GVM Intercept Layer for NVIDIA CUDA Driver

This is the source release of the GVM Intercept Layer for NVIDIA CUDA Driver, tested with CUDA Driver 12.9 and GPU Driver 575.57.08.

Requirements

1. gcc
2. nvcc
3. make
4. python3
5. python3-venv
6. patchelf

Setup build environment

python3 -m venv GVMCUDADriverBuildEnv
source GVMCUDADriverBuildEnv/bin/activate
pip3 install lief

How to build

Easiest way to build:

make

To specify output dir of build:

make BUILD=<path to dir>

To specify CUDA driver the intercept layer is attaching to:

make CUDA=<path to cuda driver>

How to install

Easiest way to install:

make install

To specify output dir of install:

make install INSTALL=<path to dir>

Note that is will first backup libcuda.so if exists in specified install dir, then remove all symlinks to the existing cuda driver in that driver if exists.

To specify CUDA driver the intercept layer is attaching to:

make install CUDA=<path to cuda driver>

How to use

For the destination dir you choose to install the intercept layer:

make install INSTALL=<path to dir>

You can work with any CUDA programs using:

LD_LIBRARY_PATH=<path to dir>:$LD_LIBRARY_PATH <cuda programms>

For example:

LD_LIBRARY_PATH=<path to dir>:$LD_LIBRARY_PATH vllm serve meta-llama/Llama-3.2-3B