CrypTorch is an MPC-based ML framework built on top of PyTorch, developed and maintained by the PAWS lab @Penn State. CrypTorch enables you to run your PyTorch-written ML program using multi-party computing (MPC), allowing model training and inference while ensuring the confidentiality of input data and model weights. 

• 🚀 Efficient: Due to its better engineering, CrypTorch is 11.1x faster for ResNet18 inference and 6.4x faster for BERT-Base inference compared to the popular library, CrypTen. If we compare individual operators, CrypTorch is 14.4-96.2x faster for Conv/Linear layers, 5.4x faster for ReLU, and 4.45-7.74x faster for Max/MaxPool.

• 🧑‍💻 Pure Python- and PyTorch-based: CrypTorch is implemented entirely in Python (except for its GPU compute kernels), which makes it easier to play with even if you are not a C++ expert. CrypTorch is built on top of PyTorch 2's compiler infrastructure (export IR) and does not introduce a custom frontend/IR.

• 🧩 Designed to be hacked, extended, and debugged: CrypTorch is a modular, multi-stage compiler, making it easy to experiment with new ideas or swap out components in isolation. Each compilation stage produces an executable graph, so you can incrementally run and test out intermediate graphs—dramatically simplifying debugging and research prototyping.

• ⚙️ Auto-tuning: CrypTorch provides operator approximation auto-tuning, enabling additional performance by trading off accuracy.

You can find the description of CrypTorch in our arxiv paper here: https://arxiv.org/pdf/2511.19711

To be updated