ml_scripts

This is a collection of scripts including cuda installation, nightly pytorch installation, torchbench tests, and more.

torchbench scripts

For scripts that source run_base.sh, the run command is like

work_path=/home/yhao/d/p8 env1=pt_jan02 env2=pt_oct26_cu117 cuda_env1=/home/yhao/setenvs/set11.6.sh mode=train  ./run_all_speedup_cuda.sh

The detailed explanation of the arguments is in run_base.sh.

cuda installation

t_cuda=118 t_cuda_path=/install/path  ./install_cuda.sh

It will automatically download and install cuda 11.8 and cudnn 8.5.0 to the path /install/path/cuda-11.8.

nightly pytorch installation

down_nightly_torch.py --date 20230102

This script will automatically download and analyze the package dependencies of the nightly pytorch build from the given date. The date format is YYYYMMDD. e.g., torchvision built in 20230102 may rely on torch 20230101 while torchtext 20230102 may rely on torch 20230102. All packages will be downloaded to .downloads/cu116. If no conflict dependcy found, you can use the following cmomand to install the nightly torch.

cd .downloads/ && mkdir 20230102 && cd 20230102
for file in `ls ../cu116/*20230102*`; do 
ln -s $file
done
ln -s ../cu116/pytorch-triton-XXX

pip install ./*