SaiSudan/mahoraga-lite

🏎️ LiquidFormer (91M) — The Liquid Neural Network Engine

LiquidFormer-91M is a high-performance language model based on the Liquid Neural Network (LNN) architecture. It replaces the standard static Feed-Forward Networks (FFN) found in traditional transformers with Liquid Time-Constant (LTC) cells, allowing for adaptive temporal dynamics and efficient context handling.

🏗️ Architecture Detail

• Model Size: 91 Million Parameters
• Core Cell: Liquid Time-Constant (LTC) ODE-based integration steps.
• Attention: Grouped-Query Attention (GQA) with an 8:2 (4:1) Query-to-KV head ratio.
• Positional Embedding: Rotary Positional Embeddings (RoPE).
• Optimization: Custom "Super Monkey Patch" for dual Tesla T4 GPU environments, achieving up to 18,000 tok/s.

📚 Training Data

The model underwent a dual-stage training regime to ensure both logical reasoning and conversational fluency:

1. Stage 1: Logic & Code Foundation
   • Trained on a curated set of 20,000 Gemini-driven code tasks focused on Python development, algorithmic logic, and system design.
2. Stage 2: Conversational Alignment
   • Fine-tuned on the OpenAssistant Guanaco dataset to align precisely with human instructions and maintain conversational context.

🚀 Usage

You can use this model locally using the LiquidFormer class provided in the repository.

Loading the Model

from architecture.liquidformer import LiquidFormer

# Load from the vault (safetensors format)
model = LiquidFormer.from_pretrained("path/to/LiquidFormer-91M")

Text Generation

from inference.generator import TextGenerator
from tokenizer.tokenizer import LiquidTokenizer

tokenizer = LiquidTokenizer("path/to/tokenizer.json")
generator = TextGenerator(model, tokenizer)

response = generator.generate(
    prompt="[USER] Write a Python function for binary search.\n[ASSISTANT]",
    max_new_tokens=256,
    temperature=0.7
)
print(response)

📊 Performance

LiquidFormer is designed for efficiency. On a standard GTX 1650 (4GB VRAM), it maintains a minimal memory footprint while delivering low-latency inference thanks to its optimized GQA and LTC dynamics.

---

Built with ❤️ for advanced LNN research.