torchlite-cpp

A lightweight C++ tensor library implementing core PyTorch-like operations for CPU. Designed as a minimal foundation for building neural networks without the complexity of full frameworks.

Core Features

Tensor Operations

• N-dimensional tensor data structure with automatic stride calculation
• Element-wise operations (add, sub, mul, div, neg, exp, log, pow, sqrt, abs, clamp) with NumPy-style broadcasting
• Matrix multiplication with batch dimension support; flash attention
• Convolution and pooling: conv2d, max_pool2d, avg_pool2d
• Reductions: sum, mean, variance, argmax, softmax
• Reshape, transpose, cat, stack, slice, pad

Neural Network Modules

• Linear, Conv2d, MaxPool2d, AvgPool2d, Flatten
• LayerNorm, BatchNorm2d, InputNormalize, Dropout
• MultiHeadAttention, TransformerEncoderLayer, TransformerEncoder
• PositionalEncoding
• Activation functions: ReLU, GELU, Sigmoid

Training

• Reverse-mode autograd (backward computation graph) over all operations
• Optimizers: SGD, Adam, AdamW
• Loss functions: MSE, MAE, BCE, NLL, Cross-entropy

Design

• CPU-only, float32 operations
• Weight loading from raw binary files

Use in your project (CMake)

Pull torchlite into your CMake project with FetchContent:

include(FetchContent)

FetchContent_Declare(
    torchlite
    GIT_REPOSITORY https://github.com/joshuazhou744/torchlite-cpp
    GIT_TAG main
)
FetchContent_MakeAvailable(torchlite)

target_link_libraries(myapp PRIVATE torchlite)

Then #include <tl/tensor.h> in your code. The external dependencies (e.g., Eigen3) will propagate automatically.

Build from source

mkdir build
cmake -S . -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build

The Release build enables optimizations (-O3 march=native -DNDEBUG):

• -O3: aggressive compiler optimizations
• -march=native: targets CPU specific instruction set
• -DNDEBUG: disables assert() calls in hot paths

Run tests (development):

./build/run_tests

Project Structure

include/tl/         Public API headers (tensor, ops, nn, activation, factory, autograd)
include/external/   Third-party headers (LibrosaCpp)
src/                Implementation
tests/              Test executables
bench/              Operation benchmarks
examples/           Example usages of the library

Examples

CNN Binary Classifier

See here for a CNN binary classifier built using torchlite-cpp.

Dependencies

• Eigen3: required by LibrosaCpp for audio preprocessing
• LibrosaCpp: single-header mel spectrogram computation (included in include/external/)
• OpenMP: for the multithreaded GEMM kernel

Requirements

• C++17 or later
• CMake 3.16+
• Eigen3 (sudo apt install libeigen3-dev)
• OpenMP (sudo apt install libgomp1)

License

MIT License