From cc7d18eeed18f10a6475617aad47e51202dc75fd Mon Sep 17 00:00:00 2001
From: iMountTai <2506700016@qq.com>
Date: Sun, 25 Jun 2023 23:24:16 +0800
Subject: [PATCH 1/2] fix training step

---
 scripts/training/run_pt.sh  | 7 +++----
 scripts/training/run_sft.sh | 9 ++++-----
 2 files changed, 7 insertions(+), 9 deletions(-)

diff --git a/scripts/training/run_pt.sh b/scripts/training/run_pt.sh
index bfd986e..a2fef7e 100644
--- a/scripts/training/run_pt.sh
+++ b/scripts/training/run_pt.sh
@@ -11,8 +11,7 @@ dataset_dir=path/to/pt/data/dir
 data_cache=temp_data_cache_dir
 per_device_train_batch_size=1
 per_device_eval_batch_size=1
-training_steps=100
-gradient_accumulation_steps=1
+gradient_accumulation_steps=8
 output_dir=output_dir
 
 deepspeed_config_file=ds_zero2_no_offload.json
@@ -29,7 +28,7 @@ torchrun --nnodes 1 --nproc_per_node 1 run_clm_pt_with_peft.py \
     --do_train \
     --seed $RANDOM \
     --fp16 \
-    --max_steps ${training_steps} \
+    --num_train_epochs 1 \
     --lr_scheduler_type cosine \
     --learning_rate ${lr} \
     --warmup_ratio 0.05 \
@@ -38,7 +37,7 @@ torchrun --nnodes 1 --nproc_per_node 1 run_clm_pt_with_peft.py \
     --logging_steps 10 \
     --save_strategy steps \
     --save_total_limit 3 \
-    --save_steps 500 \
+    --save_steps 200 \
     --gradient_accumulation_steps ${gradient_accumulation_steps} \
     --preprocessing_num_workers 8 \
     --block_size 512 \
diff --git a/scripts/training/run_sft.sh b/scripts/training/run_sft.sh
index 13925b4..df27aef 100644
--- a/scripts/training/run_sft.sh
+++ b/scripts/training/run_sft.sh
@@ -10,8 +10,7 @@ chinese_tokenizer_path=path/to/chinese/llama/tokenizer/dir
 dataset_dir=path/to/sft/data/dir
 per_device_train_batch_size=1
 per_device_eval_batch_size=1
-training_steps=100
-gradient_accumulation_steps=1
+gradient_accumulation_steps=8
 output_dir=output_dir
 peft_model=path/to/peft/model/dir
 validation_file=validation_file_name
@@ -30,7 +29,7 @@ torchrun --nnodes 1 --nproc_per_node 1 run_clm_sft_with_peft.py \
     --do_eval \
     --seed $RANDOM \
     --fp16 \
-    --max_steps ${training_steps} \
+    --num_train_epochs 3 \
     --lr_scheduler_type cosine \
     --learning_rate ${lr} \
     --warmup_ratio 0.03 \
@@ -40,8 +39,8 @@ torchrun --nnodes 1 --nproc_per_node 1 run_clm_sft_with_peft.py \
     --save_strategy steps \
     --save_total_limit 3 \
     --evaluation_strategy steps \
-    --eval_steps 250 \
-    --save_steps 500 \
+    --eval_steps 100 \
+    --save_steps 200 \
     --gradient_accumulation_steps ${gradient_accumulation_steps} \
     --preprocessing_num_workers 8 \
     --max_seq_length 512 \

From c031387ffc1cf9ca23a0b3be4e91a593c7df7898 Mon Sep 17 00:00:00 2001
From: iMountTai <2506700016@qq.com>
Date: Mon, 26 Jun 2023 08:39:12 +0800
Subject: [PATCH 2/2] fix training steps

---
 scripts/training/run_sft.sh | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/scripts/training/run_sft.sh b/scripts/training/run_sft.sh
index df27aef..db31a6c 100644
--- a/scripts/training/run_sft.sh
+++ b/scripts/training/run_sft.sh
@@ -29,7 +29,7 @@ torchrun --nnodes 1 --nproc_per_node 1 run_clm_sft_with_peft.py \
     --do_eval \
     --seed $RANDOM \
     --fp16 \
-    --num_train_epochs 3 \
+    --num_train_epochs 1 \
     --lr_scheduler_type cosine \
     --learning_rate ${lr} \
     --warmup_ratio 0.03 \