aurestral 1.0.1

Local GGUF AI inference library built on llama-cpp-python with hardware auto-tuning

Aurestral

Placeholder Local GGUF inference for Python, powered by llama-cpp-python. Aurestral discovers models in your project’s models/ folder, auto-tunes thread counts, context size, and GPU offload for your hardware, and ships with an interactive chatbot CLI. To be evolved, stay tuned - Aurestral Console.

Requirements

• Python 3.9–3.12 (3.13+ may work; 3.14 is not supported yet — no llama-cpp-python wheels)
• A GGUF model file (e.g. from Hugging Face)

Installation

Recommended (Windows)

PyPI only ships a source tarball for llama-cpp-python, which often fails on Windows (long paths). Install a prebuilt wheel first, then Aurestral:

# Use Python 3.11 or 3.12 (not 3.14)
py -3.12 -m venv .venv
.\.venv\Scripts\Activate.ps1

# Prebuilt CPU wheel (fast, no compile)
pip install llama-cpp-python --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cpu

pip install --upgrade aurestral

NVIDIA GPU (e.g. RTX 4060) — use the CUDA wheel index instead of CPU:

pip install llama-cpp-python --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cu124
pip install --upgrade aurestral

Simple install (Linux / macOS)

pip install aurestral

On macOS, prebuilt wheels usually include Metal acceleration.

If install still fails on Windows

1. Enable long path support in Windows settings.
2. Use a short temp folder before installing:

   New-Item -ItemType Directory -Force C:\tmp | Out-Null
   $env:TEMP = "C:\tmp"
   $env:TMP = "C:\tmp"
   pip install llama-cpp-python --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cpu
   

3. Confirm Python version: python --version should show 3.12.x, not 3.14.

Project layout

Place GGUF files in a models/ directory at your project root (or set AURESTRAL_MODELS_DIR):

my-project/
├── models/
│   └── llama-3.2-3b-instruct.Q4_K_M.gguf
└── main.py

Quick start

Interactive chatbot

cd my-project
aurestral
# or explicitly:
aurestral chat -m llama-3.2-3b-instruct.Q4_K_M.gguf

Chat commands: /help, /clear, /exit

Python API

from aurestral import load_model, ChatSession, generate

# One-shot completion
text = generate("Explain quantum entanglement in one sentence.")
print(text)

# Reusable model handle
model = load_model()  # auto-picks sole GGUF, or pass name="my-model"
reply = model.chat([
    {"role": "user", "content": "Hello!"},
])
print(reply)

# Multi-turn session with streaming
session = ChatSession.create(system_prompt="You are a concise coding assistant.")
session.send("Write a Python hello world.", stream=True)

List models and hardware info

aurestral list
aurestral info
aurestral run "The capital of France is" --stream

Hardware auto-tuning

On load, Aurestral inspects CPU cores, RAM, and whether llama-cpp-python was built with GPU offload support. It sets:

Setting	Behavior
n_threads	Physical cores minus one
n_ctx	1k–8k based on available RAM
n_gpu_layers	-1 (all layers) when GPU offload is available
use_mlock	Enabled on high-RAM CPU-only setups
flash_attn	Enabled when GPU offload is available

Override defaults with InferenceConfig or auto_tune=False:

from aurestral import InferenceConfig, load_model

cfg = InferenceConfig(n_ctx=8192, n_gpu_layers=35)
model = load_model("my-model.gguf", config=cfg, auto_tune=False)

Configuration reference

Environment

• AURESTRAL_MODELS_DIR — path to models folder (instead of ./models)

InferenceConfig — load-time: n_ctx, n_batch, n_threads, n_gpu_layers, use_mmap, use_mlock, flash_attn

GenerateConfig — generation-time: max_tokens, temperature, top_p, top_k, repeat_penalty, stop, stream

License

MIT License — see LICENSE.