qkan 0.2.1

Quantum-inspired Kolmogorov Arnold Networks

QKAN: Quantum-inspired Kolmogorov-Arnold Network

"Quantum Variational Activation Functions Empower Kolmogorov-Arnold Networks"

Jiun-Cheng Jiang¹ Morris Yu-Chao Huang² Tianlong Chen² Hsi-Sheng Goan¹

¹National Taiwan University  ²UNC Chapel Hill 

This is the official repository for the paper: "Quantum Variational Activation Functions Empower Kolmogorov-Arnold Networks"

📖 Documentation: https://qkan.jimq.cc/

We provide a PyTorch implementation of QKAN with:

• Pre- and post-activation processing support
• Grouped QVAFs for efficient training
• Plot the nodes and pruning unnecessary nodes
• Layer extension for more complex features
• and more ...

A basic PennyLane version of the quantum circuit is also included for demonstration, but not optimized for performance.

2026-03: Released v0.2.0 with a more efficient quantum circuit implementation—using cuQuantum for the cutn solver and Triton for the flash solver—which significantly speeds up the activation function.

Installation

You can install QKAN using pip:

pip install qkan

If you want to install the latest development version, you can use:

pip install git+https://github.com/Jim137/qkan.git

To use the GPU-optimized solvers (including flash and cutn solver), you can install with the gpu extra:

pip install qkan[gpu]

Quick Start

Here's a minimal working example for function fitting using QKAN:

import torch

from qkan import QKAN, create_dataset

device = "cuda" if torch.cuda.is_available() else "cpu"

f = lambda x: torch.sin(20*x)/x/20 # J_0(20x)
dataset = create_dataset(f, n_var=1, ranges=[0,1], device=device, train_num=1000, test_num=1000, seed=0)

qkan = QKAN(
    [1, 1], 
    reps=3, 
    device=device, 
    seed=0,
    preact_trainable=True, 
    postact_weight_trainable=True,
    postact_bias_trainable=True, 
    ba_trainable=True,
    save_act=True, # enable to plot from saved activation
)

optimizer = torch.optim.LBFGS(qkan.parameters(), lr=5e-2)

qkan.train_(
    dataset,
    steps=100,
    optimizer=optimizer,
    reg_metric="edge_forward_dr_n",
)

qkan.plot(from_acts=True, metric=None)

You can find more examples in the examples for different tasks, such as function fitting, classification, and generative modeling.

Solver Guiding

Case	Device	Recommended solver	Why	Notes
Small models, CPU runs, debugging, or you want a trusted baseline	CPU (or GPU)	exact (default)	Simple + “reference” behavior	First run may include one-time init overhead—do a warmup step before timing.
Most training workloads (medium → large models) / inference	GPU	flash	Best overall speed / memory tradeoff in these benchmarks	Good first choice for practical GPU training.
BF16/FP8 mixed-precision training for maximum throughput	GPU	cutile	cuTile fused kernels with BF16/FP8 + coalesced state layout	Best with real ansatz.
Extremely large / memory-bound runs (near OOM, very large layers/batches)	GPU (or CPU)	cutn	Best scaling and peak-memory reduction in the extreme benchmark	Use when size/memory dominates. Or CPU case better than exact.

Ansatz choice (pz vs real)

• Default: pz — most reliable quality across tasks.
• real can be faster/smaller, but may hurt accuracy/convergence on some workloads—only use if you validate it on your task.

See #8 for more discussion on solver choices and tradeoffs.

Mixed Precision

The flash and cutile solvers support BF16 and FP8 mixed-precision via the c_dtype parameter:

qkan = QKAN([10, 10], solver="flash", c_dtype=torch.bfloat16, device="cuda")

• c_dtype controls the compute dtype for quantum simulation kernels (state vectors, trig ops).
• p_dtype controls the parameter storage dtype (theta, preacts). Keep this at float32.
• BF16 is the sweet spot: 2.3-2.5x faster training, 45% less peak memory, with identical convergence.
• FP8 (torch.float8_e4m3fn) provides additional memory savings for state checkpoints via prescaled storage.
• All ansatzes (pz, rpz, real) are supported.

See #12 for full benchmarks (GPT-2 HQKANsformer, isolated kernel timings, and dtype performance matrix).

Contributing

We are very welcome to all kinds of contributions, including but not limited to bug reports, documentation improvements, and code contributions.

To start contributing, please fork the repository and create a new branch for your feature or bug fix. Then, submit a pull request with a clear description of your changes.

In your environment, you can install the development dependencies with:

# clone your forked repository and navigate to the project directory
# for example `git clone https://github.com/Jim137/qkan.git && cd qkan`

pip install -e .[dev] # install development dependencies
pip install -e .[doc] # install documentation dependencies
pip install -e .[all] # install all optional dependencies

Citation

@article{jiang2025qkan,
  title={Quantum Variational Activation Functions Empower Kolmogorov-Arnold Networks},
  author={Jiang, Jiun-Cheng and Huang, Morris Yu-Chao and Chen, Tianlong and Goan, Hsi-Sheng},
  journal={arXiv preprint arXiv:2509.14026},
  year={2025},
  url={https://arxiv.org/abs/2509.14026}
}
@misc{jiang2025qkan_software,
  title={QKAN: Quantum-inspired Kolmogorov-Arnold Network},
  author={Jiang, Jiun-Cheng},
  year={2025},
  publisher={Zenodo},
  doi={10.5281/zenodo.17437425},
  url={https://doi.org/10.5281/zenodo.17437425}
}