Training Benchmark

Running benchmark on CUDA GPU

Run benchmark, e.g. assuming you have n NVIDIA GPUs:

python training_benchmark_cuda.py --dataset ogbn-products --model edge_cnn --num-epochs 3 --n_gpus <n>

Running benchmark on Intel GPU

Environment setup

Prerequisites

Intel Data Center GPU Max Series. You could try it through Intel DevCloud.
Verify the Intel GPU Driver is installed, refer to the guide.

docker setup

If you want to run your scripts inside a docker image, you could refer to the dockerfile and the corresponding guide.

bare-metal setup

If you prefer to run your scripts directly on the bare-metal server. We recommend the installation guidance provided by Intel® Extension for PyTorch. The following are some key steps:

Install Intel® oneAPI Base Toolkit, indluding Intel® oneAPI DPC++ Compiler, Intel® oneAPI Math Kernel Library (oneMKL), Intel® oneAPI Collective Communications Library (oneCCL), and Intel® oneCCL Bindings for PyTorch.

bash

# Install oneCCL package on Ubuntu
sudo apt install -y intel-oneapi-dpcpp-cpp-2024.1=2024.1.0-963 intel-oneapi-mkl-devel=2024.1.0-691 intel-oneapi-ccl-devel=2021.12.0-309
# Install oneccl_bindings_for_pytorch
pip install oneccl_bind_pt==2.1.300+xpu --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/
# Runtime Dynamic Linking
source /opt/intel/oneapi/setvars.sh

Install Intel® Extension for PyTorch and the corresponding version of PyTorch

bash

pip install torch==2.1.0.post2 intel-extension-for-pytorch==2.1.30+xpu --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/

Running benchmark

This guide is helpful for you to launch DDP training on intel GPU.

To Run benchmark, e.g. assuming you have n XPUs:

mpirun -np <n> python training_benchmark_xpu.py --dataset ogbn-products --model edge_cnn --num-epochs 3