TICO

TICO (Torch IR to Circle ONE) is a python library for converting Pytorch modules into a circle model that is a lightweight and efficient representation in ONE designed for optimized on-device neural network inference.

Table of Contents

For Users

• Installation
• Getting Started
  • From torch module
  • From .pt2
  • Running circle models directly in Python

For Developers

• Testing & Code Formatting
• Testing
• Code Formatting

For Users

Installation

0. Prerequisites

• Python 3.10
• one-compiler nightly
  • This project depends on ONE Compiler, and it uses nightly features that are not yet available in the official release. Until one-compiler 1.30.0 is released, you must use a prebuilt nighlty version of ONE Compiler.

We highly recommend to use a virtual env, e.g., conda.

1. Clone this repo

2. Build python package

./ccex build

This will generate build and dist directories in the root directory.

3. Install generated package

./ccex install

Available options

• --dist To install the package from .whl (without this option, TICO is installed in an editable mode)
• --torch_ver <torch version> To install a specific torch version (default: 2.6).
  • Available : 2.5, 2.6, nightly

4. Now you can convert a torch module to a .circle.

Getting started

This tutorial explains how you can use TICO to generate a circle model from a torch module.

Let's assume we have a torch module.

import tico
import torch

class AddModule(torch.nn.Module):
    def __init__(self):
        super().__init__()

    def forward(self, x, y):
        return x + y

NOTE TICO internally uses torch.export. Therefore, the torch module must be 'export'able. Please see this document if you have any trouble to export.

From torch module

You can convert a torch module to a circle model with these steps.

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

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

Compile with configuration

from test.modules.op.add import AddWithCausalMaskFolded

torch_module = AddWithCausalMaskFolded()
example_inputs = torch_module.get_example_inputs()

config = tico.CompileConfigV1()
config.legalize_causal_mask_value = True
circle_model = tico.convert(torch_module, example_inputs, config = config)
circle_model.save('add_causal_mask_m120.circle')

With legalize_causal_mask_value option on, causal mask value is converted from -inf to -120, creating a more quantization-friendly circle model with the cost of slight accuracy drop.

From .pt2

The torch module can be exported and saved as .pt2 file (from PyTorch 2.1).

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

exported_program = torch.export.export(module, example_inputs)
torch.export.save(exported_program, 'add.pt2')

There are two ways to convert .pt2 file: python api, command line tool.

• Python API

circle_model = tico.convert_from_pt2('add.pt2')
circle_model.save('add.circle')

• Command Line Tool

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

• Command Line Tool with configuration

pt2-to-circle -i add.pt2 -o add.circle -c config.yaml

# config.yaml

version: '1.0' # You must specify the config version. 
legalize_causal_mask_value: True

Running circle models directly in Python

After circle export, you can run the model directly in Python.

Note that you should install one-compiler package first.

The output types are numpy.ndarray.

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

circle_model = tico.convert(torch_module, example_inputs)
circle_model(*example_inputs)
# numpy.ndarray([2., 2., 2., 2.], dtype=float32)

For Developers

Testing & Code Formatting

Run below commands to configure testing or formatting environment.

Refer to the dedicated section to have more fine-grained control.

$ ./ccex configure                          # to set up testing & formatting environment
$ ./ccex configure format                   # to set up only formatting environment
$ ./ccex configure test                     # to set up only testing environment

Available options

• --torch_ver <torch version> To install a specific torch family package(ex. torchvision) version (default: 2.6)
  • Available : '2.5', '2.6', 'nightly'

$ ./ccex configure                          # to set up testing & formatting environment with stable2.6.x version
$ ./ccex configure test                     # to set up only testing environment with stable 2.6.x version
$ ./ccex configure test --torch_ver 2.5     # to set up only testing environment with stable 2.5.x version
$ ./ccex configure test --torch_ver nightly     # to set up only testing environment with nightly version

Testing

Test congifure

Run below commands to install requirements for testing.

NOTE TICO will be installed in an editable mode.

./ccex configure test

# without editable install
./ccex configure test --dist

Test All

Run below commands to run the all unit tests.

NOTE Unit tests don't include model test.

./ccex test
# OR
./ccex test run-all-tests

Test Subset

To run subset of test.modules.*, Run ./ccex test -k <keyword>

For example, to run tests in specific sub-directory (op, net, ..)

# To run tests in specific sub-directory (op/, net/ ..)
./ccex test -k op
./ccex test -k net

# To run tests in one file (single/op/add, single/op/sub, ...)
./ccex test -k add
./ccex test -k sub

# To run SimpleAdd test in test/modules/single/op/add.py
./ccex test -k SimpleAdd

To see the full debug log, add -v or TICO_LOG=4.

TICO_LOG=4 ./ccex test -k add
# OR
./ccex test -v -k add

Test Model

If you want to test them locally, you can do so by navigating to each model directory, installing the dependencies listed in its requirements.txt, and running the tests one by one.

$ pip install -r test/modules/model/<model_name>/requirements.txt
# Run test for a single model
$ ./ccex test -m <model_name>

For example, to run a single model

./ccex test -m InceptionV3

Code Formatting

Format configure

Run below commands to install requirements for formatting.

./ccex configure format

Format run

./ccex format