generated from SparkJiao/pytorch-transformers-template
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathrun.sh
167 lines (96 loc) · 11 KB
/
run.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
159
160
161
162
163
164
165
166
167
MASTER_HOST=$SM_MASTER
MASTER_ADDR=$SM_MASTER_ADDR
MASTER_PORT="23456"
NNODES="$NODE_NUMBER"
NODE_RANK="$NODE_INDEX"
MODEL_S3_BUCKET=s3://sagemaker-us-east-1-107457652907/pretrained-models/llama-7b
BASE_BUCKET=s3://panda-us-west-2
wandb login d9bc4cccef46949e9fdffb3df442996d803d43d2
chmod +x ./s5cmd
# git clone https://github.com/HazyResearch/flash-attention.git
# cd flash-attention
# python setup.py install
# cd ../
# ======================================
#./s5cmd sync $MODEL_S3_BUCKET/* /tmp/llama-7b/
#OUTPUT_DIR=/tmp/llama.7b.zh_instruct.10M.v1.0.seq1024.w8.adamw.NA100.0421.ds
#AWS_OUTPUT_BUCKET=s3://sagemaker-us-east-1-107457652907/experiments/llama.7b.zh_instruct.10M.v1.0.seq1024.w8.adamw.NA100.0421.ds
#./s5cmd sync $AWS_OUTPUT_BUCKET/checkpoint-1750/* /tmp/checkpoints/checkpoint-1750/
#python -m torch.distributed.run --nproc_per_node 8 --nnodes $NNODES --node_rank $NODE_RANK --master_addr $MASTER_ADDR --master_port $MASTER_PORT trainer_base_ds_mul_aws.py output_dir=$OUTPUT_DIR aws_output_bucket=$AWS_OUTPUT_BUCKET resume=/tmp/checkpoints/checkpoint-1750 -cp conf/llama/zh/ -cn llama_7b_zh_instruct_v1_0_ds
# ============ COIG SFT ============
# ./s5cmd sync s3://sagemaker-us-east-1-107457652907/experiments/llama.7b.zh_instruct.10M.v1.0.seq1024.w8.adamw.NA100.0421.ds/checkpoint-1750/* /tmp/zh_instruct_v1_0/checkpoint-1750/
# AWS_OUTPUT_BUCKET=s3://sagemaker-us-east-1-107457652907/experiments/llama.7b.zh_instruct.10M.coig.sft.v1.0.seq2048.w8.adamw.NA100.0428.ds
# python -m torch.distributed.run --nproc_per_node 8 --nnodes $NNODES --node_rank $NODE_RANK --master_addr $MASTER_ADDR --master_port $MASTER_PORT trainer_base_ds_mul_aws.py aws_output_bucket=$AWS_OUTPUT_BUCKET -cp conf/llama/zh/ -cn llama_7b_zh_instruct_coig_sft_v1_0_ds
# ============ 13B pre-train ==================
#MODEL_S3_BUCKET=s3://sagemaker-us-east-1-107457652907/pretrained-models/llama-13b
#AWS_OUTPUT_BUCKET=s3://sagemaker-us-east-1-107457652907/experiments/llama.13b.zh_instruct.10M.v1.0.seq1024.w8.adamw.NA100.0430.ds
#
#./s5cmd sync $MODEL_S3_BUCKET/* /tmp/llama-13b/
#
#python -m torch.distributed.run --nproc_per_node 8 --nnodes $NNODES --node_rank $NODE_RANK --master_addr $MASTER_ADDR --master_port $MASTER_PORT trainer_base_ds_mul_aws.py aws_output_bucket=$AWS_OUTPUT_BUCKET -cp conf/llama/zh/ -cn llama_13b_zh_instruct_v1_0_ds
#
#./s5cmd sync /tmp/log_dir/* s3://sagemaker-us-east-1-107457652907/experiments/log_dir/
# =============== 7B OpenLLaMA pre-train ===============================
# MODEL_S3_BUCKET=s3://sagemaker-us-east-1-107457652907/pretrained-models/llama-13b
# AWS_OUTPUT_BUCKET=s3://sagemaker-us-east-1-107457652907/experiments/llama.13b.merit_v91_v91.seq2seq.v5.0.3aug.gpt4all.union.w8.adamw.500steps.NA100.0516
# ./s5cmd sync $MODEL_S3_BUCKET/* /tmp/llama-13b
# python -m torch.distributed.run --nproc_per_node 8 --nnodes $NNODES --node_rank $NODE_RANK --master_addr $MASTER_ADDR --master_port $MASTER_PORT trainer_base_ds_mul_aws.py aws_output_bucket=$AWS_OUTPUT_BUCKET -cp conf/llama/wiki/ -cn llama_13b_merit_v1_pv91_v91_v5_0_gpt4all_union_v1_0
# ./s5cmd sync /tmp/log_dir/* s3://sagemaker-us-east-1-107457652907/experiments/log_dir/
# # ================== llama-65B merit pre-train =====================
# MODEL_S3_BUCKET=s3://sagemaker-us-east-1-107457652907/pretrained-models/llama-65b
# AWS_OUTPUT_BUCKET=s3://sagemaker-us-east-1-107457652907/experiments/llama.65b.q_lora.merit_v91_v91.seq2seq.v5.14.3aug.w16.adamw.500steps.NA100.0608.pad_fix.aws
# # ./s5cmd sync $AWS_OUTPUT_BUCKET/checkpoint-60/* /tmp/checkpoint-60
# ./s5cmd sync $MODEL_S3_BUCKET/* /tmp/llama-65b
# PAD_TOKEN="<unk>" python -m torch.distributed.run --nproc_per_node 8 --nnodes $NNODES --node_rank $NODE_RANK --master_addr $MASTER_ADDR --master_port $MASTER_PORT trainer_base_ds_mul_aws.py aws_output_bucket=$AWS_OUTPUT_BUCKET -cp conf/llama/wiki/ -cn llama_65b_merit_v1_pv91_v91_v5_14_aws
# ./s5cmd sync /tmp/log_dir/* s3://sagemaker-us-east-1-107457652907/experiments/log_dir/
# MODEL_S3_BUCKET=s3://sagemaker-us-east-1-107457652907/pretrained-models/llama-65b
# AWS_OUTPUT_BUCKET=s3://sagemaker-us-east-1-107457652907/experiments/llama.65b.merit_v91_v91.seq2seq.v5.0.3aug.mp2.adamw.500steps.NA100.0616.aws
# ./s5cmd sync $AWS_OUTPUT_BUCKET/checkpoint-60/* /tmp/checkpoint-60
# ./s5cmd sync $MODEL_S3_BUCKET/* /tmp/llama-65b
# PAD_TOKEN="<unk>" python -m torch.distributed.run --nproc_per_node 8 --nnodes $NNODES --node_rank $NODE_RANK --master_addr $MASTER_ADDR --master_port $MASTER_PORT trainer_base_ds_mul_aws.py aws_output_bucket=$AWS_OUTPUT_BUCKET -cp conf/llama/wiki/ -cn llama_65b_merit_v1_pv91_v91_v5_0_full_mp
# PAD_TOKEN="<unk>" python -m torch.distributed.run --nproc_per_node 1 --nnodes $NNODES --node_rank $NODE_RANK --master_addr $MASTER_ADDR --master_port $MASTER_PORT trainer_base_v3_aws.py aws_output_bucket=$AWS_OUTPUT_BUCKET -cp conf/llama/wiki/ -cn llama_65b_merit_v1_pv91_v91_v5_0_full_mp
# MODEL_S3_BUCKET=s3://sagemaker-us-east-1-107457652907/pretrained-models/llama-65b
# AWS_OUTPUT_BUCKET=s3://sagemaker-us-east-1-107457652907/experiments/llama.65b.q_lora.merit_v91_v91.seq2seq.v5.15.3aug.w16.adamw.500steps.NA100.0617.pad_fix.aws
# ./s5cmd sync $MODEL_S3_BUCKET/* /tmp/llama-65b
# PAD_TOKEN="<unk>" python -m torch.distributed.run --nproc_per_node 8 --nnodes $NNODES --node_rank $NODE_RANK --master_addr $MASTER_ADDR --master_port $MASTER_PORT trainer_base_ds_mul_aws.py aws_output_bucket=$AWS_OUTPUT_BUCKET -cp conf/llama/wiki/ -cn llama_65b_merit_v1_pv91_v91_v5_15_aws
# MODEL_S3_BUCKET=s3://sagemaker-us-east-1-107457652907/pretrained-models/llama-65b
# AWS_OUTPUT_BUCKET=s3://sagemaker-us-east-1-107457652907/experiments/llama.65b.q_lora.merit_v91_v91.seq2seq.v5.18.3aug.w16.adamw.500steps.NA100.0617.aws
# ./s5cmd sync $MODEL_S3_BUCKET/* /tmp/llama-65b
# PAD_TOKEN="<unk>" python -m torch.distributed.run --nproc_per_node 8 --nnodes $NNODES --node_rank $NODE_RANK --master_addr $MASTER_ADDR --master_port $MASTER_PORT trainer_base_ds_mul_aws.py aws_output_bucket=$AWS_OUTPUT_BUCKET -cp conf/llama/wiki/ -cn llama_65b_merit_v1_pv91_v91_v5_18_aws
# ./s5cmd sync /tmp/log_dir/* s3://sagemaker-us-east-1-107457652907/experiments/log_dir/
# ================== llama-65b mp8 dp2 merit pre-train =====================
# MODEL_S3_BUCKET=s3://sagemaker-us-east-1-107457652907/pretrained-models/llama-65b-mp8
# # MODEL_S3_BUCKET=s3://sagemaker-us-east-1-107457652907/pretrained-models/llama-65b-mp16
# AWS_OUTPUT_BUCKET=s3://sagemaker-us-east-1-107457652907/experiments/llama.65b.merit_v91_v91.seq2seq.v5.4.3aug.mp8.dp2.adamw.500steps.NA100.0630.aws
# ./s5cmd sync $MODEL_S3_BUCKET/* /tmp/llama-65b-mp8
# # PAD_TOKEN="<unk>" python -m torch.distributed.run --nproc_per_node 8 --nnodes $NNODES --node_rank $NODE_RANK --master_addr $MASTER_ADDR --master_port $MASTER_PORT trainer_base_ds_mp_aws.py aws_output_bucket=$AWS_OUTPUT_BUCKET -cp conf/llama/mp/ -cn llama_65b_merit_v1_pv91_v91_v5_0_full_aws
# # PAD_TOKEN="<unk>" python -m torch.distributed.run --nproc_per_node 8 --nnodes $NNODES --node_rank $NODE_RANK --master_addr $MASTER_ADDR --master_port $MASTER_PORT trainer_base_ds_mp_aws.py aws_output_bucket=$AWS_OUTPUT_BUCKET -cp conf/llama/mp/ -cn llama_65b_merit_v1_pv91_v91_v5_2_full_aws
# # PAD_TOKEN="<unk>" python -m torch.distributed.run --nproc_per_node 8 --nnodes $NNODES --node_rank $NODE_RANK --master_addr $MASTER_ADDR --master_port $MASTER_PORT trainer_base_ds_mp_aws.py aws_output_bucket=$AWS_OUTPUT_BUCKET -cp conf/llama/mp/ -cn llama_65b_merit_v1_pv91_v91_v5_3_full_aws
# PAD_TOKEN="<unk>" python -m torch.distributed.run --nproc_per_node 8 --nnodes $NNODES --node_rank $NODE_RANK --master_addr $MASTER_ADDR --master_port $MASTER_PORT trainer_base_ds_mp_aws.py aws_output_bucket=$AWS_OUTPUT_BUCKET -cp conf/llama/mp/ -cn llama_65b_merit_v1_pv91_v91_v5_4_full_aws
# # PAD_TOKEN="<unk>" deepspeed --num_nodes 2 --master_addr $MASTER_ADDR --master_port $MASTER_PORT trainer_base_ds_mp_aws.py aws_output_bucket=$AWS_OUTPUT_BUCKET -cp conf/llama/mp/ -cn llama_65b_merit_v1_pv91_v91_v5_0_full_aws
# # PAD_TOKEN="<unk>" python -m torch.distributed.run --nproc_per_node 8 --nnodes $NNODES --node_rank $NODE_RANK --master_addr $MASTER_ADDR --master_port $MASTER_PORT trainer_ds_mp_unify_aws.py aws_output_bucket=$AWS_OUTPUT_BUCKET -cp conf/llama/mp/ -cn llama_65b_merit_v1_pv91_v91_v5_0_full_aws
# ./s5cmd sync /tmp/log_dir/* s3://sagemaker-us-east-1-107457652907/experiments/log_dir/
# ================== falcon-40b merit pre-train =====================
# MODEL_S3_BUCKET=s3://sagemaker-us-east-1-107457652907/pretrained-models/falcon-40b
# AWS_OUTPUT_BUCKET=s3://sagemaker-us-east-1-107457652907/experiments/falcon.40b.q_lora.merit_v91_v91.seq2seq.v5.0.3aug.w16.adamw.500steps.NA100.0528.aws
# ./s5cmd sync $MODEL_S3_BUCKET/* /tmp/falcon-40b
# python -m torch.distributed.run --nproc_per_node 8 --nnodes $NNODES --node_rank $NODE_RANK --master_addr $MASTER_ADDR --master_port $MASTER_PORT trainer_base_ds_mul_aws.py aws_output_bucket=$AWS_OUTPUT_BUCKET -cp conf/rw/ -cn falcon_40b_merit_v1_pv91_v91_v5_0_aws
# ./s5cmd sync /tmp/log_dir/* s3://sagemaker-us-east-1-107457652907/experiments/log_dir/
# ========================= Llama2-70b-CoT
# # MODEL_S3_BUCKET=s3://sagemaker-us-east-1-107457652907/pretrained-models/Llama-2-70b-mp
# MODEL_S3_BUCKET=s3://sagemaker-us-east-1-107457652907/pretrained-models/Llama-2-70b-hf
# AWS_OUTPUT_BUCKET=s3://sagemaker-us-east-1-107457652907/experiments/llama2.70b.act.cot.pp8.dp2.A100.v1.0.0808
# # ./s5cmd sync $MODEL_S3_BUCKET/* /tmp/Llama-2-70b-mp
# ./s5cmd sync $MODEL_S3_BUCKET/* /tmp/Llama-2-70b-hf
# ./s5cmd sync s3://sagemaker-us-east-1-107457652907/fangkai/cot-data/* /tmp/data-train
# # python -m torch.distributed.run --nproc_per_node 8 --nnodes $NNODES --node_rank $NODE_RANK --master_addr $MASTER_ADDR --master_port $MASTER_PORT trainer_base_ds_mp_aws.py aws_output_bucket=$AWS_OUTPUT_BUCKET -cp conf/llama2/cot_actor -cn llama2_70b_cot_tk_rank_v1_0_aws
# python -m torch.distributed.run --nproc_per_node 8 --nnodes $NNODES --node_rank $NODE_RANK --master_addr $MASTER_ADDR --master_port $MASTER_PORT trainer_base_ds_mul_aws.py aws_output_bucket=$AWS_OUTPUT_BUCKET -cp conf/llama2/cot_actor -cn llama2_70b_qlora_cot_tk_rank_v1_0_aws
# ./s5cmd sync /tmp/log_dir/* s3://sagemaker-us-east-1-107457652907/experiments/log_dir/
# ============================ Llama2-70b-CoT-QLoRA
MODEL_S3_BUCKET=$BASE_BUCKET/pretrained-models/Llama-2-70b-hf
AWS_OUTPUT_BUCKET=$BASE_BUCKET/experiments/llama2.70b.q_lora.act.cot.w16.A100.v1.0.0810
./s5cmd sync $MODEL_S3_BUCKET/* /tmp/Llama-2-70b-hf
./s5cmd sync $BASE_BUCKET/fangkai/cot-data/* /tmp/data-train
# python -m torch.distributed.run --nproc_per_node 8 --nnodes $NNODES --node_rank $NODE_RANK --master_addr $MASTER_ADDR --master_port $MASTER_PORT trainer_base_ds_mp_aws.py aws_output_bucket=$AWS_OUTPUT_BUCKET -cp conf/llama2/cot_actor -cn llama2_70b_cot_tk_rank_v1_0_aws
python -m torch.distributed.run --nproc_per_node 8 --nnodes $NNODES --node_rank $NODE_RANK --master_addr $MASTER_ADDR --master_port $MASTER_PORT trainer_base_ds_mul_aws.py aws_output_bucket=$AWS_OUTPUT_BUCKET -cp conf/llama2/cot_actor -cn llama2_70b_qlora_cot_tk_rank_v1_0_aws
./s5cmd sync /tmp/log_dir/* $BASE_BUCKET/experiments/log_dir/