Qwodel 0.0.4

Production-grade model quantization SDK for enterprise custom models (AWQ, GGUF, and CoreML)

Qwodel - Production-Grade Model Quantization

Qwodel is a production-ready Python package for model quantization across multiple backends (AWQ, GGUF, CoreML). It provides a unified, intuitive API for quantizing large language models with minimal code.

Features

• Unified API - Simple interface across all quantization backends
• Multiple Backends - AWQ (GPU), GGUF (CPU), CoreML (Apple devices)
• Optional Dependencies - Install only what you need
• CLI & Python API - Use via command line or programmatically
• Type Safe - Full type hints and mypy validation
• Well Documented - Comprehensive docs with examples

Quick Start

Installation

Quick Install (All Backends)

pip install qwodel[all]

This installs all backends (GGUF, AWQ, CoreML) with PyTorch 2.1.2 (CPU version).

GPU Support (for AWQ only)

If you need GPU quantization with AWQ, install PyTorch with CUDA first:

# 1. Install PyTorch with CUDA 12.1
pip install torch==2.1.2 torchvision==0.16.2 torchaudio==2.1.2 --index-url https://download.pytorch.org/whl/cu121

# 2. Install qwodel
pip install qwodel[all]

Note: GGUF and CoreML work perfectly fine with CPU-only PyTorch!

Individual Backends

# GGUF only (CPU quantization - most popular!)
pip install qwodel[gguf]

# AWQ only (GPU quantization)
pip install qwodel[awq]

# CoreML only (Apple devices)
pip install qwodel[coreml]

Local Development

# Clone and install locally
cd /path/to/qwodel
pip install -e .[all]

Python API

from qwodel import Quantizer

# Create quantizer
quantizer = Quantizer(
    backend="gguf",
    model_path="meta-llama/Llama-2-7b-hf",
    output_dir="./quantized"
)

# Quantize model
output_path = quantizer.quantize(format="Q4_K_M")
print(f"Quantized model saved to: {output_path}")

CLI

# Quantize a model
qwodel quantize \
    --backend gguf \
    --format Q4_K_M \
    --model meta-llama/Llama-2-7b-hf \
    --output ./quantized

# List available formats
qwodel list-formats --backend gguf

Supported Backends

GGUF (CPU Quantization)

• Use Case: CPU inference, broad compatibility
• Formats: Q4_K_M, Q8_0, Q2_K, Q5_K_M, and more
• Best For: Most users, CPU-based deployment

AWQ (GPU Quantization)

• Use Case: NVIDIA GPU inference
• Formats: INT4
• Best For: GPU deployments, maximum speed
• Requires: CUDA 12.1+

CoreML (Apple Devices)

• Use Case: iOS, macOS, iPadOS deployment
• Formats: FLOAT16, INT8, INT4
• Best For: Apple device deployment

Examples

Batch Processing

from qwodel import quantize

models = ["meta-llama/Llama-2-7b-hf", "meta-llama/Llama-2-13b-hf"]

for model in models:
    quantize(
        model_path=model,
        backend="gguf",
        format="Q4_K_M",
        output_dir="./quantized"
    )

Custom Progress Callback

from qwodel import Quantizer

def progress_handler(progress: int, stage: str, message: str):
    print(f"[{progress}%] {stage}: {message}")

quantizer = Quantizer(
    backend="gguf",
    model_path="./my-model",
    output_dir="./output",
    progress_callback=progress_handler
)

quantizer.quantize(format="Q4_K_M")

Documentation

• API Reference
• CLI Reference
• API Reference
• CLI Reference
• Troubleshooting

Contributing

We welcome contributions! Please see CONTRIBUTING.md for guidelines.

License

MIT License - see LICENSE for details.

Acknowledgments

Qwodel builds upon the excellent work of:

• llama.cpp for GGUF quantization
• llm-compressor for AWQ quantization
• CoreMLTools for CoreML conversion