v0.1.0

TICO is an experimental Python library for converting PyTorch models to Circle models, designed for easy integration into on-device inference pipelines across a variety of platforms — including mobile CPUs, GPUs, and NPUs. It provides a programmatic interface for model export, graph-level optimization, and post-training quantization — all tailored for deployment on resource-constrained devices.

Version 0.1.0 marks the first public preview.

🚀 Highlights

• Single-call conversion: tico.convert() top-level helper returns a ready-to-deploy Circle binary.
• Post training Quantization: PT2E, SmoothQuant, GPTQ.
• Lightweight dependency stack: minimal setup, pure Python workflow.

Requirements

Component	Tested Version
Python	3.10 ~ 3.12
PyTorch	2.5 ~ 2.7

torch.export is under active development.

Installation

pip install tico

Quick Start

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')

Supported Models

The following models have been tested with TICO v0.1.0. Refer to test/modules/model for implementation details and configuration examples.

• Llama 3.2
• GPT2
• Florence2
• SmolVLM
• TinyLlama
• EfficientFormer
• EfficientNet
• InceptionV3
• Mamba
• MobileNetV2
• ResNet18
• DeepSeek-R1-Distill-Qwen