TAG) && \
+		git checkout FETCH_HEAD
+
+onnxrt_checkout:
+	cd onnxruntime && \
+		git checkout v$(ONNXRT_VERSION) && \
+		git reset --hard && \
+		git clean -xdf && \
+		git submodule update --init --recursive
+
+onnxrt_patch:
+	cd onnxruntime && \
+		git apply ../patches/onnxruntime/cfloat.patch

architectures/linux-cuda-12.5.0

@@ -0,0 +1,89 @@
+
+# BSD 2-Clause License
+#
+# Copyright (c) 2024, Hewlett Packard Enterprise
+# All rights reserved.
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are met:
+#
+# 1. Redistributions of source code must retain the above copyright notice, this
+#    list of conditions and the following disclaimer.
+#
+# 2. Redistributions in binary form must reproduce the above copyright notice,
+#    this list of conditions and the following disclaimer in the documentation
+#    and/or other materials provided with the distribution.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+# DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+# FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+# SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+# OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+OS = linux
+ARCHITECTURE = x64
+CUDA_VERSION = 12.5.0
+STACK=cuda-$(CUDA_VERSION)
+
+# pyTorch options
+PYTORCH_VERSION = 2.4.0
+TORCH_CMAKE_OPTIONS = -DBUILD_PYTHON=OFF
+TORCH_CMAKE_OPTIONS += -DUSE_NCCL=ON -DUSE_CUDA=ON -DUSE_STATIC_MKL=ON -DUSE_ROCM=OFF -DUSE_CUDNN=ON
+TORCH_CMAKE_OPTIONS += -DTORCH_CUDA_ARCH_LIST=${TORCH_CUDA_ARCH_LIST} -DCUDNN_LIBRARY_PATH=${CUDNN_LIBRARY} -DCUDNN_INCLUDE_PATH=${CUDNN_INCLUDE_DIR}
+PYTORCH_PREBUILD_TARGETS = pytorch_checkout
+
+# Tensorflow options
+TF_VERSION = 2.17.0
+TF_TAG = v$(TF_VERSION)
+TF_REMOTE = https://github.com/tensorflow/tensorflow.git
+TF_PREBUILD_TARGETS = tf_checkout tf_prebuild
+TF_BAZEL_OPTS = --jobs 192
+
+
+ONNXRT_VERSION = 1.17.3
+ONNXRT_OPTIONS = --use_cuda --cudnn_home=${CUDNN_INSTALL_PATH}
+ONNXRT_PREBUILD_TARGETS = onnxrt_checkout onnxrt_patch
+
+pytorch_checkout:
+	cd pytorch && \
+		git checkout v${PYTORCH_VERSION} && \
+		git submodule update --init --recursive
+
+pytorch_rocm_prebuild:
+	cd pytorch; python tools/amd_build/build_amd.py
+	sed -i 's/attr.memoryType/attr.type/g' pytorch/aten/src/ATen/hip/detail/HIPHooks.cpp
+	cd pytorch; git apply ../patches/pytorch/caffe2_rocm_path.patch
+
+tf_prebuild:
+	cd tensorflow; \
+		USE_DEFAULT_PYTHON_LIB_PATH=1 \
+		PYTHON_BIN_PATH=$$(which python) \
+		TF_NEED_CLANG=1 \
+		TF_NEED_ROCM=0 \
+		TF_NEED_CUDA=1 \
+		TF_NEED_TENSORRT=0 \
+		TF_CUDA_CLANG=0 \
+		CC_OPT_FLAGS="-Wno-sign-compare" \
+		TF_SET_ANDROID_WORKSPACE=0 \
+		python configure.py
+
+tf_checkout:
+	cd tensorflow; \
+		git fetch $(TF_REMOTE) $(TF_TAG) && \
+		git checkout FETCH_HEAD
+
+onnxrt_checkout:
+	cd onnxruntime && \
+		git checkout v$(ONNXRT_VERSION) && \
+		git reset --hard && \
+		git clean -xdf && \
+		git submodule update --init --recursive
+
+onnxrt_patch:
+	cd onnxruntime && \
+		git apply ../patches/onnxruntime/cfloat.patch