First use apps_create_split.py
to create the train.json
and test.json
. Note the paths specified in apps_create_split.py
should point to relative paths from training directory or absolute paths.
We use the following command to run and train. Note the configuration file is called deepspeed_config.json.
USE_TF=NO deepspeed tune_apps_gpt.py \
--save-dir=/path/to/save_dir \
--load=/path/to/model \ # Can be used to restart from checkpoint
--apps-train-files ~/apps/train \
--apps-dataroot ~/apps/train/ \
--grad-acc-steps=8 \
--epochs=10 \
--fp16 \
--deepspeed deepspeed_config.json \
--batch-size-per-replica=2