Project PRIMAL: 4-bit Prime-Harmonic Training Engine

Status: Active Research / Proof of Concept Hardware Target: Consumer GPUs (e.g., GTX 1080 Ti, RTX 3060) License: MIT

🚀 The 11GB Challenge

Training Large Language Models (LLMs) usually requires massive VRAM because of the **Shadow Weight Tax- Neural Vocoder: HiFi-GAN v3 integration via SpeechBrain.

• Automation: voice_test.py for batch inference verification.

🚀 Quick Start (Inference)

To generate speech from text, ensure you are in the project root and run:

# Single Sentence
python tts_inference.py --text "Project Trinity is alive." --checkpoint "checkpoints/ghost_tts/best_sentinel.pt" --output "output.wav"

# Automated Batch Test
python voice_test.py

Results will be saved in tests/voice_samples/. ing (QAT) keeps the model in 4-bit but maintains a full FP16/FP32 copy of the weights for updates, effectively doubling memory usage.

Project PRIMAL removes the shadow weights entirely.

It implements a Discrete Optimization Loop that trains a 0.1B parameter model directly on a rigid 4-bit integer grid. This allows for massive batch sizes and high throughput on older cards like the GTX 1080 Ti.

---

⚡ Key Features

1. Prime Harmonic Grid (v3.0.0)

Instead of linear INT4 quantization (which wastes precision on large numbers), PRIMAL uses a custom 13-value Look-Up Table (LUT) derived from prime reciprocals. This concentrates precision around zero, where 90% of LLM weights reside.

Index	Value	Description
0-6	±1, ±0.5, ±0.33...	Coarse adjustment for "body" layers.
7	0.0	Exact zero for sparsity.
Fine	±0.66, ±0.25...	(Layer 12 Only) High-precision bridge.

2. The "Poltergeist"# Antigravity: Shadowless 8-bit Discrete Training (Linear Protocol)

Discrete training often fails due to "stochastic thrashing" during gradient accumulation. PRIMAL solves this with Decoupled Flipping:

• Backward Pass: No updates. Gradients cast "votes" (+1 or -1) into an int8 buffer.
• Optimizer Step: Votes are aggregated. Weights only flip if there is a consensus across micro-batches (e.g., Batch 64).
• Adaptive Probability: Weights flip stochastically based on their magnitude (Z-Score filtering).

3. Efficiency Benchmarks (GTX 1080 Ti)

Metric	Model Size	Result	Notes
Training VRAM	0.1B Params	10.3 GB	@ Batch Size 64 (Full Saturation)
Inference VRAM	1.1B Params	~550 MB	4-bit Loading Verified
Throughput	0.1B Params	~6,000 TPS	Training (Python Loop)

---

🛠️ Installation & Usage

Prerequisites

• NVIDIA GPU (Pascal or newer)
• CUDA 11.8+
• Python 3.10+

Quick Start

# Clone the repo
git clone https://github.com/batteryphil/Primal-Discrete-LLM-Training.git
cd Primal-Discrete-LLM-Training

# Install dependencies
pip install -r requirements.txt

# Run the training demo (0.1B Model)
python primal_train_ghost.py