TorchLessCUDA is a deep learning framework built entirely from scratch using Modern C++20 and CUDA, implementing tensor operations, GPU training, and inference without relying on existing machine learning frameworks.
Modern deep learning frameworks abstract away GPU programming, memory management, tensor operations, automatic differentiation, and optimization.
TorchLessCUDA focuses on implementing these components manually while maintaining a modular, extensible, and production-inspired architecture.
The framework supports end-to-end neural network training and inference entirely on the GPU through custom CUDA kernels, providing a deeper understanding of how modern deep learning systems operate internally.
- NVIDIA GPU with CUDA support
- C++20 compatible compiler
- NVIDIA CUDA Toolkit 12.x
- CMake 3.25+
- Visual Studio 2022 (Windows)
TorchLessCUDA was built as a systems-oriented learning project to understand how deep learning frameworks work beneath high-level APIs.
Instead of relying on existing ML libraries, the framework implements its own tensor engine, execution pipeline, CUDA kernels, neural network layers, optimization algorithms, and GPU memory management using Modern C++20 and CUDA.
The primary goal is to bridge the gap between deep learning theory and low-level systems implementation while providing a modular foundation for future experimentation and extension.