/pytorch-dataloader-segfault-investigation

Environments to help reproduct DataLoader segfault issue

Primary LanguagePython

This repo was created to help me attempt to reproduce an issue in PyTorch: link to issue

Prerequisite: Conda

Install Miniconda3: link

Prerequisite: CUDA toolkits

The different environments in this repo require different CUDA versions. I used these instructions to install multiple versions CUDA versions on my system: link

When creating an environment, if the required CUDA version is not found, an error is thrown.

Creating environments

Each of the environments in this repo have their own directory. They all contain a create-conda-env script which creates the conda environment, sets up variables to point to the proper CUDA version, and installs the corresponding PyTorch version.

To create, for instance, env0, run the following:

$ ./env0/create-conda-env

This will create the conda environment pytorch-dataloader-env0. To activate it, run:

$ conda activate pytorch-dataloader-env0

Running the DataLoader segfault test

To run the DataLoader test, make sure one of the environments has been activated and then run:

$ python test.py