XBTorch

---

XBTorch is a PyTorch-native framework for simulating crossbar-based deep neural networks with emerging memory technologies such as ReRAM, FeFETs, PCM, and MTJs.

It enables researchers and engineers to:

• Model realistic device-level behavior (variability, noise, nonlinearity),
• Perform hardware-aware training with quantization and gradient decomposition,
• Evaluate fault-tolerant inference on simulated crossbar arrays,
• Seamlessly integrate with existing PyTorch models with minimal code changes.

👉 For detailed guides, please see the XBTorch Documentation.

---

Dependencies & Installation

The recommended installation method is to create a lightweight virtual environment and install XBTorch in editable mode:

$ python -m venv .env
$ source .env/bin/activate
(.env) $ pip install -e xbtorch

This will install XBTorch in editable mode, allowing you to modify the source code directly.

For more detailed instructions (including optional dependencies and troubleshooting), review our documentation.

---

Getting Started

Minimal code changes are needed to adapt PyTorch models for XBTorch:

import xbtorch
import xbtorch.optim as xboptim
from xbtorch.patches import xbtorch_model
import torch.nn as nn

# Define a simple 2-layer perceptron network
class SimpleMLP(nn.Module):
    def __init__(self, input_size, hidden_size, output_size):
        self.input_size = input_size
        super(SimpleMLP, self).__init__()
        self.model = nn.Sequential(
            nn.Linear(input_size, hidden_size, bias=False),
            nn.ReLU(),
            nn.Linear(hidden_size, output_size, bias=False),
        )

    def forward(self, x):
        x = x.view(-1, self.input_size)  # Flatten the image
        x = self.model(x)
        return x

# Initialize
xbtorch.initialize()

# Define your model
model = SimpleMLP(10, 5, 2)
model = xbtorch_model(model)   # patch with XBTorch

# Optimizer
optimizer = xboptim.SGD(model.parameters(), lr=0.01)
criterion = nn.CrossEntropyLoss()

# ... Implement your training loop as usual!

For full examples (e.g., hardware-aware training and inference, fault-tolerance, etc.), see the examples/ directory or the documentation.

---

Citation

If you use this library, please cite this repository according to the information in CITATION.cff and/or the introductory paper:

@inproceedings{yousuf2025xbtorch,
  author    = {Yousuf, Osama and Glasmann, Andreu L. and Lueker-Boden, Martin and Najmaei, Sina and Adam, Gina C.},
  title     = {XBTorch: A Unified Framework for Modeling and Co-Design of Crossbar-Based Deep Learning Accelerators},
  booktitle = {arXiv},
  year      = {2026},
  url       = {https://arxiv.org/abs/2601.07086}
}

---

Acknowledgements

This library was developed as a collaboration between:

• The George Washington University (GWU)
• DEVCOM Army Research Laboratory (ARL)
• Western Digital Research

The project is licensed under the BSD-3 license (see LICENSE).

---

Contact and Collaboration

Research groups interested in collaborating are encouraged to reach out:

Osama Yousuf
Osama.Yousuf1@wdc.com
Western Digital Research
R&D Engineering, Memory Technology

Prof. Gina Adam
GinaAdam@gwu.edu
Adaptive Devices and Microsystems Group
Department of Electrical and Computer Engineering
George Washington University

Andreu L. Glasmann
Andreu.L.Glasmann.Civ@army.mil
DEVCOM Army Research Lab

---

License

BSD-3 License. See the LICENSE file for details.