organ-architecture/README.md

57 lines
2.0 KiB
Markdown

<div align="center">
# Organ Architecture
**Neural network surgery and checkpoint manipulation for GGUF models**
[![License: BSL 1.1](https://img.shields.io/badge/License-BSL%201.1-orange.svg)](LICENSE)
[**inference-x.com**](https://inference-x.com) · [**Community**](https://git.inference-x.com/inference-x-community)
</div>
---
## What is Organ Architecture?
Organ Architecture treats neural networks like biological organisms — with organs (layers, heads, MLP blocks) that can be extracted, analyzed, transplanted, and recombined.
## Features
- **Model inspection** — visualize layer shapes, attention patterns, weight distributions
- **Selective extraction** — extract specific layers or attention heads as "organs"
- **Cross-model transplant** — graft organs between compatible GGUF models
- **Quantization surgery** — selectively re-quantize individual layers
- **Architecture diff** — compare two model checkpoints layer-by-layer
## Quick Start
```bash
git clone https://git.inference-x.com/inference-x/organ-architecture
cd organ-architecture && make
# Inspect a model
./organ inspect model.gguf
# Extract attention layers 16-24
./organ extract model.gguf --layers 16-24 --output head_block.gguf
# Transplant into another model
./organ transplant donor.gguf recipient.gguf --source-layers 16-24 --target-layers 16-24
```
## Supported Operations
| Operation | Description |
|---|---|
| `inspect` | Show model architecture, layer shapes, parameter count |
| `extract` | Pull specific organs (layers, heads, MLPs) |
| `transplant` | Merge organs from one model into another |
| `diff` | Compare two models layer by layer |
| `requant` | Re-quantize selected layers |
| `prune` | Remove layers to create smaller variants |
---
*Part of the [Inference-X ecosystem](https://inference-x.com) · See also: [organ-store](https://git.inference-x.com/inference-x-community/organ-store), [organ-architect](https://git.inference-x.com/inference-x-community/organ-architect)*