-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathiter_pipeline.sh
158 lines (141 loc) · 5.56 KB
/
iter_pipeline.sh
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
SCRIPT_ABS_PATH=$(readlink -f "$0")
SCRIPT_DIR=$(dirname "$SCRIPT_ABS_PATH")
# todo
LOG_FILE="$SCRIPT_DIR/logs/YOUR_TIME_LOG_PATH.log"
ERROR_LOG_FILE="$SCRIPT_DIR/logs/YOUR_LOG_PATH.log"
bm_data_format_dir="$SCRIPT_DIR/data/bm/YOUR_PATH_TO_bm_data_format_dir"
Finetune_file_name="YOUR_FINETUNE_FILE_NAME"
model_name_or_path="YOUR_PATH_TO_LLM"
# todo: this is example path format
model_save_path="$SCRIPT_DIR/model/cpm_finetuned_model"
kb_path="data/loogle/KB/KB_{id}.json"
kb_emb_path="data/loogle/embeddings/{bge_type}/sentence_embeddings_{id}_{chunk_size}.pkl"
kb_sentence_path="data/loogle/sentences/sentence_sentences_{id}_{chunk_size}.pkl"
# todo
total_CUDA=0,1,2,3
CUDA_1=0,1
CUDA_2=2,3
MASTER_PORT=12245
log_time() {
local cmd_description="$1"
shift
local cmd="$@"
local start_time=$(date +"%Y-%m-%d %H:%M:%S")
echo "START: [$cmd_description] at $start_time" >> "$LOG_FILE"
{ time bash -c "$cmd"; } 2>> "$ERROR_LOG_FILE"
local end_time=$(date +"%Y-%m-%d %H:%M:%S")
echo "END: [$cmd_description] at $end_time" >> "$LOG_FILE"
echo "----------------------------------------" >> "$LOG_FILE"
}
# todo: you can give iter_num, train_step_num and split_num directly or give total_train_step, KB_size and to calculate others.
# total_train_step= # annotation_data/(batch_size*gradient_accumulation_steps*cuda_num)
# pair_w_tokens=
# KB_size=
# iter_num=3
# # 8=batch_size*gradient_accumulation_steps
# train_step_num=$(echo "scale=0; $KB_size/8 * $pair_w_tokens / $iter_num" | bc)
# split_num=$((total_train_step / train_step_num))
iter_num=1
train_step_num=1
split_num=1
# 分割数据集
cd verify
OPTS=""
OPTS+=" --in_path ${bm_data_format_dir}/YOUR_DATASET_PATH.jsonl" # todo
OPTS+=" --output_dir ${bm_data_format_dir}/${split_num}_${train_step_num}"
OPTS+=" --num $split_num"
CMD="python split_verify.py ${OPTS}"
echo "split dataset..."
log_time "${split_num}_${train_step_num} split_verify.py" "$CMD"
# 分割的数量循环
for ((step_num=1; step_num<=$iter_num; step_num+=1)); do
export CUDA_VISIBLE_DEVICES=$total_CUDA
cd ../finetune
# 第一轮用原始的训练
last_num=$((step_num - 1))
next_num=$((step_num + 1))
echo "current path"
pwd
MASTER_ADDR=localhost
NNODES=1
NODE_RANK=0
GPUS_PER_NODE=4
DISTRIBUTED_ARGS="--nproc_per_node $GPUS_PER_NODE \
--nnodes $NNODES \
--node_rank $NODE_RANK \
--master_addr $MASTER_ADDR \
--master_port $MASTER_PORT"
OPTS=""
OPTS+=" --gradient_accumulation_steps 4" # todo
OPTS+=" --logging_step 5" # todo
OPTS+=" --batch_size_per_device 2" # todo
OPTS+=" --save_step $train_step_num"
OPTS+=" --epochs 1" # todo
# OPTS+=" --seed 42" # todo
OPTS+=" --lr 1e-5" # todo
OPTS+=" --max_seq_length 3000" # todo
OPTS+=" --weight-decay 0.1" # todo
if [ "$step_num" -eq 1 ]; then
OPTS+=" --warmup_iters 50" # todo
else
OPTS+=" --warmup_iters 0" # todo
fi
OPTS+=" --lr-decay-iters $(($iter_num * $train_step_num / 4))" # todo
# OPTS+=" --lr-decay-iters $(($train_step_num))"
OPTS+=" --lr-decay-style cosine" # todo
OPTS+=" --start-step 0"
OPTS+=" --model_name_or_path $YOUR_PATH_TO_LLM"
OPTS+=" --tensorboard ${model_save_path}/$train_step_num/iter$step_num/logs"
OPTS+=" --model mc_gold_v1"
OPTS+=" --save_dir ${model_save_path}/$train_step_num/iter$step_num"
OPTS+=" --data_dir ${bm_data_format_dir}/${split_num}_${train_step_num}"
if [ "$step_num" -eq 1 ]; then
OPTS+=" --ultra_split part$step_num"
else
OPTS+=" --load_ckpt ${model_save_path}/$train_step_num/iter$last_num/mc_gold_v1/step_$train_step_num"
OPTS+=" --ultra_split part${step_num}_verify"
fi
OPTS+=" --save_limit 5"
OPTS+=" --loss-scale 32768000"
echo "export PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:512"
CMD="torchrun ${DISTRIBUTED_ARGS} mc_finetune/src/train_bm_modelcenter.py ${OPTS}"
echo "Starting training..."
echo $CMD
log_time "train_bm_modelcenter.py (iter $step_num)" "$CMD"
# tranform the model
echo "current path"
pwd
OPTS=""
OPTS+=" --inpath ${model_save_path}/$train_step_num/iter$step_num/mc_gold_v1/step_$train_step_num"
OPTS+=" --inpath2 $YOUR_PATH_TO_LLM"
OPTS+=" --outpath ${model_save_path}/$train_step_num/iter$step_num/mc/v1_$train_step_num"
OPTS+=" --has_special false"
CMD="python3 ./bmtrainMiniCPM_hugMiniCPM.py ${OPTS}"
echo "-------final CMD is------"
echo "${CMD}"
echo "-------final CMD end------"
log_time "bmtrainMiniCPM_hugMiniCPM.py (iter $step_num)" "$CMD" &
wait
# verify
if [ "$step_num" -eq $iter_num ]; then
echo "next"
else
export CUDA_VISIBLE_DEVICES=$CUDA_1
cd ../verify
OPTS=""
OPTS+=" --kb_path $kb_path"
OPTS+=" --kb_emb_path $kb_emb_path"
OPTS+=" --kb_sentence_path $kb_sentence_path"
OPTS+=" --in_path ${bm_data_format_dir}/${split_num}_${train_step_num}/part${next_num}.jsonl"
OPTS+=" --out_path ${bm_data_format_dir}/${split_num}_${train_step_num}/part${next_num}_verify.jsonl"
OPTS+=" --model_path ${model_save_path}/$train_step_num/iter$step_num/mc/v1_$train_step_num"
OPTS+=" --language English" # todo
CMD="python3 ./gen_iterate_verify.py ${OPTS}"
echo "-------final CMD is------"
echo "${CMD}"
echo "-------final CMD end------"
log_time "gen_iterate_verify.py (iter $step_num)" "$CMD" &
fi
wait
# rm -r "${model_save_path}/$train_step_num/iter$step_num/mc/v1_$train_step_num"
done