ethicallama 0.1.2

ethical, local-only inference wrapper for llama.cpp

ethicallama

A local-first, privacy-respecting LLM inference wrapper for running large language models entirely on your own hardware.

Features

• Local-Only Inference: Everything runs on your machine. No data ever leaves your computer unless you explicitly configure it otherwise.
• Multi-Engine Support: Use llama.cpp, whisper.cpp, or any custom inference engine via Jinja2-templated configuration.
• Multiple GPU Backends: Choose between Vulkan, ROCm, CUDA, or CPU inference.
• Built-in HTTP API: Optional FastAPI-powered REST API for remote inference.
• Model Indexing: Automatically discover and manage models across your configured directories.
• Configurable Telemetry: Telemetry is DISABLED by default. Opt-in only with explicit confirmation.

Quick Start

Prerequisites

• Python 3.10+
• Rust/Cargo (for building the native core)
• Git

Installation

Pick the install method that matches your workflow. ethicallama is local-only: nothing is sent to the network and no telemetry runs unless you explicitly enable it.

Using pip (recommended, once on PyPI)

The standard, works-everywhere install. Wheels are pre-built for Linux and macOS on Python 3.10+.

pip install ethicallama

# With API server support:
pip install "ethicallama[api]"

# With everything (API, HF pulling, Safetensors conversion):
pip install "ethicallama[all]"

Using uv (10-100× faster than pip)

uv is a Rust-based Python package manager — drop-in faster pip. Use it inside an existing virtualenv or for one-off pip-style installs.

# Inside a uv-managed venv
uv pip install ethicallama
uv pip install "ethicallama[api]"

# Run without permanently installing (one-shot)
uv run --with ethicallama ethllama run llama3.2

Using pipx (isolated CLI, no venv management)

pipx installs Python CLI tools in their own isolated virtualenv and exposes the ethllama executable on your PATH globally — ideal if you just want the CLI on your machine.

pipx install ethicallama

# With API server support:
pipx install "ethicallama[api]"

# Upgrade later:
pipx upgrade ethicallama

From source (latest code, full Rust core)

For the latest unreleased code, custom Rust core builds, or contributing to the project:

git clone --recursive https://github.com/luluthehungrycat/ethicallama
cd ethicallama

# Set up a venv (uv or stdlib venv both work)
uv venv && source .venv/bin/activate
# (or:  python3 -m venv .venv && source .venv/bin/activate)

uv pip install maturin ".[all]"
# (or:  pip install maturin ".[all]")

# Build the Rust extension and install the Python package
maturin develop --release

After installing, initialize the user config:

ethllama config --init

Optional extras

Extra	Adds
[api]	FastAPI server (ethllama serve) + uvicorn + pydantic
[pull]	HuggingFace Hub model pulling (ethllama pull)
[convert]	Safetensors → GGUF conversion (ethllama convert)
[all]	All of the above

Extras are stacked with commas, e.g. pip install "ethicallama[api,pull]".

Basic Usage

# Run a model with default settings
ethllama run ~/models/qwen2.5-7b-q4_k_m.gguf --prompt "What is the capital of France?"

# Use a specific GPU backend
ethllama run ~/models/model.gguf --gpu cuda --gpu-layers 32

# Enable the HTTP API
ethllama serve --host 127.0.0.1 --port 8080

# List all indexed models
ethllama index list

# Index a directory of models
ethllama index add ~/models

CLI Reference

Global Options

Option	Description
--config	Path to config file (default: ~/.ethllama/config.yaml)
--verbose	Enable verbose logging

Commands

Command	Description
run	Run inference with a model
serve	Start the HTTP API server
config	Manage configuration
index	Manage model index

run

ethllama run <model> [options]

Options:

Option	Default	Description
--prompt, -p	"Hello"	Input prompt
--temperature, -t	0.7	Sampling temperature
--top-p	0.95	Top-p sampling
--top-k	40	Top-k sampling
--threads	4	Number of CPU threads
--gpu	cpu	GPU backend: vulkan, rocm, cuda, cpu
--gpu-layers	0	Number of layers offloaded to GPU
--engine	llama-cpp	Engine to use
--output, -o	None	Output file path

Examples:

# Basic inference
ethllama run model.gguf --prompt "Write a poem about AI"

# With GPU acceleration
ethllama run model.gguf --gpu cuda --gpu-layers 35 --threads 8

# Using a custom engine
ethllama run model.safetensors --engine my-custom-engine

# Save output to file
ethllama run model.gguf --prompt "Translate to French: Hello" --output result.txt

serve

ethllama serve [options]

Options:

Option	Default	Description
--host	127.0.0.1	Bind address
--port	8080	Port number
--api-key	""	API key for authentication

Example:

ethllama serve --host 0.0.0.0 --port 8080 --api-key "my-secret-key"

Configuration

Configuration is stored in ~/.ethllama/config.yaml:

gpu:
  backend: vulkan
  fallback: true
api:
  enabled: false
  host: 127.0.0.1
  port: 8080
  api_key: ""
telemetry:
  enabled: false
model_dirs:
  - /home/user/models

Run the interactive setup:

ethllama config --init

Architecture

┌─────────────────────────────────────────────┐
│              ethllama (Python CLI)           │
│  ┌─────────┐ ┌──────────┐ ┌──────────────┐ │
│  │ Config  │ │  Index   │ │  Engines     │ │
│  │ Manager │ │  Manager │ │  (YAML defs) │ │
│  └────┬────┘ └────┬─────┘ └──────┬───────┘ │
└───────┼───────────┼──────────────┼─────────┘
        │           │              │
┌───────┴───────────┴──────────────┴──────────┐
│           ethllama-core (Rust/PyO3)          │
│  ┌─────────────┐  ┌──────────────────┐      │
│  │ Model Loader │  │  Inference Engine │     │
│  └──────┬──────┘  └────────┬─────────┘      │
└─────────┼──────────────────┼────────────────┘
          │                  │
┌─────────┴──────────────────┴────────────────┐
│  External Backends (llama.cpp, whisper.cpp)  │
│  ┌──────────┐ ┌──────────┐ ┌──────────┐    │
│  │  Vulkan  │ │   ROCm   │ │   CUDA   │    │
│  └──────────┘ └──────────┘ └──────────┘    │
└─────────────────────────────────────────────┘

Components

• ethllama (Python): CLI interface, configuration management, model indexing, and engine orchestration using Jinja2-templated YAML engine definitions.
• ethllama-core (Rust): Native model loading and inference via PyO3 bindings to llama.cpp.
• Engines (YAML): Pluggable engine configurations that define how to invoke external inference binaries (llama.cpp, whisper.cpp, or custom).

Engine Configuration

Engines are defined as YAML files placed in ~/.ethllama/engines/. See docs/examples/ for ready-to-use templates.

Each engine config specifies:

• The binary to invoke
• A Jinja2 args template for building CLI commands
• Environment variables
• A pre-check command for validation
• Streaming support flag
• Supported model file extensions

Development

Setup

# Clone and enter the project
git clone https://github.com/luluthehungrycat/ethicallama.git
cd ethicallama

# Initialize submodules (for llama.cpp dependency)
git submodule update --init --recursive

# Build the Rust core
cargo build --release -p ethllama-core

# Install the Python package in editable mode
pip install -e ".[dev]"

Project Structure

ethicallama/
├── ethllama/                 # Python package
│   ├── __init__.py           # Package exports
│   ├── config.py             # Configuration management
│   ├── engines.py            # Engine config loading/running
│   └── index.py              # Model index management
├── ethllama-core/            # Rust core (PyO3 bindings)
│   ├── src/
│   │   ├── lib.rs            # PyO3 module definition
│   │   ├── llama.rs          # llama.cpp FFI bindings
│   │   └── utils.rs          # Utilities
│   ├── build.rs              # Build script (links llama.cpp)
│   └── Cargo.toml
├── docs/                     # Documentation
│   ├── USAGE.md              # Detailed usage guide
│   ├── PRIVACY.md            # Privacy policy
│   └── examples/             # Example engine configs
├── scripts/                  # Helper scripts
│   ├── setup.sh              # Development setup
│   └── benchmark.sh          # Simple benchmark
├── CREDITS.md                # Open-source credits
├── LICENSE                   # MIT License
└── README.md                 # This file

Testing

# Run Python tests
pytest

# Run Rust tests
cargo test -p ethllama-core

Using uv? The entire project works seamlessly with uv — see AGENTS.md for the full uv development workflow (venv creation, maturin builds, and the recommended test commands).

Credits

ethicallama is built on the shoulders of several excellent open-source projects. See CREDITS.md for the full list.

Key dependencies:

• llama.cpp - GGUF model loading and inference
• whisper.cpp - Speech-to-text (planned)
• PyO3 - Rust-Python bindings
• FastAPI - HTTP API server

License

MIT License. See LICENSE for details.