A transparent look at the 522-commit C++ AI runtime
Project Overview
SNEPPX-Alg is an open-source (MIT) AI runtime written in C++ with Python bindings. Its primary differentiator is a 10-layer security model (S0 to S9) that includes memory encryption (S1), control-flow obfuscation (S2), and runtime monitoring (S4).
Directory Layout (What goes where)
kernel/: Core tensor operations, automatic differentiation, and memory allocators.
algorithms/: Model implementations (Transformer, MoE, Mamba-2, GAN, Diffusion, RL). Currently, the Transformer and base MLP modules are functional. The rest are in active development.
drivers/: Hardware abstraction for CUDA, ROCm, Vulkan, Metal, TPU, and Intel oneAPI.
security/: The S0–S9 layers. S0 (Dilithium crypto) and S1 (AES-GCM memory) are implemented. S8 (Formal Verification) and S9 (Penetration Testing) are currently placeholders for future workflows.
net/: Distributed training coordinator using gRPC.
bindings/python/: Pybind11 wrappers exposing the engine to Python.
Current Stability Status
Stable: CPU backend, core tensor operations (MatMul, ReLU, Softmax, LayerNorm), Dilithium signing/verification, ONNX import (basic ops).
Experimental: CUDA/ROCm backends, ZeRO-1 distribution, Mamba-2 and Diffusion model stubs.
Need Help: The AWQ/GPTQ quantization passes and the ONNX exporter for complex graphs (Resize, Gather) need debugging.
How to Build and Test (Right Now)
Clone the repository and run the minimal CPU build to verify your environment:
bash
git clone https://github.com/ammar49-cyber/sneppx-alg.git
cd sneppx-alg
cmake -B build -G Ninja -DCMAKE_BUILD_TYPE=Release
cmake --build build --config Release
cd build && ctest -C Release --output-on-failure
If this passes, you have the core engine running.
Specific Contributions Needed
Python Bindings: The from_pretrained() API works but lacks support for safetensors files. If you know Python/C++ interop, this is a high-impact, 200-line fix.
Kernel Optimization: The kernel/ops/matmul.cpp currently uses naive loops for the CPU fallback. Implementing a blocked matrix multiplication (tiled) would improve CPU inference speed by ~40%.
Documentation: The Doxygen comments are sparse. I need help adding clear API documentation for the public functions in bindings/python/src/.
My Commitment as Maintainer
I review every pull request within 48 hours. I tag issues with good-first-issue and help-wanted to make onboarding easier. This is a solo project, but I am actively building it in public and responding to every contributor.
Repository:
ammar49-cyber
/
sneppx-alg
The Universal open source AI algorithm!
SNEPPX-Alg: Secure Neural Architecture (ARIX_Algo)
ARIX_Algo — Secure, composable, production-grade AI algorithm pipeline with 10 security layers (S0–S9), model zoo, distributed training, quantization, and advanced architectures.
This directory contains the SNEPPX-Alg cognitive processing system — a next-generation AI framework with security built into the foundation.
See the top-level README for the full overview
For complete documentation, start at docs/index.md.
Features
- 5-component algorithm pipeline: HSS (SSM), SER (MoE), ARC (Adversarial Guard), NPE (Neural VM), FM (Federated Memory)
- 10 security layers (S0–S9): Crypto, Secure Memory, Obfuscation, Monitoring, Network, AI Sanitizer, Key Vault, Updates, Formal Verification, Penetration Testing
-
Model Zoo:
from_pretrained()API with ModelHub, weight management, model cards, converter presets for LLaMA 2/3, Mistral, Qwen 2, DeepSeek V2 - Distributed Training: ZeRO-1/2/3, pipeline/tensor/expert parallelism, elastic training, fault tolerance
- Quantization: INT8/INT4/FP8, AWQ, GPTQ
- Advanced Architectures: Differential Attention, Mamba-2 SSM, FlexAttention, Mixture of Depth
- ONNX import/export: standalone numpy-only…
Top comments (0)