Muennighoff commited on
Commit
1539ad6
1 Parent(s): e663995
4b284b6bc4seed1/3484067.err ADDED
The diff for this file is too large to render. See raw diff
 
4b284b6bc4seed1/3484067.out ADDED
The diff for this file is too large to render. See raw diff
 
4b284b6bc4seed1/sbatch_4b284b6bseed1val.sh ADDED
@@ -0,0 +1,169 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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=40
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=4b284b6bc4seed1val
17
+ VARIANT_CKPT=lm1-4b2-84b-c4seeds/4b284b6bc4seed1
18
+
19
+ # if run without sbatch, invoke here
20
+ if [ -z $SLURM_JOB_ID ]; then
21
+ mkdir -p logs
22
+ sbatch "$0"
23
+ exit
24
+ fi
25
+
26
+ set -euo pipefail
27
+
28
+ # symlink logs/latest.out and logs/latest.err
29
+ ln -f -s $SLURM_JOB_ID.out logs/latest.out
30
+ ln -f -s $SLURM_JOB_ID.err logs/latest.err
31
+
32
+ KILL_SWITCH_PATH=kill-switch-$VARIANT
33
+ CHECKPOINT_PATH=$VARIANT_CKPT
34
+ TENSORBOARD_PATH=tensorboard_$VARIANT
35
+
36
+ # Data
37
+ VOCAB_FILE="gpt2/vocab.json"
38
+ MERGE_FILE="gpt2/merges.txt"
39
+ #DATA_PATH="/scratch/project_462000119/data/pile/megatron_data/meg-gpt2_pile_text_document"
40
+ TRAIN_DATA_PATH=train1b5.txt
41
+ # "train: 1.0 0:1 /scratch/project_462000119/data/c4_subsampled/gpt2tok_c4_en_12B_text_document"
42
+ VALID_DATA_PATH=val.txt
43
+ # "validation: 1.0 0:1 /scratch/project_462000119/data/c4_validation/gpt2tok_c4validation_rerun_text_document"
44
+
45
+ PP_SIZE=1
46
+ TP_SIZE=2
47
+
48
+ MICRO_BATCH_SIZE=2
49
+ GRADIENT_ACCUMULATION_STEPS=2
50
+ WORLD_SIZE=$((SLURM_GPUS_ON_NODE*SLURM_JOB_NUM_NODES))
51
+ GLOBAL_BATCH_SIZE=$((MICRO_BATCH_SIZE*WORLD_SIZE*GRADIENT_ACCUMULATION_STEPS))
52
+
53
+ # Model parameters
54
+ source model_params.sh
55
+ MODEL_PARAM=("${PARAM_4516M[@]}")
56
+ NHIDDEN=${MODEL_PARAM[0]}
57
+ FFN_HIDDEN_SIZE=${MODEL_PARAM[1]}
58
+ KV_SIZE=${MODEL_PARAM[2]}
59
+ NHEADS=${MODEL_PARAM[3]}
60
+ NLAYERS=${MODEL_PARAM[4]}
61
+ SEQ_LEN=2048
62
+
63
+ echo "Model parameters: d_model $NHIDDEN ffw_size $FFN_HIDDEN_SIZE kv_size $KV_SIZE n_heads $NHEADS n_layers $NLAYERS"
64
+
65
+ SAVE_INTERVAL=1000
66
+
67
+ # Tokens: 25611230000
68
+ # -> Samples: 12505484
69
+ TRAIN_SAMPLES=1
70
+
71
+ OPTIMIZER_ARGS=" \
72
+ --optimizer adam \
73
+ --adam-beta1 0.9 \
74
+ --adam-beta2 0.999 \
75
+ --adam-eps 1e-8 \
76
+ --lr 2e-4 \
77
+ --min-lr 2e-5 \
78
+ --lr-decay-style cosine \
79
+ --lr-decay-samples $TRAIN_SAMPLES \
80
+ --lr-warmup-samples 0 \
81
+ --clip-grad 1.0 \
82
+ --weight-decay 1e-1 \
83
+ --override-lr-scheduler \
84
+ --reset-progress \
85
+ --no-load-optim \
86
+ "
87
+
88
+ GPT_ARGS=" \
89
+ --num-layers $NLAYERS \
90
+ --hidden-size $NHIDDEN \
91
+ --num-attention-heads $NHEADS \
92
+ --kv-channels $KV_SIZE \
93
+ --ffn-hidden-size $FFN_HIDDEN_SIZE \
94
+ --seq-length $SEQ_LEN \
95
+ --max-position-embeddings $SEQ_LEN \
96
+ --micro-batch-size $MICRO_BATCH_SIZE \
97
+ --global-batch-size $GLOBAL_BATCH_SIZE \
98
+ --train-samples $TRAIN_SAMPLES \
99
+ --vocab-file $VOCAB_FILE \
100
+ --merge-file $MERGE_FILE \
101
+ --clip-grad 1.0 \
102
+ --kill-switch-path $KILL_SWITCH_PATH \
103
+ --bf16 \
104
+ $OPTIMIZER_ARGS \
105
+ "
106
+
107
+ OUTPUT_ARGS=" \
108
+ --log-interval 10 \
109
+ --save-interval $SAVE_INTERVAL \
110
+ --eval-interval 1 \
111
+ --eval-iters 100 \
112
+ --eval-only true \
113
+ --tensorboard-dir $TENSORBOARD_PATH \
114
+ --tensorboard-queue-size 5 \
115
+ --log-timers-to-tensorboard \
116
+ --log-batch-size-to-tensorboard \
117
+ --log-validation-ppl-to-tensorboard \
118
+ "
119
+
120
+ ZERO_STAGE=0
121
+
122
+ mkdir -p ds_configs
123
+ DS_CONFIG_PATH="ds_configs/$SLURM_JOB_ID.json"
124
+
125
+ cat <<EOF > $DS_CONFIG_PATH
126
+ {
127
+ "train_micro_batch_size_per_gpu": $MICRO_BATCH_SIZE,
128
+ "train_batch_size": $GLOBAL_BATCH_SIZE,
129
+ "gradient_clipping": 1.0,
130
+ "zero_optimization": {
131
+ "stage": $ZERO_STAGE
132
+ },
133
+ "bf16": {
134
+ "enabled": true
135
+ },
136
+ "steps_per_print": 2000,
137
+ "wall_clock_breakdown": false
138
+ }
139
+ EOF
140
+
141
+ DEEPSPEED_ARGS=" \
142
+ --deepspeed \
143
+ --deepspeed_config $DS_CONFIG_PATH \
144
+ --zero-stage $ZERO_STAGE \
145
+ "
146
+
147
+ CMD=" \
148
+ Megatron-DeepSpeed/pretrain_gpt.py \
149
+ --tensor-model-parallel-size $TP_SIZE \
150
+ --pipeline-model-parallel-size $PP_SIZE \
151
+ $GPT_ARGS \
152
+ $OUTPUT_ARGS \
153
+ --save $CHECKPOINT_PATH \
154
+ --load $CHECKPOINT_PATH \
155
+ --train-weighted-split-paths-path $TRAIN_DATA_PATH \
156
+ --valid-weighted-split-paths-path $VALID_DATA_PATH \
157
+ --data-impl mmap \
158
+ $DEEPSPEED_ARGS \
159
+ --seed 1 \
160
+ "
161
+
162
+ echo $CMD
163
+
164
+ echo "START $SLURM_JOBID: $(date)"
165
+
166
+ # bash launch_srun_32.sh $CMD
167
+ srun --label launch.sh $CMD
168
+
169
+ echo "END $SLURM_JOBID: $(date)"
4b284b6bc4seed1/tensorboard_4b284b6bc4seed1val/events.out.tfevents.1683628794.nid005245.82784.0 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3a6085e9adfdcefbe8cf66585b46e0329295c6269e7ffd04c5ea3db31f0bea53
3
+ size 980
4b284b6bc4seed2/3484232.err ADDED
The diff for this file is too large to render. See raw diff
 
4b284b6bc4seed2/3484232.out ADDED
The diff for this file is too large to render. See raw diff
 
4b284b6bc4seed2/sbatch_4b284b6bc4seed2val.sh ADDED
@@ -0,0 +1,169 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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=40
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=4b284b6bc4seed2val
17
+ VARIANT_CKPT=lm1-4b2-84b-c4seeds/4b284b6bc4seed2
18
+
19
+ # if run without sbatch, invoke here
20
+ if [ -z $SLURM_JOB_ID ]; then
21
+ mkdir -p logs
22
+ sbatch "$0"
23
+ exit
24
+ fi
25
+
26
+ set -euo pipefail
27
+
28
+ # symlink logs/latest.out and logs/latest.err
29
+ ln -f -s $SLURM_JOB_ID.out logs/latest.out
30
+ ln -f -s $SLURM_JOB_ID.err logs/latest.err
31
+
32
+ KILL_SWITCH_PATH=kill-switch-$VARIANT
33
+ CHECKPOINT_PATH=$VARIANT_CKPT
34
+ TENSORBOARD_PATH=tensorboard_$VARIANT
35
+
36
+ # Data
37
+ VOCAB_FILE="gpt2/vocab.json"
38
+ MERGE_FILE="gpt2/merges.txt"
39
+ #DATA_PATH="/scratch/project_462000119/data/pile/megatron_data/meg-gpt2_pile_text_document"
40
+ TRAIN_DATA_PATH=train1b5.txt
41
+ # "train: 1.0 0:1 /scratch/project_462000119/data/c4_subsampled/gpt2tok_c4_en_12B_text_document"
42
+ VALID_DATA_PATH=val.txt
43
+ # "validation: 1.0 0:1 /scratch/project_462000119/data/c4_validation/gpt2tok_c4validation_rerun_text_document"
44
+
45
+ PP_SIZE=1
46
+ TP_SIZE=2
47
+
48
+ MICRO_BATCH_SIZE=2
49
+ GRADIENT_ACCUMULATION_STEPS=2
50
+ WORLD_SIZE=$((SLURM_GPUS_ON_NODE*SLURM_JOB_NUM_NODES))
51
+ GLOBAL_BATCH_SIZE=$((MICRO_BATCH_SIZE*WORLD_SIZE*GRADIENT_ACCUMULATION_STEPS))
52
+
53
+ # Model parameters
54
+ source model_params.sh
55
+ MODEL_PARAM=("${PARAM_4516M[@]}")
56
+ NHIDDEN=${MODEL_PARAM[0]}
57
+ FFN_HIDDEN_SIZE=${MODEL_PARAM[1]}
58
+ KV_SIZE=${MODEL_PARAM[2]}
59
+ NHEADS=${MODEL_PARAM[3]}
60
+ NLAYERS=${MODEL_PARAM[4]}
61
+ SEQ_LEN=2048
62
+
63
+ echo "Model parameters: d_model $NHIDDEN ffw_size $FFN_HIDDEN_SIZE kv_size $KV_SIZE n_heads $NHEADS n_layers $NLAYERS"
64
+
65
+ SAVE_INTERVAL=1000
66
+
67
+ # Tokens: 25611230000
68
+ # -> Samples: 12505484
69
+ TRAIN_SAMPLES=1
70
+
71
+ OPTIMIZER_ARGS=" \
72
+ --optimizer adam \
73
+ --adam-beta1 0.9 \
74
+ --adam-beta2 0.999 \
75
+ --adam-eps 1e-8 \
76
+ --lr 2e-4 \
77
+ --min-lr 2e-5 \
78
+ --lr-decay-style cosine \
79
+ --lr-decay-samples $TRAIN_SAMPLES \
80
+ --lr-warmup-samples 0 \
81
+ --clip-grad 1.0 \
82
+ --weight-decay 1e-1 \
83
+ --override-lr-scheduler \
84
+ --reset-progress \
85
+ --no-load-optim \
86
+ "
87
+
88
+ GPT_ARGS=" \
89
+ --num-layers $NLAYERS \
90
+ --hidden-size $NHIDDEN \
91
+ --num-attention-heads $NHEADS \
92
+ --kv-channels $KV_SIZE \
93
+ --ffn-hidden-size $FFN_HIDDEN_SIZE \
94
+ --seq-length $SEQ_LEN \
95
+ --max-position-embeddings $SEQ_LEN \
96
+ --micro-batch-size $MICRO_BATCH_SIZE \
97
+ --global-batch-size $GLOBAL_BATCH_SIZE \
98
+ --train-samples $TRAIN_SAMPLES \
99
+ --vocab-file $VOCAB_FILE \
100
+ --merge-file $MERGE_FILE \
101
+ --clip-grad 1.0 \
102
+ --kill-switch-path $KILL_SWITCH_PATH \
103
+ --bf16 \
104
+ $OPTIMIZER_ARGS \
105
+ "
106
+
107
+ OUTPUT_ARGS=" \
108
+ --log-interval 10 \
109
+ --save-interval $SAVE_INTERVAL \
110
+ --eval-interval 1 \
111
+ --eval-iters 100 \
112
+ --eval-only true \
113
+ --tensorboard-dir $TENSORBOARD_PATH \
114
+ --tensorboard-queue-size 5 \
115
+ --log-timers-to-tensorboard \
116
+ --log-batch-size-to-tensorboard \
117
+ --log-validation-ppl-to-tensorboard \
118
+ "
119
+
120
+ ZERO_STAGE=0
121
+
122
+ mkdir -p ds_configs
123
+ DS_CONFIG_PATH="ds_configs/$SLURM_JOB_ID.json"
124
+
125
+ cat <<EOF > $DS_CONFIG_PATH
126
+ {
127
+ "train_micro_batch_size_per_gpu": $MICRO_BATCH_SIZE,
128
+ "train_batch_size": $GLOBAL_BATCH_SIZE,
129
+ "gradient_clipping": 1.0,
130
+ "zero_optimization": {
131
+ "stage": $ZERO_STAGE
132
+ },
133
+ "bf16": {
134
+ "enabled": true
135
+ },
136
+ "steps_per_print": 2000,
137
+ "wall_clock_breakdown": false
138
+ }
139
+ EOF
140
+
141
+ DEEPSPEED_ARGS=" \
142
+ --deepspeed \
143
+ --deepspeed_config $DS_CONFIG_PATH \
144
+ --zero-stage $ZERO_STAGE \
145
+ "
146
+
147
+ CMD=" \
148
+ Megatron-DeepSpeed/pretrain_gpt.py \
149
+ --tensor-model-parallel-size $TP_SIZE \
150
+ --pipeline-model-parallel-size $PP_SIZE \
151
+ $GPT_ARGS \
152
+ $OUTPUT_ARGS \
153
+ --save $CHECKPOINT_PATH \
154
+ --load $CHECKPOINT_PATH \
155
+ --train-weighted-split-paths-path $TRAIN_DATA_PATH \
156
+ --valid-weighted-split-paths-path $VALID_DATA_PATH \
157
+ --data-impl mmap \
158
+ $DEEPSPEED_ARGS \
159
+ --seed 2 \
160
+ "
161
+
162
+ echo $CMD
163
+
164
+ echo "START $SLURM_JOBID: $(date)"
165
+
166
+ # bash launch_srun_32.sh $CMD
167
+ srun --label launch.sh $CMD
168
+
169
+ echo "END $SLURM_JOBID: $(date)"
4b284b6bc4seed2/tensorboard_4b284b6bc4seed2val/events.out.tfevents.1683631386.nid005245.103217.0 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c96a278778e738f489cd608ef9f2aba84c489140ec4156da04bc475b5d405f50
3
+ size 980
4b284b6bc4seed3/3484239.err ADDED
The diff for this file is too large to render. See raw diff
 
4b284b6bc4seed3/3484239.out ADDED
The diff for this file is too large to render. See raw diff
 
4b284b6bc4seed3/sbatch_4b284b6bc4seed3val.sh ADDED
@@ -0,0 +1,169 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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=40
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=4b284b6bc4seed3val
17
+ VARIANT_CKPT=lm1-4b2-84b-c4seeds/4b284b6bc4seed3
18
+
19
+ # if run without sbatch, invoke here
20
+ if [ -z $SLURM_JOB_ID ]; then
21
+ mkdir -p logs
22
+ sbatch "$0"
23
+ exit
24
+ fi
25
+
26
+ set -euo pipefail
27
+
28
+ # symlink logs/latest.out and logs/latest.err
29
+ ln -f -s $SLURM_JOB_ID.out logs/latest.out
30
+ ln -f -s $SLURM_JOB_ID.err logs/latest.err
31
+
32
+ KILL_SWITCH_PATH=kill-switch-$VARIANT
33
+ CHECKPOINT_PATH=$VARIANT_CKPT
34
+ TENSORBOARD_PATH=tensorboard_$VARIANT
35
+
36
+ # Data
37
+ VOCAB_FILE="gpt2/vocab.json"
38
+ MERGE_FILE="gpt2/merges.txt"
39
+ #DATA_PATH="/scratch/project_462000119/data/pile/megatron_data/meg-gpt2_pile_text_document"
40
+ TRAIN_DATA_PATH=train1b5.txt
41
+ # "train: 1.0 0:1 /scratch/project_462000119/data/c4_subsampled/gpt2tok_c4_en_12B_text_document"
42
+ VALID_DATA_PATH=val.txt
43
+ # "validation: 1.0 0:1 /scratch/project_462000119/data/c4_validation/gpt2tok_c4validation_rerun_text_document"
44
+
45
+ PP_SIZE=1
46
+ TP_SIZE=2
47
+
48
+ MICRO_BATCH_SIZE=2
49
+ GRADIENT_ACCUMULATION_STEPS=2
50
+ WORLD_SIZE=$((SLURM_GPUS_ON_NODE*SLURM_JOB_NUM_NODES))
51
+ GLOBAL_BATCH_SIZE=$((MICRO_BATCH_SIZE*WORLD_SIZE*GRADIENT_ACCUMULATION_STEPS))
52
+
53
+ # Model parameters
54
+ source model_params.sh
55
+ MODEL_PARAM=("${PARAM_4516M[@]}")
56
+ NHIDDEN=${MODEL_PARAM[0]}
57
+ FFN_HIDDEN_SIZE=${MODEL_PARAM[1]}
58
+ KV_SIZE=${MODEL_PARAM[2]}
59
+ NHEADS=${MODEL_PARAM[3]}
60
+ NLAYERS=${MODEL_PARAM[4]}
61
+ SEQ_LEN=2048
62
+
63
+ echo "Model parameters: d_model $NHIDDEN ffw_size $FFN_HIDDEN_SIZE kv_size $KV_SIZE n_heads $NHEADS n_layers $NLAYERS"
64
+
65
+ SAVE_INTERVAL=1000
66
+
67
+ # Tokens: 25611230000
68
+ # -> Samples: 12505484
69
+ TRAIN_SAMPLES=1
70
+
71
+ OPTIMIZER_ARGS=" \
72
+ --optimizer adam \
73
+ --adam-beta1 0.9 \
74
+ --adam-beta2 0.999 \
75
+ --adam-eps 1e-8 \
76
+ --lr 2e-4 \
77
+ --min-lr 2e-5 \
78
+ --lr-decay-style cosine \
79
+ --lr-decay-samples $TRAIN_SAMPLES \
80
+ --lr-warmup-samples 0 \
81
+ --clip-grad 1.0 \
82
+ --weight-decay 1e-1 \
83
+ --override-lr-scheduler \
84
+ --reset-progress \
85
+ --no-load-optim \
86
+ "
87
+
88
+ GPT_ARGS=" \
89
+ --num-layers $NLAYERS \
90
+ --hidden-size $NHIDDEN \
91
+ --num-attention-heads $NHEADS \
92
+ --kv-channels $KV_SIZE \
93
+ --ffn-hidden-size $FFN_HIDDEN_SIZE \
94
+ --seq-length $SEQ_LEN \
95
+ --max-position-embeddings $SEQ_LEN \
96
+ --micro-batch-size $MICRO_BATCH_SIZE \
97
+ --global-batch-size $GLOBAL_BATCH_SIZE \
98
+ --train-samples $TRAIN_SAMPLES \
99
+ --vocab-file $VOCAB_FILE \
100
+ --merge-file $MERGE_FILE \
101
+ --clip-grad 1.0 \
102
+ --kill-switch-path $KILL_SWITCH_PATH \
103
+ --bf16 \
104
+ $OPTIMIZER_ARGS \
105
+ "
106
+
107
+ OUTPUT_ARGS=" \
108
+ --log-interval 10 \
109
+ --save-interval $SAVE_INTERVAL \
110
+ --eval-interval 1 \
111
+ --eval-iters 100 \
112
+ --eval-only true \
113
+ --tensorboard-dir $TENSORBOARD_PATH \
114
+ --tensorboard-queue-size 5 \
115
+ --log-timers-to-tensorboard \
116
+ --log-batch-size-to-tensorboard \
117
+ --log-validation-ppl-to-tensorboard \
118
+ "
119
+
120
+ ZERO_STAGE=0
121
+
122
+ mkdir -p ds_configs
123
+ DS_CONFIG_PATH="ds_configs/$SLURM_JOB_ID.json"
124
+
125
+ cat <<EOF > $DS_CONFIG_PATH
126
+ {
127
+ "train_micro_batch_size_per_gpu": $MICRO_BATCH_SIZE,
128
+ "train_batch_size": $GLOBAL_BATCH_SIZE,
129
+ "gradient_clipping": 1.0,
130
+ "zero_optimization": {
131
+ "stage": $ZERO_STAGE
132
+ },
133
+ "bf16": {
134
+ "enabled": true
135
+ },
136
+ "steps_per_print": 2000,
137
+ "wall_clock_breakdown": false
138
+ }
139
+ EOF
140
+
141
+ DEEPSPEED_ARGS=" \
142
+ --deepspeed \
143
+ --deepspeed_config $DS_CONFIG_PATH \
144
+ --zero-stage $ZERO_STAGE \
145
+ "
146
+
147
+ CMD=" \
148
+ Megatron-DeepSpeed/pretrain_gpt.py \
149
+ --tensor-model-parallel-size $TP_SIZE \
150
+ --pipeline-model-parallel-size $PP_SIZE \
151
+ $GPT_ARGS \
152
+ $OUTPUT_ARGS \
153
+ --save $CHECKPOINT_PATH \
154
+ --load $CHECKPOINT_PATH \
155
+ --train-weighted-split-paths-path $TRAIN_DATA_PATH \
156
+ --valid-weighted-split-paths-path $VALID_DATA_PATH \
157
+ --data-impl mmap \
158
+ $DEEPSPEED_ARGS \
159
+ --seed 3 \
160
+ "
161
+
162
+ echo $CMD
163
+
164
+ echo "START $SLURM_JOBID: $(date)"
165
+
166
+ # bash launch_srun_32.sh $CMD
167
+ srun --label launch.sh $CMD
168
+
169
+ echo "END $SLURM_JOBID: $(date)"
4b284b6bc4seed3/sbatch_4b284b6bc4seed4val.sh ADDED
@@ -0,0 +1,169 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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=40
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=4b284b6bc4seed4val
17
+ VARIANT_CKPT=lm1-4b2-84b-c4seeds/4b284b6bc4seed4
18
+
19
+ # if run without sbatch, invoke here
20
+ if [ -z $SLURM_JOB_ID ]; then
21
+ mkdir -p logs
22
+ sbatch "$0"
23
+ exit
24
+ fi
25
+
26
+ set -euo pipefail
27
+
28
+ # symlink logs/latest.out and logs/latest.err
29
+ ln -f -s $SLURM_JOB_ID.out logs/latest.out
30
+ ln -f -s $SLURM_JOB_ID.err logs/latest.err
31
+
32
+ KILL_SWITCH_PATH=kill-switch-$VARIANT
33
+ CHECKPOINT_PATH=$VARIANT_CKPT
34
+ TENSORBOARD_PATH=tensorboard_$VARIANT
35
+
36
+ # Data
37
+ VOCAB_FILE="gpt2/vocab.json"
38
+ MERGE_FILE="gpt2/merges.txt"
39
+ #DATA_PATH="/scratch/project_462000119/data/pile/megatron_data/meg-gpt2_pile_text_document"
40
+ TRAIN_DATA_PATH=train1b5.txt
41
+ # "train: 1.0 0:1 /scratch/project_462000119/data/c4_subsampled/gpt2tok_c4_en_12B_text_document"
42
+ VALID_DATA_PATH=val.txt
43
+ # "validation: 1.0 0:1 /scratch/project_462000119/data/c4_validation/gpt2tok_c4validation_rerun_text_document"
44
+
45
+ PP_SIZE=1
46
+ TP_SIZE=2
47
+
48
+ MICRO_BATCH_SIZE=2
49
+ GRADIENT_ACCUMULATION_STEPS=2
50
+ WORLD_SIZE=$((SLURM_GPUS_ON_NODE*SLURM_JOB_NUM_NODES))
51
+ GLOBAL_BATCH_SIZE=$((MICRO_BATCH_SIZE*WORLD_SIZE*GRADIENT_ACCUMULATION_STEPS))
52
+
53
+ # Model parameters
54
+ source model_params.sh
55
+ MODEL_PARAM=("${PARAM_4516M[@]}")
56
+ NHIDDEN=${MODEL_PARAM[0]}
57
+ FFN_HIDDEN_SIZE=${MODEL_PARAM[1]}
58
+ KV_SIZE=${MODEL_PARAM[2]}
59
+ NHEADS=${MODEL_PARAM[3]}
60
+ NLAYERS=${MODEL_PARAM[4]}
61
+ SEQ_LEN=2048
62
+
63
+ echo "Model parameters: d_model $NHIDDEN ffw_size $FFN_HIDDEN_SIZE kv_size $KV_SIZE n_heads $NHEADS n_layers $NLAYERS"
64
+
65
+ SAVE_INTERVAL=1000
66
+
67
+ # Tokens: 25611230000
68
+ # -> Samples: 12505484
69
+ TRAIN_SAMPLES=1
70
+
71
+ OPTIMIZER_ARGS=" \
72
+ --optimizer adam \
73
+ --adam-beta1 0.9 \
74
+ --adam-beta2 0.999 \
75
+ --adam-eps 1e-8 \
76
+ --lr 2e-4 \
77
+ --min-lr 2e-5 \
78
+ --lr-decay-style cosine \
79
+ --lr-decay-samples $TRAIN_SAMPLES \
80
+ --lr-warmup-samples 0 \
81
+ --clip-grad 1.0 \
82
+ --weight-decay 1e-1 \
83
+ --override-lr-scheduler \
84
+ --reset-progress \
85
+ --no-load-optim \
86
+ "
87
+
88
+ GPT_ARGS=" \
89
+ --num-layers $NLAYERS \
90
+ --hidden-size $NHIDDEN \
91
+ --num-attention-heads $NHEADS \
92
+ --kv-channels $KV_SIZE \
93
+ --ffn-hidden-size $FFN_HIDDEN_SIZE \
94
+ --seq-length $SEQ_LEN \
95
+ --max-position-embeddings $SEQ_LEN \
96
+ --micro-batch-size $MICRO_BATCH_SIZE \
97
+ --global-batch-size $GLOBAL_BATCH_SIZE \
98
+ --train-samples $TRAIN_SAMPLES \
99
+ --vocab-file $VOCAB_FILE \
100
+ --merge-file $MERGE_FILE \
101
+ --clip-grad 1.0 \
102
+ --kill-switch-path $KILL_SWITCH_PATH \
103
+ --bf16 \
104
+ $OPTIMIZER_ARGS \
105
+ "
106
+
107
+ OUTPUT_ARGS=" \
108
+ --log-interval 10 \
109
+ --save-interval $SAVE_INTERVAL \
110
+ --eval-interval 1 \
111
+ --eval-iters 100 \
112
+ --eval-only true \
113
+ --tensorboard-dir $TENSORBOARD_PATH \
114
+ --tensorboard-queue-size 5 \
115
+ --log-timers-to-tensorboard \
116
+ --log-batch-size-to-tensorboard \
117
+ --log-validation-ppl-to-tensorboard \
118
+ "
119
+
120
+ ZERO_STAGE=0
121
+
122
+ mkdir -p ds_configs
123
+ DS_CONFIG_PATH="ds_configs/$SLURM_JOB_ID.json"
124
+
125
+ cat <<EOF > $DS_CONFIG_PATH
126
+ {
127
+ "train_micro_batch_size_per_gpu": $MICRO_BATCH_SIZE,
128
+ "train_batch_size": $GLOBAL_BATCH_SIZE,
129
+ "gradient_clipping": 1.0,
130
+ "zero_optimization": {
131
+ "stage": $ZERO_STAGE
132
+ },
133
+ "bf16": {
134
+ "enabled": true
135
+ },
136
+ "steps_per_print": 2000,
137
+ "wall_clock_breakdown": false
138
+ }
139
+ EOF
140
+
141
+ DEEPSPEED_ARGS=" \
142
+ --deepspeed \
143
+ --deepspeed_config $DS_CONFIG_PATH \
144
+ --zero-stage $ZERO_STAGE \
145
+ "
146
+
147
+ CMD=" \
148
+ Megatron-DeepSpeed/pretrain_gpt.py \
149
+ --tensor-model-parallel-size $TP_SIZE \
150
+ --pipeline-model-parallel-size $PP_SIZE \
151
+ $GPT_ARGS \
152
+ $OUTPUT_ARGS \
153
+ --save $CHECKPOINT_PATH \
154
+ --load $CHECKPOINT_PATH \
155
+ --train-weighted-split-paths-path $TRAIN_DATA_PATH \
156
+ --valid-weighted-split-paths-path $VALID_DATA_PATH \
157
+ --data-impl mmap \
158
+ $DEEPSPEED_ARGS \
159
+ --seed 4 \
160
+ "
161
+
162
+ echo $CMD
163
+
164
+ echo "START $SLURM_JOBID: $(date)"
165
+
166
+ # bash launch_srun_32.sh $CMD
167
+ srun --label launch.sh $CMD
168
+
169
+ echo "END $SLURM_JOBID: $(date)"
4b284b6bc4seed3/tensorboard_4b284b6bc4seed3val/events.out.tfevents.1683631386.nid005684.108633.0 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1860572eb4d04275e354928309637db4fdb92b66d450eb1e2f2a5fff301a0e60
3
+ size 980
4b284b6bc4seed3/tensorboard_4b284b6bc4seed4val/events.out.tfevents.1683632942.nid005245.117796.0 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:dbb484665f35fc7e7dda8608d14b81744e886cb6c2b30e59b569b685fbf0ed4a
3
+ size 980
4b284b6bc4seed4/3484368.err ADDED
The diff for this file is too large to render. See raw diff
 
4b284b6bc4seed4/3484368.out ADDED
The diff for this file is too large to render. See raw diff