Skip to content

endurehero/Tensile_multiprocess_train

Repository files navigation

Tensile_multiprocess_train

Description: A tool for tensile training(multicard support) based on python3.

Usages:
/bin/train is the top-level script. Below command line could be used for launching...

cd ~/Tenslie_multiporcess_train #enter the root folder
scl enable devtoolset-7 bash
python ./bin/train config_path output_path tensile_path [gpu]
~ config_path : required, specify the input yaml file.
~ output_path : required, sepcify the output folder.
~ tensile_path : required, specify the root folder of tensile.
~ gpu : optional, specify the gpu index which will be used to trian, default is use all the gpu in that environment.

example:
python train ./rocblas_dgemm_nt_inc0.yaml ./output ~/Tensile/ -g 0 1 2 3 # use gpu 0 1 2 3 to train tensile

after training, you will find "output" folder in the workdir, and it's tree struct as below:

~ output
0_SplitConfigYaml #Split config file for every gpu
1_BenchmarkProblems #as folder name
2_BenchmarkData #as folder name
3_LibraryLogic #as folder name
4_LibraryClient #as folder name
5_SplitTrainingLog #Tensile log for every gpu

About

python tool for tensile training

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages