Installation

Tutorial video

Requirements

We recommend using a conda environment manager.

System requirements:

• 3.8 <= Python <= 3.11
• 2.0 <= PyTorch <= 2.3
• CUDA (run the nvidia-smi command and ensure it shows your available GPUs)

Optional requirements

Graph visualization

If you wish to visualize the model architecture graph before and after the compression process, you can install the dependencies graphviz and pydot using:

conda install graphviz
pip install pydot

Once the dependencies are installed, two .svg files containing the visualizations of the original model and the quantized model will be generated in the outputs folder when running ACE. For more information, see graph visualization.

CLIKA Compression installation

Installing clika-ace is similar to installing any other Python package using the pip package manager; the only special requirement is to add the license key to gain access to CLIKA's PIP index.

You can install the package using the following lines:

# set your license key as a environment varialbe
export CC_LICENSE_KEY=YOUR_LICENSE_KEY

pip install --force "clika-ace" --extra-index-url \
https://license:$CC_LICENSE_KEY@license.clika.io/simple

Making the license key globally available

Since every usage of the clika-ace package requires the availability of a valid license key, we will provide clika-ace access to the license key within your system.

Execute the following command from any directory:

clika-init-license

and insert your license key when prompted.

This process will generate a file containing your license key in ~/.clika/.cc_license, which can be accessed by the clika-ace package, facilitating its usage whenever required.

Congratulations! You're ready to start using the clika-ace Python package.

---

note

(Optional) An alternative option to keep your license key globally available is to set it as a permanent environment variable instead of using the clika-init-license command.

tip

If you wish to add clika-ace to a requirements.txt file, you may add the following lines to it:

--extra-index-url https://license:$CC_LICENSE_KEY@license.clika.io/simple
clika-ace

Common issues

CUDA-related issues

In the case of CUDA-related issues we recommend performing the following steps:

Verify CUDA installation

You can validate the system CUDA installation using the command:

nvcc --version

The expected output of the command is the CUDA version installed on your system (we recommend the latest available CUDA version). If the command does not work as expected, please refer to the CUDA installation guide before using the SDK.

Verify PyTorch installation

Test if the GPU is accessible to PyTorch using the command:

python3 -c "import torch; print(torch.cuda.is_available())"

Expected output:

>>> True

If the output is False please refer to the PyTorch installation guide and CUDA installation guide before using the SDK.