QTensorAI

A hybrid quantum-classical neural network simulation platform. Quantum simulation uses QTensor, a state-of-the-art tensor network-based simulator that usually has linear complexity in the number of qubits for shallow circuits, instead of exponential complexity. This opens up the possibility to simulate large hybrid models with many qubits. The hybrid model is a PyTorch model, batch-parallelized, GPU compatible and fully differentiable.

We demonstrate examples of how to use our code in \examples. A Jupyter notebook Example.ipynb shows the CPU usage. The script cuda.py shows the GPU usage. The script cuda_graph.py shows how to most efficiently use the GPU resource by tracing the graph, achieving a 35% speed up. When reading these files in the example folder, the comments are cut down to those that are relevant to your usage of our library without extraneous details.

This librar is based on QTensor at https://github.com/danlkv/QTensor.git, which is built on qtree at https://github.com/Huawei-HiQ/qtree.git. However, these libraries have heavy dependencies, and we restructured the code base and removed unnecessary components to create a minimal implementation that is functional.

Installation

Creat your pytorch environment with conda first:

conda create --name qtensor_ai pytorch cudatoolkit=11.3 -c pytorch

If you wish to run qtensor_ai on cpu only, then you do not have to include cudatoolkit in the installation command. Make sure you choose CUDA version 11 and above if you wish to use CUDAGraph acceleration. Otherwise, CUDA 10 is tested to work as well. Clone the repository QTensorAI:

git clone https://github.com/sss441803/QTensorAI.git

Install the qtensor_ai library:

cd QTensorAI
python setup.py install

Then, install the Tamaki optimizer. This is optional but recommended. If you do not install it, you will need to remove it's import from the script. More details for installation can be found at https://github.com/danlkv/QTensor.git.

> cd qtensor_ai/qtensor/qtree/thirdparty/tamaki_treewidth
> make heuristic 
javac tw/heuristic/*.java

If javac is not already available, install openjdk with conda:

conda install -c anaconda openjdk

Finally, add the directory QTensorAI/qtensor_ai/qtensor/qtree/thirdparty/tamaki_treewidth to $PATH.