Hands-on machine learning with C++ 2nd edition

cuda version on local

$ nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2026 NVIDIA Corporation
Built on Tue_Jun_09_02:43:40_PM_PDT_2026
Cuda compilation tools, release 13.3, V13.3.73
Build cuda_13.3.r13.3/compiler.38244171_0

local development environment

installing several dependencies

# Basic tools and development packages
sudo dnf install -y unzip
sudo dnf install -y gcc gcc-c++ make  # equivalent to build-essential
sudo dnf install -y gdb
sudo dnf install -y git
sudo dnf install -y cmake
sudo dnf install -y cmake-gui  # equivalent to cmake-curses-gui
sudo dnf install -y python3
sudo dnf install -y python3-pip
sudo dnf install -y blas-devel  # equivalent to libblas-dev
sudo dnf install -y openblas-devel  # equivalent to libopenblas-dev
sudo dnf install -y fftw-devel  # equivalent to libfftw3-dev
sudo dnf install -y atlas-devel  # equivalent to libatlas-base-dev
sudo dnf install -y lapack-devel  # equivalent to liblapack-dev and liblapacke-dev
sudo dnf install -y boost-devel  # equivalent to libboost-all-dev

# OpenCV packages - these may need adjustment based on version available in repos
sudo dnf install -y opencv-core  # equivalent to libopencv-core4.5d
sudo dnf install -y opencv-devel # development files
sudo dnf install -y opencv  # main package

# Additional libraries
sudo dnf install -y hdf5-devel  # equivalent to libhdf5-dev
sudo dnf install -y json-c-devel  # equivalent to libjson-c-dev
sudo dnf install -y libX11-devel  # equivalent to libx11-dev
sudo dnf install -y java-1.8.0-openjdk-devel  # equivalent to openjdk-8-jdk
sudo dnf install -y wget
sudo dnf install -y ninja-build
sudo dnf install -y gnuplot
sudo dnf install -y vim
sudo dnf install -y python3-virtualenv  # equivalent to python3-venv


sudo dnf install ocl-icd-devel opencl-headers

sudo dnf install fmt-devel
sudo dnf install nnpack-devel
sudo dnf install mlpack-devel

# arrayfire
sudo dnf install spdlog-devel

pip3 install pyyaml
pip3 install typing
pip3 install typing_extensions

flashlight

cat > flashlight_algo_fix.patch << 'EOF'
diff --git a/flashlight/fl/tensor/backend/af/ArrayFireReductions.cpp b/flashlight/fl/tensor/backend/af/ArrayFireReductions.cpp
index xxxxxxx..xxxxxxx 100644
--- a/flashlight/fl/tensor/backend/af/ArrayFireReductions.cpp
+++ b/flashlight/fl/tensor/backend/af/ArrayFireReductions.cpp
@@ -13,6 +13,7 @@
 #include "flashlight/fl/tensor/backend/af/ArrayFireTensor.h"

 #include <stdexcept>
+#include <algorithm>

 #include "flashlight/fl/tensor/Index.h"
 #include "flashlight/fl/tensor/backend/af/Utils.h"
diff --git a/flashlight/fl/tensor/TensorBase.cpp b/flashlight/fl/tensor/TensorBase.cpp
index xxxxxxx..xxxxxxx 100644
--- a/flashlight/fl/tensor/TensorBase.cpp
+++ b/flashlight/fl/tensor/TensorBase.cpp
@@ -10,6 +10,7 @@
 #include "flashlight/fl/tensor/TensorBase.h"

 #include <cstring>
+#include <algorithm>
 #include <memory>
 #include <numeric>
 #include <sstream>
EOF

cd /media/wdhome/github/khofesh/machine-learning-cpp/development/libs/sources/flashlight/
patch -p1 < /path/to/flashlight_algo_fix.patch

flashlight (CUDA 13)

main (commit 439bb526, "Fix CUDA 13 nvToolsExt build break") builds with CUDA 13.3 through ArrayFire's CUDA backend. The GCC 16 host compiler still needs a few fixes; they are kept as development/flashlight_main_cuda.patch:

CMakeLists.txt: CMAKE_CXX_STANDARD 17 -> 20 (std::unordered_map::contains)
cmake/flashlightConfig.cmake.in: find_dependency(CUDAToolkit) so consumers get CUDA::cublas
TensorBase.h: add <cstdint> (GCC 16 no longer includes it transitively)
backend/jit/ShapeInference.cpp: add <algorithm> for std::reverse

Build/install (NVCC 13.3 rejects GCC > 15, hence NVCC_PREPEND_FLAGS; this also applies to CMake's CUDA compiler detection, where CMAKE_CUDA_FLAGS is not used):

cd development
patch -p1 -d libs/sources/flashlight < flashlight_main_cuda.patch   # if starting fresh
export NVCC_PREPEND_FLAGS="-allow-unsupported-compiler"
FORCE_REBUILD=1 . ./install_lib.sh https://github.com/flashlight/flashlight.git main \
  -DFL_BUILD_TESTS=OFF -DFL_BUILD_EXAMPLES=OFF \
  -DFL_USE_CUDA=ON -DFL_USE_CPU=OFF -DFL_USE_ONEDNN=OFF \
  -DArrayFire_DIR=/opt/ArrayFire-3.10.0-Linux/share/ArrayFire/cmake/ \
  -DFL_ARRAYFIRE_USE_CUDA=ON -DFL_ARRAYFIRE_USE_CPU=OFF \
  -DFL_BUILD_DISTRIBUTED=OFF -DCMAKE_CUDA_ARCHITECTURES=86

At runtime ArrayFire's bundled libnvrtc.so.12 needs libnvrtc-builtins.so.12.2 from /opt/ArrayFire-3.10.0-Linux/lib64, which is not in the loader cache, so add it once:

echo '/opt/ArrayFire-3.10.0-Linux/lib64' | sudo tee /etc/ld.so.conf.d/arrayfire.conf
sudo ldconfig

Only one ArrayFire backend may be linked at a time: FL_ARRAYFIRE_USE_CUDA=ON requires FL_ARRAYFIRE_USE_CPU=OFF (and vice versa), otherwise CMake errors out.

nccl

download nccl from here https://developer.nvidia.com/nccl

# adjust the +cudaXX suffix to match the installed toolkit (CUDA 13.3 here)
sudo dnf install nccl-local-repo-rhel9-2.26.5-cuda12.9-1.0-1.x86_64.rpm
sudo dnf  install libnccl-2.26.5-1+cuda12.9 libnccl-devel-2.26.5-1+cuda12.9 libnccl-static-2.26.5-1+cuda12.9

pytorch

down the zip file here https://pytorch.org/get-started/locally/

or https://download.pytorch.org/libtorch/cu132/libtorch-shared-with-deps-2.14.0%2Bcu132.zip

wget https://download.pytorch.org/libtorch/cu132/libtorch-shared-with-deps-2.14.0%2Bcu132.zip
mv libtorch-shared-with-deps-2.14.0+cu132.zip development/libs
cd development/libs
unzip libtorch-shared-with-deps-2.14.0+cu132.zip

libtorch 2.14 requires C++20 (check_torch is built with CXX_STANDARD 20). CUDA 13.3 rejects GCC > 15 as host compiler, so the root CMakeLists.txt adds -allow-unsupported-compiler to CMAKE_CUDA_FLAGS.

run check_torch if the lib is working

arrayfire

download arrayfire here https://arrayfire.com/binaries/

cd development
mkdir arrayfire
wget https://arrayfire.s3.amazonaws.com/3.10.0/ArrayFire-v3.10.0_Linux_x86_64.sh
chmod +x ArrayFire-v3.10.0_Linux_x86_64.sh

# https://arrayfire.org/docs/installing.htm#gsc.tab=0
sudo ./ArrayFire-v3.10.0_Linux_x86_64.sh --include-subdir --prefix=/opt
# env
echo 'export LD_LIBRARY_PATH=/opt/ArrayFire-3.10.0-Linux/lib64:$LD_LIBRARY_PATH' >> ~/.bashrc
source ~/.bashrc
# or
sudo echo '/opt/ArrayFire-3.10.0-Linux/lib64' | sudo tee /etc/ld.so.conf.d/arrayfire.conf
sudo ldconfig

vcpkg

git clone https://github.com/Microsoft/vcpkg.git
cd vcpkg

./bootstrap-vcpkg.sh  # on linux
./vcpkg install dlib

VS Code Settings

Open VS Code settings (Ctrl+, or Cmd+,)
Search for "cmake configure environment"
Add the environment variable in CMake: Configure Environment:

json{
  "VCPKG_ROOT": "/home/fahmad/vcpkg"
}

chapter 2

# cd to your vcpkg
cd /to/your/path
# optional
./vcpkg upgrade --no-dry-run

# json
./vcpkg install nlohmann-json

# fast-cpp-csv-parser
./vcpkg install fast-cpp-csv-parser

# mlpack
./vcpkg install mlpack

# HighFive
./vcpkg install highfive

# hdf5
./vcpkg install hdf5

dlib (GUI support)

vcpkg's dlib port is built with DLIB_NO_GUI_SUPPORT=ON, so the GUI example (chapter002/img/dlib) cannot link. Build the bundled dlib source with X11 GUI support and install it into development/libs:

cmake -S development/libs/sources/dlib -B development/libs/sources/dlib/build \
  -DCMAKE_BUILD_TYPE=Release \
  -DDLIB_NO_GUI_SUPPORT=OFF \
  -DBUILD_SHARED_LIBS=OFF \
  -DCMAKE_INSTALL_PREFIX=$PWD/development/libs
cmake --build development/libs/sources/dlib/build --parallel
cmake --install development/libs/sources/dlib/build

homlcpp/chapter002/img/dlib/CMakeLists.txt points dlib_DIR at this install (development/libs/lib64/cmake/dlib); the other dlib examples keep using vcpkg's build.

build notes for newer packages

HDF5 2.2.0 (vcpkg): use config mode (find_package(HDF5 CONFIG REQUIRED)) so the imported target pulls in zlib/libaec transitively. Module mode only links libhdf5.a, which fails with undefined inflate/SZ_* symbols.
mlpack 4.8: the mlpack::data:: namespace was removed; use mlpack::DatasetInfo, mlpack::Load, mlpack::MinMaxScaler, mlpack::StandardScaler.
libtorch 2.14: requires C++20.
CUDA 13.3 + GCC 16: nvcc rejects GCC > 15 as host compiler; the root CMakeLists.txt appends -allow-unsupported-compiler to CMAKE_CUDA_FLAGS.
GUI examples (img-opencv, img-dlib) are interactive; run them under xvfb-run for headless smoke tests.

I haven't succeeded in installing flashlight-cuda or flashlight-cpu using vcpkg, this is the steps I've tried

https://www.if-not-true-then-false.com/2024/install-nvidia-cudnn-on-fedora/

mkdir cudnn
cd cudnn
wget https://developer.download.nvidia.com/compute/cudnn/redist/cudnn/linux-x86_64/cudnn-linux-x86_64-9.10.2.21_cuda12-archive.tar.xz # https://developer.nvidia.com/cudnn-downloads?target_os=Linux&target_arch=x86_64&Distribution=Agnostic&cuda_version=12
tar xvf cudnn-linux-x86_64-9.10.2.21_cuda12-archive.tar.xz
cd cudnn-linux-x86_64-9.10.2.21_cuda12-archive
# pick the archive matching your toolkit (cuda13 for CUDA 13.3); /usr/local/cuda points at the active one
sudo cp include/cudnn*.h /usr/local/cuda/include/
sudo cp lib/libcudnn* /usr/local/cuda/lib64/
sudo chmod a+r /usr/local/cuda/include/cudnn*.h /usr/local/cuda/lib64/libcudnn*
export ArrayFire_DIR=/media/wdhome/github/khofesh/machine-learning-cpp/development/arrayfire
export LD_LIBRARY_PATH=/media/wdhome/github/khofesh/machine-learning-cpp/development/arrayfire/lib64:$LD_LIBRARY_PATH
export CXXFLAGS="-Wno-template-id-cdtor"
export CFLAGS="-Wno-template-id-cdtor"
cd ~/vcpkg
./vcpkg install flashlight-cuda

# flashlight cpu

folder structure

tree -I "Hands-on-Machine-learning-with-C-Second-Edition|LearningModCppFinance" -L 4

references

C++ Implementation of PyTorch Tutorials for Everyone - https://github.com/prabhuomkar/pytorch-cpp
Learning Modern C++ for Finance - https://github.com/QuantDevHacks/LearningModCppFinance/
distrobox - https://github.com/89luca89/distrobox?tab=readme-ov-file

Name		Name	Last commit message	Last commit date
.claude		.claude
.vscode		.vscode
Hands-on-Machine-learning-with-C-Second-Edition @ 0fec0f0		Hands-on-Machine-learning-with-C-Second-Edition @ 0fec0f0
LearningModCppFinance @ 471a459		LearningModCppFinance @ 471a459
check_torch		check_torch
datasets		datasets
development		development
env_scripts		env_scripts
graphify-out		graphify-out
homlcpp		homlcpp
.gitattributes		.gitattributes
.gitignore		.gitignore
.gitmodules		.gitmodules
.graphifyignore		.graphifyignore
AGENTS.md		AGENTS.md
CLAUDE.md		CLAUDE.md
CMakeLists.txt		CMakeLists.txt
README.md		README.md
requirements.txt		requirements.txt

Navigation Menu

Repository files navigation

Hands-on machine learning with C++ 2nd edition

cuda version on local

local development environment

flashlight

flashlight (CUDA 13)

nccl

pytorch

arrayfire

vcpkg

dlib (GUI support)

build notes for newer packages

folder structure

references

About

Releases

Packages

Used by

Contributors

Languages

Latest commit

History

Folders and files

Repository files navigation

Hands-on machine learning with C++ 2nd edition

cuda version on local

local development environment

flashlight

flashlight (CUDA 13)

nccl

pytorch

arrayfire

vcpkg

dlib (GUI support)

build notes for newer packages

folder structure

references

About

Resources

Stars

Watchers

Forks

Releases

Packages

Used by

Contributors

Languages