This is the implementation of zkGPT, which is a SNARK for LLM inference. This project partially references code from the work of zkCNN and Lasso. Current implementation only supports GPT-2.

• C++14
• cmake >= 3.10
• GMP library

To ensure smooth execution, we recommend using a server with the following specifications:

• Operating System (OS): Linux (e.g., Ubuntu 18.04 or later)
• Processor (CPU): Multi-core CPU, preferably with 16 cores or more
• Memory (RAM): at least 200GB

To run the code, make sure you clone with

git clone --recurse-submodules [email protected]:security-Anonymous/zkGPT.git

since the mcl library is included as a submodule.

sudo apt install libgmp-dev

The script to run LLM inference proving.

./llm.sh

The circuit initiation phase may take minutes because brute-forcely computing plaintext matrix multiplication using CPU single thread. We will optimize this part in future. Note this part will not become bottleneck in real world, due to it can be computed extremely efficiently on GPU.

• zkCNN: Zero knowledge proofs for convolutional neural network predictions and accuracy.

• Unlocking the lookup singularity with Lasso

• mcl