Deep_Learning_Cuda

This project provides an singularity container setup for developing and experimenting with custom CUDA kernels for PyTorch neural networks.

Container Setup

To build the container in an Apptainer environment, run:

apptainer build DL_cuda.sif DL_cuda.def

Development

For interactive development with CUDA support:

apptainer exec --nv DL_cuda.sif bash

Inside the container, verify that CUDA is properly configured:

nvcc --version

build custom cuda kernels with pip

To build and install custom CUDA kernels:

cd cuda_kernels
pip install --user -e .

Build cuda kernals

You need to compile cuda kernals before using them

cd cuda_kernals
pip install --user -e .
cd ..

To Run the model with custom functions

Configure the yaml file with preferred settings.

python main.py

Experiments

To run performance experiments first give preffered model configurations in config.yaml.

Then run

python experiements/experiments.py

To do compute profiling experiments first give preffered data configurations in config.yaml.

Then run

python experiements/compute_experiments.py

Tests

Checkout to test branch.

Run

pytest

or (to avoid specific wardnings if packages are outdated.)

pytest -W ignore::DeprecationWarning