Axon

A minimal deep learning framework built in C++20 from scratch with eager autograd, AVX2 SIMD CPU backend, and GGML-style quantization

Axon is a minimal deep learning framework in C++20 that implements a complete training and inference pipeline with zero external ML framework dependencies.

GitHub Repository

Architectural Highlights

  • C++20 Core & Eager Autograd: Polymorphic Node computational graph recording reverse-mode automatic differentiation natively in C++20.
  • AVX2 SIMD CPU Acceleration: Hand-tuned vectorization kernels via KernelRegistry for dense matrix multiplications (GEMM) and fused elementwise activations.
  • GGML-Style Quantization: Custom integer quantization routines enabling low-latency CPU inference without third-party runtimes.
  • End-to-End Deep Learning Pipeline: Complete implementations of Tensor abstractions, Neural Network Modules (Linear, Conv2d, BatchNorm), Optimizers (SGD, AdamW), and custom binary model serialization.