PipeLLM

This repository is a prototype of paper PipeLLM: Fast and Confidential Large Language Model Services with Speculative Pipelined Encryption (https://arxiv.org/abs/2411.03357).

Basic Steps

Before compiling PipeLLM, the hard-code filepaths and parameters in the source files should be modified.

Compile PipeLLM:

make -j $(nproc)

This would generate libcrypto.so.3.

To use PipeLLM, add this file to LD_PRELOAD:

export LD_PRELOAD=$(pwd)/libcrypto.so.3
export LD_LIBRARY_PATH=$(pwd)

Before running the first program, PipeLLM should get the pattern of the IV with:

./profile_iv.sh

Use PipeLLM in CUDA Applications

Apart from the basic steps, the following should be done:

To use PipeLLM with applications compiled with nvcc, the applications should be recompiled adding --cudart shared flag (see profile.sh as an example).

Use PipeLLM in Pytorch

Apart from the basic steps, the following should be done:

First, copy $(pwd)/libcudart_static.a to /usr/local/cuda/lib64/ (please backup the original file first).

Next, compile pytorch from source.

Use PipeLLM in conda environment

Apart from the basic steps, the following should be done:

• Move the libcrypto.so.3 in the lib directory of the conda venv (possibly in anaconda3/envs/vllm/lib) to libcrypto.so.4
• Move the libcrypto.so.3 in this repo to the lib directory of the conda venv.