amf-core 0.1.0

Atomic Model Fragmentation (AMF) - Universal LLM Decomposition Library

Atomic Model Fragmentation (AMF)

المحرك الجزيئي — Molecular Engine

A revolutionary approach to running Large Language Models on resource-constrained hardware by breaking them down into functional, on-demand cells.

Installation

# Clone the repository
cd AI_NEW_GEN

# Install as a library (development mode)
pip install -e .

Quick Start (CLI)

1. Fragment a model:

   amf fragment --model qwen2.5:0.5b
   

2. Start Interactive Engine:

   amf chat
   

Python Library Usage (amf-core)

The AMF system provides a universal Python library abstraction allowing you to easily embed this technology in any Python application.

import amf

# 1. Load Universal Model (GGUF, Safetensors auto-detection)
model = amf.load_universal("/path/to/model.gguf")

# 2. Apply Genetic Fragmentation Strategy
cells = amf.fragment(
    model, 
    strategy="functional", 
    output_dir="./cells"
)

# 3. Explore Cellular DNA
for cell in cells.cells:
    print(f"Cell ID: {cell.cell_id}")
    print(f"DNA Tag: {cell.dna_tag}")
    print(f"Size: {cell.size_mb:.1f} MB")

System Architecture

AMF operates in 6 distinct phases:

1. Universal Parsing: ModelLoader reads formats like GGUF and translates them into a universal tensor abstraction.
2. Weight Analysis: WeightAnalyzer classifies weights by layer zone, functional role, and computes statistics.
3. DNA Tagging & Fragmentation: SortingAlgorithm chunks weights into semantic groups (e.g., A-L-003-Q for Attention-Linguistic-Layer3-QBlock).
4. Intent Analysis: IntentAnalyzer parses user prompts to determine required reasoning/semantic layer zones.
5. Molecular Orchestration: MolecularEngine triggers selective mmap loading of only the necessary cells into memory.
6. Dynamic Inference: InferenceEngine executes the forward pass on the loaded functional subset using NumPy.