IbrahimAnaz/pure-lie-engine

Pure Algebraic Lie Engine with Multi-Token Lookahead

A breakthrough neural language engine operating strictly on continuous Lie groups and Lie algebras without standard neural network heuristics:

• Zero Elementwise Nonlinear Activations: Exactly 0 ReLU, GELU, Sigmoid, or Tanh.
• Zero Additive Biases: Exactly 0 linear shift biases (+ b).
• Lie Commutator Memory: Dynamic routing via matrix commutators $[H, K] = HK - KH$.
• Hamiltonian Symplectic Phase Clock: Continuous rotation group trajectories in $\mathrm{SO}(16)$.
• Multi-Token Lookahead (MTP): 4-step simultaneous horizon prediction ($x_{t+1}, x_{t+2}, x_{t+3}, x_{t+4}$).

---

Environment Setup & CUDA Installation

Prerequisites

• Python: 3.10, 3.11, 3.12, or 3.13
• GPU: Any NVIDIA GPU with Tensor Cores (Turing, Ampere, Ada Lovelace, Blackwell sm_120)

Quick Install

CUDA 12.4 (Recommended for RTX 40-series, RTX 50-series, T4):

pip install torch torchvision --index-url https://download.pytorch.org/whl/cu124
pip install transformers datasets huggingface_hub

CUDA 12.1:

pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121
pip install transformers datasets huggingface_hub

---

Loading Model Weights

import torch
from huggingface_hub import hf_hub_download

ckpt_path = hf_hub_download(
    repo_id="IbrahimAnaz/pure-lie-engine",
    filename="pure_lie_engine_epoch_040.pt"
)
weights = torch.load(ckpt_path, map_location="cpu", weights_only=True)
print("Successfully loaded Lie Engine weights:", list(weights.keys())[:5])