Muennighoff commited on
Commit
77b513d
1 Parent(s): efafe69
perplexity50/sbatch_2b855bc4perplexity50.sh ADDED
@@ -0,0 +1,165 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/bin/bash
2
+ #SBATCH --exclude=nid007571,nid007112,nid006774,nid007502,nid007506,nid007507,nid005145,nid006692,nid007218,nid007123,nid006124,nid006123,nid007496,nid007237,nid006852,nid007206,nid006947,nid007212,nid006977,nid007222,nid005444,nid007219,nid007493,nid007221,nid005300,nid005619,nid006118,nid005203,nid006113,nid006481,nid007077,nid005208,nid005207,nid005879,nid005901
3
+ #SBATCH --nodes=32
4
+ #SBATCH --ntasks-per-node=1
5
+ #SBATCH --cpus-per-task=32
6
+ #SBATCH --mem=256G
7
+ #SBATCH -p standard-g
8
+ #SBATCH -t 48:00:00
9
+ #SBATCH --gpus-per-node=mi250:8
10
+ #SBATCH --exclusive=user
11
+ #SBATCH --hint=nomultithread
12
+ #SBATCH --account=project_462000119
13
+ #SBATCH -o logs/%j.out
14
+ #SBATCH -e logs/%j.err
15
+
16
+ VARIANT=2b855b55bc4perplexity50
17
+
18
+ # if run without sbatch, invoke here
19
+ if [ -z $SLURM_JOB_ID ]; then
20
+ mkdir -p logs
21
+ sbatch "$0"
22
+ exit
23
+ fi
24
+
25
+ set -euo pipefail
26
+
27
+ # symlink logs/latest.out and logs/latest.err
28
+ ln -f -s $SLURM_JOB_ID.out logs/latest.out
29
+ ln -f -s $SLURM_JOB_ID.err logs/latest.err
30
+
31
+ KILL_SWITCH_PATH=kill-switch-$VARIANT
32
+ CHECKPOINT_PATH=checkpoints_$VARIANT
33
+ TENSORBOARD_PATH=tensorboard_$VARIANT
34
+
35
+ # Data
36
+ VOCAB_FILE="gpt2/vocab.json"
37
+ MERGE_FILE="gpt2/merges.txt"
38
+ #TRAIN_DATA_PATH="/scratch/project_462000119/data/c4_subsampled/gpt2tok_c4_en_10B_text_document"
39
+ #VAL_DATA_PATH="/scratch/project_462000119/data/c4_validation/gpt2tok_c4validation_rerun_text_document"
40
+
41
+ TRAIN_DATA_PATH=trainperplexity50.txt
42
+ # "train: 1.0 0:1 /scratch/project_462000119/data/c4_subsampled/gpt2tok_c4_en_55B_text_document"
43
+ VALID_DATA_PATH=val.txt
44
+ # "validation: 1.0 0:1 /scratch/project_462000119/data/c4_validation/gpt2tok_c4validation_rerun_text_document"
45
+
46
+ PP_SIZE=1
47
+ TP_SIZE=1
48
+
49
+ MICRO_BATCH_SIZE=2
50
+ GRADIENT_ACCUMULATION_STEPS=1
51
+ WORLD_SIZE=$((SLURM_GPUS_ON_NODE*SLURM_JOB_NUM_NODES))
52
+ GLOBAL_BATCH_SIZE=$((MICRO_BATCH_SIZE*WORLD_SIZE*GRADIENT_ACCUMULATION_STEPS))
53
+
54
+ # Model parameters
55
+ source model_params.sh
56
+ MODEL_PARAM=("${PARAM_2980M[@]}")
57
+ NHIDDEN=${MODEL_PARAM[0]}
58
+ FFN_HIDDEN_SIZE=${MODEL_PARAM[1]}
59
+ KV_SIZE=${MODEL_PARAM[2]}
60
+ NHEADS=${MODEL_PARAM[3]}
61
+ NLAYERS=${MODEL_PARAM[4]}
62
+ SEQ_LEN=2048
63
+
64
+ echo "Model parameters: d_model $NHIDDEN ffw_size $FFN_HIDDEN_SIZE kv_size $KV_SIZE n_heads $NHEADS n_layers $NLAYERS"
65
+
66
+ SAVE_INTERVAL=10000
67
+
68
+ # Tokens: 55000000000
69
+ # -> Samples: 26855469
70
+ TRAIN_SAMPLES=26_855_469
71
+
72
+ OPTIMIZER_ARGS=" \
73
+ --optimizer adam \
74
+ --adam-beta1 0.9 \
75
+ --adam-beta2 0.95 \
76
+ --adam-eps 1e-8 \
77
+ --lr 2e-4 \
78
+ --min-lr 2e-5 \
79
+ --lr-decay-style cosine \
80
+ --lr-decay-samples $TRAIN_SAMPLES \
81
+ --lr-warmup-samples 268_555 \
82
+ --clip-grad 1.0 \
83
+ --weight-decay 1e-1 \
84
+ "
85
+
86
+ GPT_ARGS=" \
87
+ --num-layers $NLAYERS \
88
+ --hidden-size $NHIDDEN \
89
+ --num-attention-heads $NHEADS \
90
+ --kv-channels $KV_SIZE \
91
+ --ffn-hidden-size $FFN_HIDDEN_SIZE \
92
+ --seq-length $SEQ_LEN \
93
+ --max-position-embeddings $SEQ_LEN \
94
+ --micro-batch-size $MICRO_BATCH_SIZE \
95
+ --global-batch-size $GLOBAL_BATCH_SIZE \
96
+ --train-samples $TRAIN_SAMPLES \
97
+ --vocab-file $VOCAB_FILE \
98
+ --merge-file $MERGE_FILE \
99
+ --clip-grad 1.0 \
100
+ --kill-switch-path $KILL_SWITCH_PATH \
101
+ --bf16 \
102
+ $OPTIMIZER_ARGS \
103
+ "
104
+
105
+ OUTPUT_ARGS=" \
106
+ --log-interval 10 \
107
+ --save-interval $SAVE_INTERVAL \
108
+ --eval-interval 100 \
109
+ --eval-iters 100 \
110
+ --tensorboard-dir $TENSORBOARD_PATH \
111
+ --tensorboard-queue-size 5 \
112
+ --log-timers-to-tensorboard \
113
+ --log-batch-size-to-tensorboard \
114
+ --log-validation-ppl-to-tensorboard \
115
+ "
116
+
117
+ ZERO_STAGE=0
118
+
119
+ mkdir -p ds_configs
120
+ DS_CONFIG_PATH="ds_configs/$SLURM_JOB_ID.json"
121
+
122
+ cat <<EOF > $DS_CONFIG_PATH
123
+ {
124
+ "train_micro_batch_size_per_gpu": $MICRO_BATCH_SIZE,
125
+ "train_batch_size": $GLOBAL_BATCH_SIZE,
126
+ "gradient_clipping": 1.0,
127
+ "zero_optimization": {
128
+ "stage": $ZERO_STAGE
129
+ },
130
+ "bf16": {
131
+ "enabled": true
132
+ },
133
+ "steps_per_print": 2000,
134
+ "wall_clock_breakdown": false
135
+ }
136
+ EOF
137
+
138
+ DEEPSPEED_ARGS=" \
139
+ --deepspeed \
140
+ --deepspeed_config $DS_CONFIG_PATH \
141
+ --zero-stage $ZERO_STAGE \
142
+ "
143
+
144
+ CMD=" \
145
+ Megatron-DeepSpeed/pretrain_gpt.py \
146
+ --tensor-model-parallel-size $TP_SIZE \
147
+ --pipeline-model-parallel-size $PP_SIZE \
148
+ $GPT_ARGS \
149
+ $OUTPUT_ARGS \
150
+ --save $CHECKPOINT_PATH \
151
+ --load $CHECKPOINT_PATH \
152
+ --train-weighted-split-paths-path $TRAIN_DATA_PATH \
153
+ --valid-weighted-split-paths-path $VALID_DATA_PATH \
154
+ --data-impl mmap \
155
+ $DEEPSPEED_ARGS \
156
+ "
157
+
158
+ echo $CMD
159
+
160
+ echo "START $SLURM_JOBID: $(date)"
161
+
162
+ # bash launch_srun.sh $CMD
163
+ srun --label launch.sh $CMD
164
+
165
+ echo "END $SLURM_JOBID: $(date)"