| # OpenVinayaka Engine (v1.0) |
|
|
| **The Universal Hallucination-Free Inference Engine.** |
|
|
| OpenVinayaka Engine is a C++ inference runtime designed to replace `llama.cpp` and `vLLM` for mission-critical applications. Unlike standard engines that focus only on token generation speed, OpenVinayaka prioritizes **Factual Integrity** by mathematically intervening in the model's internal state using the OV-Memory Priority Formula. |
|
|
| ## 🏗️ Supported Architectures |
| This engine provides "Universal Kernels" to fix hallucinations in: |
|
|
| 1. **Transformers (Llama 3, Gemma 2, GPT-Neo):** |
| * **Mechanism:** Attention Steering. |
| * **Logic:** Injects a bias mask into $QK^T$ to force attention onto verified context. |
| 2. **State Space Models (Mamba 1/2):** |
| * **Mechanism:** State Correction. |
| * **Logic:** Linearly interpolates the hidden state $h_t$ towards a "Truth Vector" to prevent drift. |
| 3. **Mixture of Experts (DeepSeek-V3, Mixtral):** |
| * **Mechanism:** Router Bias. |
| * **Logic:** Identifies "Factual Experts" and biases the Gating Network to select them. |
| 4. **Hybrid Architectures (Jamba, Samba):** |
| * **Mechanism:** Interleaved Correction. |
| * **Logic:** Applies State Correction in SSM layers and Attention Steering in Transformer layers. |
| |
| ## 🚀 Key Features |
| * **Zero-Hallucination Guarantee:** (When supported by OV-Memory Graph). |
| * **CPU/GPU Hybrid:** Graph walk on CPU, Matrix Math on GPU. |
| * **MIT Licensed:** Free for research and commercial use. |
| * **Single File Deployment:** Compatible with future `OV-GGUF` format. |
| |
| ## 🛠️ Building |
| ```bash |
| make |
| ./ov_engine_full |
| ``` |
| |
| ## 📂 Structure |
| * `kernels/`: The math (Universal Kernel header). |
| * `src/`: The engine logic and block manager. |
| * `include/`: Shared headers. |
| * `examples/`: Sample integrations. |
| |
| --- |
| *Dedicated to Om Vinayaka.* |
| |
