Hugging Face's logo

datablations
/
lm5-2b8-55b-c4

TensorBoard
Model card Files Metrics Community
lm5-2b8-55b-c4
File size: 4,254 Bytes 
b757584
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
 
#!/bin/bash
#SBATCH --exclude=nid005159
#SBATCH --nodes=1
#SBATCH --ntasks-per-node=1
#SBATCH --cpus-per-task=32
#SBATCH --mem=256G
#SBATCH -p small-g
#SBATCH -t 2-0:00:00
#SBATCH --gpus-per-node=mi250:1
#SBATCH --exclusive=user
#SBATCH --hint=nomultithread
#SBATCH --account=project_462000119
#SBATCH -o logs/%j.out
#SBATCH -e logs/%j.err

# if run without sbatch, invoke here
if [ -z $SLURM_JOB_ID ]; then
    mkdir -p logs
    sbatch "$0"
    exit
fi

set -euo pipefail

# symlink logs/latest_eval.out and logs/latest_eval.err
ln -f -s $SLURM_JOB_ID.out logs/latest_eval.out
ln -f -s $SLURM_JOB_ID.err logs/latest_eval.err

# Data
#CHECKPOINT_PATH=/scratch/project_462000119/muennighoff/nov-2022-optimization/checkpoints/global_step10
#VARIANT=global_step10

CHECKPOINT_PATH=lm1-220m/global_step14324
VARIANT=lm1-220m
CHECKPOINT_PATH=lm1-220m-7b5-oscar/global_step14324
VARIANT=lm1-220m-7b5-oscar

#CHECKPOINT_PATH=lm1-280m/global_step11269
#VARIANT=lm1-280m-5b9
#CHECKPOINT_PATH=lm1-280m-5b9-oscar/global_step11269
#VARIANT=lm1-280m-5b9-oscar

CHECKPOINT_PATH=lm1-1b1-21b-oscar/global_step39672
VARIANT=lm1-1b1-21b-oscar
#CHECKPOINT_PATH=lm1-1b1-21b/global_step39672
#VARIANT=lm1-1b1-21b

#CHECKPOINT_PATH=lm1-2b8-55b-oscar/global_step52452
#VARIANT=lm1-2b8-55b-oscar
#CHECKPOINT_PATH=lm1-2b8-55b/global_step52452
#VARIANT=lm1-2b8-55b

#CHECKPOINT_PATH=lm1-2b8-55b-oscar/global_step52452
#VARIANT=lm1-2b8-55b-oscar
#CHECKPOINT_PATH=lm1-3b9-77b/global_step73814
#VARIANT=lm1-3b9-77b

#CHECKPOINT_PATH=lm1-1b1-21b-c4/global_step39672
#VARIANT=lm1-1b1-21b-c4


# tensorboard_2b855b11bc4  tensorboard_2b855b14bc4  tensorboard_2b855b18bc4  tensorboard_2b855b28bc4  tensorboard_2b855b9bc4
#2b855b50c4py  2b855b60c4py  2b855b70c4py  2b855b80c4py  2b855b90c4py
VARIANT=2b855b70c4py
CHECKPOINT_PATH=lm1-2b8-55b-c4py/$VARIANT/global_step52452
#2b855b11bc4  2b855b14bc4  2b855b18bc4  2b855b28bc4  2b855b9bc4
#VARIANT=2b855b9boscar
#CHECKPOINT_PATH=lm1-2b8-55b-oscar-repetitions/$VARIANT/global_step52452

#VARIANT=realtasky
#CHECKPOINT_PATH=checkpoints_2b855brealtasky/global_step52452
#CHECKPOINT_PATH=lm1-2b8-55b-c4-repetitions/2b855b55bc4/global_step52452


CHECKPOINT_PATH=checkpoints_2b855b55bc4ul2valfast/global_step52452
VARIANT=ul2valfast

CHECKPOINT_PATH=lm2-2b8-55b-c4-new/global_step52452
VARIANT=ul2new

export HF_DATASETS_OFFLINE=1
export HF_DATASETS_CACHE=/scratch/project_462000119/ds_cache

VOCAB_FILE="gpt2/vocab.json"
MERGE_FILE="gpt2/merges.txt"

PP_SIZE=1
TP_SIZE=1
# different from the training MICRO_BATCH_SIZE - no optim memory, so can do bigger BS
# make as big as it can fit into gpu w/o OOM, but not too close to 100%
EVAL_MICRO_BATCH_SIZE=1
MICRO_BS_MULTIPLIER=1

# Model parameters
SEQ_LEN=2048

# Dummy arguments
MEGATRON_REQUIRED_ARGS=" \
    --num-layers -1 \
    --hidden-size -1 \
    --num-attention-heads -1 \
    --seq-length -1  \
    --max-position-embeddings -1 \
"

ZERO_STAGE=0

mkdir -p ds_configs
DS_CONFIG_PATH="ds_configs/$SLURM_JOB_ID.json"

cat <<EOF > $DS_CONFIG_PATH
{
    "train_micro_batch_size_per_gpu": 1,
    "train_batch_size": 1,
    "gradient_clipping": 1.0,
    "zero_optimization": {
        "stage": $ZERO_STAGE
    },
    "bf16": {
        "enabled": true
    },
    "steps_per_print": 2000,
    "wall_clock_breakdown": false
}
EOF

DEEPSPEED_ARGS=" \
    --deepspeed \
    --deepspeed_config $DS_CONFIG_PATH \
    --zero-stage $ZERO_STAGE \
    "

CMD="Megatron-DeepSpeed/tasks/eval_harness/evaluate.py \
    --load $CHECKPOINT_PATH \
    --results_path $VARIANT-results.json \
    --tensor-model-parallel-size $TP_SIZE  \
    --pipeline-model-parallel-size $PP_SIZE \
    --vocab-file $VOCAB_FILE \
    --merge-file $MERGE_FILE \
    --micro-batch-size $EVAL_MICRO_BATCH_SIZE \
    --no-load-optim \
    --no-load-rng \
    --bf16 \
    --inference \
    --seq-length $SEQ_LEN \
    --task_list anli_r1,anli_r2,anli_r3,cb,copa,hellaswag,rte,winogrande,storycloze_2016,boolq,arc_easy,arc_challenge,sciq,piqa \
    --intermed_results \
    --adaptive_seq_len \
    --add_denoiser \
    --micro_bs_multiplier $MICRO_BS_MULTIPLIER \
    $MEGATRON_REQUIRED_ARGS \
    $DEEPSPEED_ARGS \
    "

echo $CMD

echo "START $SLURM_JOBID: $(date)"

srun --label launch.sh $CMD

echo "END $SLURM_JOBID: $(date)"