wfw-ai 1.0.0

WANI — Wave-Attractor Neural Inference Framework

WANI — Wave-Attractor Neural Inference

Powered by Zain Ali

Run any GGUF model locally. Drop model in folder, run one command.

---

Folder Structure

WANI/
├── wani.py          ← Main tool (run this)
├── requirements.txt
├── models/          ← Put your .gguf files HERE
│   └── (empty)
├── logs/
│   └── wani.log     ← Auto-created
└── core/
    ├── config.py
    ├── scanner.py
    ├── loader.py
    └── __init__.py

---

Setup (3 steps)

Step 1 — Install Python 3.8+

python --version

Step 2 — Install backend

# CPU only (works on all systems including ARM/Android)
pip install llama-cpp-python

# OR with GPU support (Vulkan — works on Adreno 610/650)
CMAKE_ARGS="-DLLAMA_VULKAN=on" pip install llama-cpp-python --force-reinstall

Step 3 — Add your model

Download any GGUF model from:
  https://huggingface.co/models?search=gguf

Put the .gguf file inside:  WANI/models/

Example models to try:
  • Llama-3.2-1B-Instruct-Q4_K_M.gguf    (~700 MB)
  • Llama-3.2-3B-Instruct-Q4_K_M.gguf    (~1.8 GB)
  • Mistral-7B-Instruct-v0.3.Q4_K_M.gguf (~4.1 GB)
  • Phi-3-mini-4k-instruct-Q4_K_M.gguf   (~2.2 GB)

---

Usage

# Interactive mode (auto-scan + pick model)
python wani.py

# List all models
python wani.py --list

# Load specific model, start chat
python wani.py --model Llama-3.2-1B-Instruct-Q4_K_M.gguf --chat

# Single prompt
python wani.py --model yourmodel.gguf --prompt "Pakistan ki capital kya hai?"

# With custom settings
python wani.py --model yourmodel.gguf --threads 4 --ctx 4096 --temp 0.8

---

Options

Flag	Default	Description
--model FILE	auto-pick	Model filename from models/
--prompt TEXT	—	Single prompt, then exit
--chat	—	Force chat mode
--ctx N	2048	Context window (tokens)
--threads N	4	CPU threads
--gpu N	0	GPU offload layers (0=CPU)
--temp F	0.7	Temperature (creativity)
--max N	512	Max tokens to generate
--list	—	List models and exit
--install	—	Show install instructions

---

Recommended Models by RAM

RAM	Model	Download
1 GB	Llama-3.2-1B Q4	HuggingFace
2 GB	Phi-3-mini Q4	HuggingFace
3 GB	Llama-3.2-3B Q4	HuggingFace
4 GB	Mistral-7B Q4	HuggingFace
8 GB	LLaMA-13B Q4	HuggingFace

---

Snapdragon 685 Tips

# Use 4 big cores only
python wani.py --threads 4

# Small context saves RAM
python wani.py --ctx 1024

# 1B or 3B models run best
# Use Q4_K_M quantization (best quality/size ratio)

---

Chat Commands

While in chat mode:

• stats — show session stats (TPS, tokens generated)
• clear — reset conversation history
• help — show commands
• quit — exit

---

WANI Framework v1.0.0 — Powered by Zain Ali