transformers-benchmark-23.07.def

Bootstrap: docker
From: nvcr.io/nvidia/pytorch:23.07-py3

%files
    $PWD/run_clm_bench.sh /opt/run_clm_bench.sh
    $PWD/requirements-23.07.txt requirements.txt

%post
    # Install other python deps into the venv
    python3 -m pip install --pre -r requirements.txt
    # Clone transformers for the run_clm example
    git clone --depth 1 --branch v4.37.0 https://github.com/huggingface/transformers.git /opt/transformers/

%test
    # Ensure expected directories and files have been copied into the image
    if [ ! -d /opt/transformers ]; then
        exit 1
    fi
    if [ ! -f /opt/run_clm_bench.sh ]; then
        exit 1
    fi
    # Ensure torch imports, and print the cuda/cuda arch list versions for visual checking
    python3 -c "import torch;print(torch.cuda.is_available()); print(torch.cuda.get_arch_list())"
    # Print the pytorch version
    python3 -c "import torch;print(torch.__version__)"
    # Ensure that the pytorch was built with the c++ abi 
    python3 -c "import torch;print(torch.compiled_with_cxx11_abi())"
    # Ensure that transformers.pipelines can be imported successfully. If not, there's likely a conflict between the NGC built packages and pip installed packages (or host python packages mixing into the environment)
    python3 -c "import transformers.pipelines"

# Define the default run script, to run the benchmark script as-is
%runscript
    /opt/run_clm_bench.sh /opt/transformers/examples/pytorch/language-modeling/run_clm.py