github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-09-11 05:49:55 +00:00
Code Issues Packages Projects Releases Wiki Activity

upgrade ppo dpo rm script

Browse Source
This commit is contained in:
YeAnbang
2024-05-28 03:04:39 +00:00
parent 7a7e86987d
commit 929e1e3da4
15 changed files with 169 additions and 139 deletions
Download Patch File Download Diff File Expand all files Collapse all files

38
applications/ColossalChat/examples/training_scripts/train_sft.sh
View File

@@ -15,24 +15,24 @@ set_n_least_used_CUDA_VISIBLE_DEVICES() {
# export CUDA_VISIBLE_DEVICES=4,5,6
set_n_least_used_CUDA_VISIBLE_DEVICES 4
set_n_least_used_CUDA_VISIBLE_DEVICES 2
PROJECT_NAME="sft"
PARENT_SAVE_DIR="/home/yeanbang/data/experiment/output/model" # Path to a folder to save checkpoints
PARENT_TENSORBOARD_DIR="/home/yeanbang/data/experiment/logs/tensorboard" # Path to a folder to save logs
PARENT_CONFIG_FILE="/home/yeanbang/data/experiment/logs/config" # Path to a folder to save training config logs
PRETRAINED_MODEL_PATH="/mnt/jfs-hdd/share/models/Llama-2-7b-chat-hf" # huggingface or local model path
PRETRAINED_TOKENIZER_PATH="/mnt/jfs-hdd/share/models/Llama-2-7b-chat-hf" # huggingface or local tokenizer path
PRETRAINED_MODEL_PATH="/mnt/jfs-hdd/share/models/Yi-1.5-6B" # huggingface or local model path
PRETRAINED_TOKENIZER_PATH="/mnt/jfs-hdd/share/models/Yi-1.5-6B" # huggingface or local tokenizer path
declare -a dataset=(
/home/yeanbang/data/experiment/dataset/alpaca/test/Llama-2-7b-chat-hf/arrow/part-00000
/home/yeanbang/data/experiment/dataset/alpaca/test/Llama-2-7b-chat-hf/arrow/part-00001
/home/yeanbang/data/experiment/dataset/alpaca/test/Llama-2-7b-chat-hf/arrow/part-00002
/home/yeanbang/data/experiment/dataset/alpaca/test/Llama-2-7b-chat-hf/arrow/part-00003
/home/yeanbang/data/experiment/dataset/alpaca/test/Llama-2-7b-chat-hf/arrow/part-00004
/home/yeanbang/data/experiment/dataset/alpaca/test/Llama-2-7b-chat-hf/arrow/part-00005
/home/yeanbang/data/experiment/dataset/alpaca/test/Llama-2-7b-chat-hf/arrow/part-00006
/home/yeanbang/data/experiment/dataset/alpaca/test/Llama-2-7b-chat-hf/arrow/part-00007
/home/yeanbang/data/experiment/dataset/alpaca/test/Llama-2-7b-chat-hf/arrow/part-00008
/home/yeanbang/data/experiment/dataset/alpaca/test/Llama-2-7b-chat-hf/arrow/part-00009
/home/yeanbang/data/experiment/dataset/alpaca/test/Yi-1.5-6B/arrow/part-00000
/home/yeanbang/data/experiment/dataset/alpaca/test/Yi-1.5-6B/arrow/part-00001
/home/yeanbang/data/experiment/dataset/alpaca/test/Yi-1.5-6B/arrow/part-00002
/home/yeanbang/data/experiment/dataset/alpaca/test/Yi-1.5-6B/arrow/part-00003
/home/yeanbang/data/experiment/dataset/alpaca/test/Yi-1.5-6B/arrow/part-00004
/home/yeanbang/data/experiment/dataset/alpaca/test/Yi-1.5-6B/arrow/part-00005
/home/yeanbang/data/experiment/dataset/alpaca/test/Yi-1.5-6B/arrow/part-00006
/home/yeanbang/data/experiment/dataset/alpaca/test/Yi-1.5-6B/arrow/part-00007
/home/yeanbang/data/experiment/dataset/alpaca/test/Yi-1.5-6B/arrow/part-00008
/home/yeanbang/data/experiment/dataset/alpaca/test/Yi-1.5-6B/arrow/part-00009
)
TIMESTAMP=$(date +%Y-%m-%d-%H-%M-%S)
@@ -43,7 +43,7 @@ CONFIG_FILE="${PARENT_CONFIG_FILE}-${FULL_PROJECT_NAME}.json"
echo $(which colossalai)
echo $(which python)
# the real batch size for gradient descent is number_of_node_in_hostfile * nproc_per_node * train_batch_size
colossalai run --nproc_per_node 4 --master_port 31312 --hostfile ./hostfile train_sft.py \
colossalai run --nproc_per_node 2 --master_port 31312 --hostfile ./hostfile train_sft.py \
--pretrain $PRETRAINED_MODEL_PATH \
--tokenizer_dir $PRETRAINED_TOKENIZER_PATH \
--save_interval 4000 \
@@ -51,13 +51,13 @@ colossalai run --nproc_per_node 4 --master_port 31312 --hostfile ./hostfile trai
--save_path $SAVE_DIR \
--config_file $CONFIG_FILE \
--lora_rank 0 \
--plugin zero2 \
--tp 1 \
--plugin 3d \
--tp 2 \
--pp 1 \
--zero_stage 2 \
--batch_size 4 \
--zero_stage 0 \
--batch_size 2 \
--max_epochs 3 \
--accumulation_steps 4 \
--accumulation_steps 1 \
--lr 5e-5 \
--max_len 400 \
--grad_checkpoint \