File size: 1,848 Bytes
31d9d4f | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 | # 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.*
|