rriftt_ai.h is a minimal, dependency-free AI library implemented in pure C, designed for developers who require a straightforward and efficient solution for creating neural networks. This single-header C23 library serves as a bare-metal neural network engine, alleviating the complexities associated with modern deep learning frameworks that often rely on extensive dependencies and complex build systems.

Key Features

• Zero Dependencies: No need for external libraries or complex installations. Simply include rriftt_ai.h in your project to get started.
• Control Over Memory: Utilizes a custom memory arena through RaiArena, ensuring efficient memory management without any dynamic allocation during execution. This guarantees predictable performance in your applications.
• Modern C Architecture: Built with strict C23 compliance, this library integrates advanced modern programming practices.
• Comprehensive Transformer Support: Features robust implementations of RoPE, RMSNorm, SwiGLU, and Scaled Dot-Product Attention, allowing for sophisticated model designs.
• Integrated Training Engine: Offers built-in routines for backpropagation along with Cross-Entropy loss and AdamW optimizer, all encapsulated within C structures, facilitating seamless training workflows.
• Native Tokenization: Supports Byte-Pair Encoding (BPE) directly in C for efficient data preprocessing.

Quick Start

Developers can easily start utilizing the library with limited code, as demonstrated in the following example:

#define RRIFTT_AI_IMPLEMENTATION
#include "rriftt_ai.h"

int main(void) {
    RaiArena arena = rai_arena_create(1024 * 1024);
    RaiTensor A = RAI_TENSOR_ALLOC_FILL(&arena, 2.0f, 2, 2);
    RaiTensor B = RAI_TENSOR_ALLOC_FILL(&arena, 3.0f, 2, 2);
    RaiTensor C = rai_tensor_add(&arena, A, B);
    rai_arena_destroy(&arena);
    return 0;
}

Compile your code using a standard C compiler with a simple command:

gcc main.c -o engine -lm
./engine

Active Development

rriftt_ai.h is actively developed, welcoming contributions for enhancements, optimizations, and new features. The project aims for scalability, allowing community involvement in its evolution.

This library is ideal for those who need a lightweight, efficient solution for AI model implementation without the overhead of traditional machine learning frameworks.