TICO

TICO (Torch IR to Circle ONE) is a Python library that converts PyTorch modules into Circle models — a lightweight and efficient representation in ONE designed for optimized on-device neural network inference.

Highlights

• One-call conversion — tico.convert() turns an nn.Module into a ready-to-deploy .circle binary.
• .pt2 support — convert saved torch.export programs via the Python API or the pt2-to-circle command-line tool.
• Run Circle models in Python — execute converted models directly for quick parity checks against PyTorch.
• Quantization toolkit — a unified prepare/convert API with GPTQ, PTQ (WrapQ), SmoothQuant, SpinQuant, and CLE, plus config-driven CLI recipes for LLMs and VLMs.

Installation

Prerequisites

• Python 3.10+
• (Optional) one-compiler — only required to run inference with converted Circle models. Conversion itself does not need it.

We highly recommend using a virtual environment (e.g., conda, venv).

From PyPI

pip install tico

From source

git clone https://github.com/Samsung/TICO.git
cd TICO

./ccex build     # generates build/ and dist/
./ccex install   # installs the package

./ccex install options

Option	Description
--dist	Install from the built wheel (default is editable mode)
--torch_ver <ver>	Torch version to install: a family (2.5 ~ 2.10), an exact version (e.g. 2.7.0+cu118), or nightly. Default: 2.7
--cuda_ver <maj.min>	Override the detected CUDA version (e.g. 12.1)
--cpu_only	Force a CPU-only Torch installation

Quick start

import tico
import torch

class AddModule(torch.nn.Module):
    def forward(self, x, y):
        return x + y

torch_module = AddModule()
example_inputs = (torch.ones(4), torch.ones(4))

circle_model = tico.convert(torch_module.eval(), example_inputs)
circle_model.save('add.circle')

Note

Call eval() on the module before conversion. TICO internally uses torch.export, so the module must be export-able.

Converting a saved .pt2 file from the command line:

pt2-to-circle -i add.pt2 -o add.circle

See the Getting Started guide for compile configurations, .pt2 conversion, and running Circle models directly in Python.

Quantization

The tico.quantization module provides a unified, modular interface for quantizing neural networks — including large language models — through a simple two-step prepare → convert workflow:

from tico.quantization import prepare, convert
from tico.quantization.config.gptq import GPTQConfig

prepared_model = prepare(model.eval(), GPTQConfig())

for d in dataset:          # calibration
    prepared_model(d)

quantized_model = convert(prepared_model, GPTQConfig())

• Quantization overview — API, architecture, and how to add a new algorithm
• Quantization algorithms — GPTQ, SmoothQuant, SpinQuant, CLE, …
• Config-driven CLI examples — quantize, evaluate, and inspect LLM/VLM recipes from the command line

Documentation

For users

Document	Description
Getting Started	Converting modules and .pt2 files, compile configuration, running Circle models in Python
Quantization	The prepare/convert quantization API and toolkit
Quantization examples	Command-line quantization, evaluation, and debugging workflows

For developers

Document	Description
Development guide	Environment setup, testing, and code formatting with ./ccex
System design	Architecture, pass pipeline, invariants, and behavior design
Requirements	Functional and non-functional requirements
System tests	System-level test coverage

Contributing

Contributions are welcome! For quantization algorithms, start with the quantization contribution guide and the recipes developer guide. Before submitting a PR, set up the development environment and run the tests and formatter as described in the development guide.

License

Licensed under the Apache License 2.0 — see LICENSE.