Add

4b284b12bc4seed1/latest

@@ -0,0 +1 @@
+global_step80108

4b284b12bc4seed1/sbatch_4b284b12bseed1val.sh

@@ -0,0 +1,169 @@
+#!/bin/bash
+#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
+#SBATCH --nodes=32
+#SBATCH --ntasks-per-node=1
+#SBATCH --cpus-per-task=40
+#SBATCH --mem=256G
+#SBATCH -p standard-g
+#SBATCH -t 48:00:00
+#SBATCH --gpus-per-node=mi250:8
+#SBATCH --exclusive=user
+#SBATCH --hint=nomultithread
+#SBATCH --account=project_462000119
+#SBATCH -o logs/%j.out
+#SBATCH -e logs/%j.err
+
+VARIANT=4b284b12bc4seed1val
+VARIANT_CKPT=lm1-4b2-84b-c4seeds/4b284b12bc4seed1
+
+# 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.out and logs/latest.err
+ln -f -s $SLURM_JOB_ID.out logs/latest.out
+ln -f -s $SLURM_JOB_ID.err logs/latest.err
+
+KILL_SWITCH_PATH=kill-switch-$VARIANT
+CHECKPOINT_PATH=$VARIANT_CKPT
+TENSORBOARD_PATH=tensorboard_$VARIANT
+
+# Data
+VOCAB_FILE="gpt2/vocab.json"
+MERGE_FILE="gpt2/merges.txt"
+#DATA_PATH="/scratch/project_462000119/data/pile/megatron_data/meg-gpt2_pile_text_document"
+TRAIN_DATA_PATH=train1b5.txt
+# "train: 1.0 0:1 /scratch/project_462000119/data/c4_subsampled/gpt2tok_c4_en_12B_text_document"
+VALID_DATA_PATH=val.txt
+# "validation: 1.0 0:1 /scratch/project_462000119/data/c4_validation/gpt2tok_c4validation_rerun_text_document"
+
+PP_SIZE=1
+TP_SIZE=2
+
+MICRO_BATCH_SIZE=2
+GRADIENT_ACCUMULATION_STEPS=2
+WORLD_SIZE=$((SLURM_GPUS_ON_NODE*SLURM_JOB_NUM_NODES))
+GLOBAL_BATCH_SIZE=$((MICRO_BATCH_SIZE*WORLD_SIZE*GRADIENT_ACCUMULATION_STEPS))
+
+# Model parameters
+source model_params.sh
+MODEL_PARAM=("${PARAM_4516M[@]}")
+NHIDDEN=${MODEL_PARAM[0]}
+FFN_HIDDEN_SIZE=${MODEL_PARAM[1]}
+KV_SIZE=${MODEL_PARAM[2]}
+NHEADS=${MODEL_PARAM[3]}
+NLAYERS=${MODEL_PARAM[4]}
+SEQ_LEN=2048
+
+echo "Model parameters: d_model $NHIDDEN ffw_size $FFN_HIDDEN_SIZE kv_size $KV_SIZE n_heads $NHEADS n_layers $NLAYERS"
+
+SAVE_INTERVAL=1000
+
+# Tokens: 25611230000
+# -> Samples: 12505484
+TRAIN_SAMPLES=1
+
+OPTIMIZER_ARGS=" \
+    --optimizer adam \
+    --adam-beta1 0.9 \
+    --adam-beta2 0.999 \
+    --adam-eps 1e-8 \
+    --lr 2e-4 \
+    --min-lr 2e-5 \
+    --lr-decay-style cosine \
+    --lr-decay-samples $TRAIN_SAMPLES \
+    --lr-warmup-samples 0 \
+    --clip-grad 1.0 \
+    --weight-decay 1e-1 \
+    --override-lr-scheduler \
+    --reset-progress \
+    --no-load-optim \
+    "
+
+GPT_ARGS=" \
+    --num-layers $NLAYERS \
+    --hidden-size $NHIDDEN \
+    --num-attention-heads $NHEADS \
+    --kv-channels $KV_SIZE \
+    --ffn-hidden-size $FFN_HIDDEN_SIZE \
+    --seq-length $SEQ_LEN \
+    --max-position-embeddings $SEQ_LEN \
+    --micro-batch-size $MICRO_BATCH_SIZE \
+    --global-batch-size $GLOBAL_BATCH_SIZE \
+    --train-samples $TRAIN_SAMPLES \
+    --vocab-file $VOCAB_FILE \
+    --merge-file $MERGE_FILE \
+    --clip-grad 1.0 \
+    --kill-switch-path $KILL_SWITCH_PATH \
+    --bf16 \
+    $OPTIMIZER_ARGS \
+    "
+
+OUTPUT_ARGS=" \
+    --log-interval 10 \
+    --save-interval $SAVE_INTERVAL \
+    --eval-interval 1 \
+    --eval-iters 100 \
+    --eval-only true \
+    --tensorboard-dir $TENSORBOARD_PATH \
+    --tensorboard-queue-size 5 \
+    --log-timers-to-tensorboard \
+    --log-batch-size-to-tensorboard \
+    --log-validation-ppl-to-tensorboard \
+    "
+
+ZERO_STAGE=0
+
+mkdir -p ds_configs
+DS_CONFIG_PATH="ds_configs/$SLURM_JOB_ID.json"
+
+cat <<EOF > $DS_CONFIG_PATH
+{
+    "train_micro_batch_size_per_gpu": $MICRO_BATCH_SIZE,
+    "train_batch_size": $GLOBAL_BATCH_SIZE,
+    "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/pretrain_gpt.py \
+    --tensor-model-parallel-size $TP_SIZE \
+    --pipeline-model-parallel-size $PP_SIZE \
+    $GPT_ARGS \
+    $OUTPUT_ARGS \
+    --save $CHECKPOINT_PATH \
+    --load $CHECKPOINT_PATH \
+    --train-weighted-split-paths-path $TRAIN_DATA_PATH \
+    --valid-weighted-split-paths-path $VALID_DATA_PATH \
+    --data-impl mmap \
+     $DEEPSPEED_ARGS \
+    --seed 1 \
+    "
+
+echo $CMD
+
+echo "START $SLURM_JOBID: $(date)"
+
+# bash launch_srun_32.sh $CMD
+srun --label launch.sh $CMD
+
+echo "END $SLURM_JOBID: $(date)"

4b284b12bc4seed1/tensorboard_4b284b12bc4seed1val/events.out.tfevents.1683551547.nid005684.5347.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d5f4eef1b0b38a28b547a29951bf5f6f12ff41af507c6a1266e4067da5f69d09
+size 40

4b284b12bc4seed1/tensorboard_4b284b12bc4seed1val/events.out.tfevents.1683577102.nid007109.10132.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:469a8f524a071ade99de6ba915db23130873de8e12af4eea445c255771bf94e3
+size 980

4b284b12bc4seed2/latest

@@ -0,0 +1 @@
+global_step80108

4b284b12bc4seed2/sbatch_4b284b12bc4seed2val.sh

@@ -0,0 +1,169 @@
+#!/bin/bash
+#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
+#SBATCH --nodes=32
+#SBATCH --ntasks-per-node=1
+#SBATCH --cpus-per-task=40
+#SBATCH --mem=256G
+#SBATCH -p standard-g
+#SBATCH -t 48:00:00
+#SBATCH --gpus-per-node=mi250:8
+#SBATCH --exclusive=user
+#SBATCH --hint=nomultithread
+#SBATCH --account=project_462000119
+#SBATCH -o logs/%j.out
+#SBATCH -e logs/%j.err
+
+VARIANT=4b284b12bc4seed2val
+VARIANT_CKPT=lm1-4b2-84b-c4seeds/4b284b12bc4seed2
+
+# 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.out and logs/latest.err
+ln -f -s $SLURM_JOB_ID.out logs/latest.out
+ln -f -s $SLURM_JOB_ID.err logs/latest.err
+
+KILL_SWITCH_PATH=kill-switch-$VARIANT
+CHECKPOINT_PATH=$VARIANT_CKPT
+TENSORBOARD_PATH=tensorboard_$VARIANT
+
+# Data
+VOCAB_FILE="gpt2/vocab.json"
+MERGE_FILE="gpt2/merges.txt"
+#DATA_PATH="/scratch/project_462000119/data/pile/megatron_data/meg-gpt2_pile_text_document"
+TRAIN_DATA_PATH=train1b5.txt
+# "train: 1.0 0:1 /scratch/project_462000119/data/c4_subsampled/gpt2tok_c4_en_12B_text_document"
+VALID_DATA_PATH=val.txt
+# "validation: 1.0 0:1 /scratch/project_462000119/data/c4_validation/gpt2tok_c4validation_rerun_text_document"
+
+PP_SIZE=1
+TP_SIZE=2
+
+MICRO_BATCH_SIZE=2
+GRADIENT_ACCUMULATION_STEPS=2
+WORLD_SIZE=$((SLURM_GPUS_ON_NODE*SLURM_JOB_NUM_NODES))
+GLOBAL_BATCH_SIZE=$((MICRO_BATCH_SIZE*WORLD_SIZE*GRADIENT_ACCUMULATION_STEPS))
+
+# Model parameters
+source model_params.sh
+MODEL_PARAM=("${PARAM_4516M[@]}")
+NHIDDEN=${MODEL_PARAM[0]}
+FFN_HIDDEN_SIZE=${MODEL_PARAM[1]}
+KV_SIZE=${MODEL_PARAM[2]}
+NHEADS=${MODEL_PARAM[3]}
+NLAYERS=${MODEL_PARAM[4]}
+SEQ_LEN=2048
+
+echo "Model parameters: d_model $NHIDDEN ffw_size $FFN_HIDDEN_SIZE kv_size $KV_SIZE n_heads $NHEADS n_layers $NLAYERS"
+
+SAVE_INTERVAL=1000
+
+# Tokens: 25611230000
+# -> Samples: 12505484
+TRAIN_SAMPLES=1
+
+OPTIMIZER_ARGS=" \
+    --optimizer adam \
+    --adam-beta1 0.9 \
+    --adam-beta2 0.999 \
+    --adam-eps 1e-8 \
+    --lr 2e-4 \
+    --min-lr 2e-5 \
+    --lr-decay-style cosine \
+    --lr-decay-samples $TRAIN_SAMPLES \
+    --lr-warmup-samples 0 \
+    --clip-grad 1.0 \
+    --weight-decay 1e-1 \
+    --override-lr-scheduler \
+    --reset-progress \
+    --no-load-optim \
+    "
+
+GPT_ARGS=" \
+    --num-layers $NLAYERS \
+    --hidden-size $NHIDDEN \
+    --num-attention-heads $NHEADS \
+    --kv-channels $KV_SIZE \
+    --ffn-hidden-size $FFN_HIDDEN_SIZE \
+    --seq-length $SEQ_LEN \
+    --max-position-embeddings $SEQ_LEN \
+    --micro-batch-size $MICRO_BATCH_SIZE \
+    --global-batch-size $GLOBAL_BATCH_SIZE \
+    --train-samples $TRAIN_SAMPLES \
+    --vocab-file $VOCAB_FILE \
+    --merge-file $MERGE_FILE \
+    --clip-grad 1.0 \
+    --kill-switch-path $KILL_SWITCH_PATH \
+    --bf16 \
+    $OPTIMIZER_ARGS \
+    "
+
+OUTPUT_ARGS=" \
+    --log-interval 10 \
+    --save-interval $SAVE_INTERVAL \
+    --eval-interval 1 \
+    --eval-iters 100 \
+    --eval-only true \
+    --tensorboard-dir $TENSORBOARD_PATH \
+    --tensorboard-queue-size 5 \
+    --log-timers-to-tensorboard \
+    --log-batch-size-to-tensorboard \
+    --log-validation-ppl-to-tensorboard \
+    "
+
+ZERO_STAGE=0
+
+mkdir -p ds_configs
+DS_CONFIG_PATH="ds_configs/$SLURM_JOB_ID.json"
+
+cat <<EOF > $DS_CONFIG_PATH
+{
+    "train_micro_batch_size_per_gpu": $MICRO_BATCH_SIZE,
+    "train_batch_size": $GLOBAL_BATCH_SIZE,
+    "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/pretrain_gpt.py \
+    --tensor-model-parallel-size $TP_SIZE \
+    --pipeline-model-parallel-size $PP_SIZE \
+    $GPT_ARGS \
+    $OUTPUT_ARGS \
+    --save $CHECKPOINT_PATH \
+    --load $CHECKPOINT_PATH \
+    --train-weighted-split-paths-path $TRAIN_DATA_PATH \
+    --valid-weighted-split-paths-path $VALID_DATA_PATH \
+    --data-impl mmap \
+     $DEEPSPEED_ARGS \
+    --seed 2 \
+    "
+
+echo $CMD
+
+echo "START $SLURM_JOBID: $(date)"
+
+# bash launch_srun_32.sh $CMD
+srun --label launch.sh $CMD
+
+echo "END $SLURM_JOBID: $(date)"

4b284b12bc4seed2/tensorboard_4b284b12bc4seed2val/events.out.tfevents.1683577715.nid007109.19228.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7673810ad0ea4fa428bcb8c04a951ec214355401eb9881aa780c4a42542920c6
+size 980

4b284b12bc4seed3/latest

@@ -0,0 +1 @@
+global_step80108

4b284b12bc4seed3/sbatch_4b284b12bc4seed3val.sh

@@ -0,0 +1,169 @@
+#!/bin/bash
+#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
+#SBATCH --nodes=32
+#SBATCH --ntasks-per-node=1
+#SBATCH --cpus-per-task=40
+#SBATCH --mem=256G
+#SBATCH -p standard-g
+#SBATCH -t 48:00:00
+#SBATCH --gpus-per-node=mi250:8
+#SBATCH --exclusive=user
+#SBATCH --hint=nomultithread
+#SBATCH --account=project_462000119
+#SBATCH -o logs/%j.out
+#SBATCH -e logs/%j.err
+
+VARIANT=4b284b12bc4seed3val
+VARIANT_CKPT=lm1-4b2-84b-c4seeds/4b284b12bc4seed3
+
+# 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.out and logs/latest.err
+ln -f -s $SLURM_JOB_ID.out logs/latest.out
+ln -f -s $SLURM_JOB_ID.err logs/latest.err
+
+KILL_SWITCH_PATH=kill-switch-$VARIANT
+CHECKPOINT_PATH=$VARIANT_CKPT
+TENSORBOARD_PATH=tensorboard_$VARIANT
+
+# Data
+VOCAB_FILE="gpt2/vocab.json"
+MERGE_FILE="gpt2/merges.txt"
+#DATA_PATH="/scratch/project_462000119/data/pile/megatron_data/meg-gpt2_pile_text_document"
+TRAIN_DATA_PATH=train1b5.txt
+# "train: 1.0 0:1 /scratch/project_462000119/data/c4_subsampled/gpt2tok_c4_en_12B_text_document"
+VALID_DATA_PATH=val.txt
+# "validation: 1.0 0:1 /scratch/project_462000119/data/c4_validation/gpt2tok_c4validation_rerun_text_document"
+
+PP_SIZE=1
+TP_SIZE=2
+
+MICRO_BATCH_SIZE=2
+GRADIENT_ACCUMULATION_STEPS=2
+WORLD_SIZE=$((SLURM_GPUS_ON_NODE*SLURM_JOB_NUM_NODES))
+GLOBAL_BATCH_SIZE=$((MICRO_BATCH_SIZE*WORLD_SIZE*GRADIENT_ACCUMULATION_STEPS))
+
+# Model parameters
+source model_params.sh
+MODEL_PARAM=("${PARAM_4516M[@]}")
+NHIDDEN=${MODEL_PARAM[0]}
+FFN_HIDDEN_SIZE=${MODEL_PARAM[1]}
+KV_SIZE=${MODEL_PARAM[2]}
+NHEADS=${MODEL_PARAM[3]}
+NLAYERS=${MODEL_PARAM[4]}
+SEQ_LEN=2048
+
+echo "Model parameters: d_model $NHIDDEN ffw_size $FFN_HIDDEN_SIZE kv_size $KV_SIZE n_heads $NHEADS n_layers $NLAYERS"
+
+SAVE_INTERVAL=1000
+
+# Tokens: 25611230000
+# -> Samples: 12505484
+TRAIN_SAMPLES=1
+
+OPTIMIZER_ARGS=" \
+    --optimizer adam \
+    --adam-beta1 0.9 \
+    --adam-beta2 0.999 \
+    --adam-eps 1e-8 \
+    --lr 2e-4 \
+    --min-lr 2e-5 \
+    --lr-decay-style cosine \
+    --lr-decay-samples $TRAIN_SAMPLES \
+    --lr-warmup-samples 0 \
+    --clip-grad 1.0 \
+    --weight-decay 1e-1 \
+    --override-lr-scheduler \
+    --reset-progress \
+    --no-load-optim \
+    "
+
+GPT_ARGS=" \
+    --num-layers $NLAYERS \
+    --hidden-size $NHIDDEN \
+    --num-attention-heads $NHEADS \
+    --kv-channels $KV_SIZE \
+    --ffn-hidden-size $FFN_HIDDEN_SIZE \
+    --seq-length $SEQ_LEN \
+    --max-position-embeddings $SEQ_LEN \
+    --micro-batch-size $MICRO_BATCH_SIZE \
+    --global-batch-size $GLOBAL_BATCH_SIZE \
+    --train-samples $TRAIN_SAMPLES \
+    --vocab-file $VOCAB_FILE \
+    --merge-file $MERGE_FILE \
+    --clip-grad 1.0 \
+    --kill-switch-path $KILL_SWITCH_PATH \
+    --bf16 \
+    $OPTIMIZER_ARGS \
+    "
+
+OUTPUT_ARGS=" \
+    --log-interval 10 \
+    --save-interval $SAVE_INTERVAL \
+    --eval-interval 1 \
+    --eval-iters 100 \
+    --eval-only true \
+    --tensorboard-dir $TENSORBOARD_PATH \
+    --tensorboard-queue-size 5 \
+    --log-timers-to-tensorboard \
+    --log-batch-size-to-tensorboard \
+    --log-validation-ppl-to-tensorboard \
+    "
+
+ZERO_STAGE=0
+
+mkdir -p ds_configs
+DS_CONFIG_PATH="ds_configs/$SLURM_JOB_ID.json"
+
+cat <<EOF > $DS_CONFIG_PATH
+{
+    "train_micro_batch_size_per_gpu": $MICRO_BATCH_SIZE,
+    "train_batch_size": $GLOBAL_BATCH_SIZE,
+    "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/pretrain_gpt.py \
+    --tensor-model-parallel-size $TP_SIZE \
+    --pipeline-model-parallel-size $PP_SIZE \
+    $GPT_ARGS \
+    $OUTPUT_ARGS \
+    --save $CHECKPOINT_PATH \
+    --load $CHECKPOINT_PATH \
+    --train-weighted-split-paths-path $TRAIN_DATA_PATH \
+    --valid-weighted-split-paths-path $VALID_DATA_PATH \
+    --data-impl mmap \
+     $DEEPSPEED_ARGS \
+    --seed 3 \
+    "
+
+echo $CMD
+
+echo "START $SLURM_JOBID: $(date)"
+
+# bash launch_srun_32.sh $CMD
+srun --label launch.sh $CMD
+
+echo "END $SLURM_JOBID: $(date)"

4b284b12bc4seed3/tensorboard_4b284b12bc4seed3val/events.out.tfevents.1683577715.nid006149.14735.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:03ef1e06adf22e1a14e97f46b8c68019de7843f099a8264d27467276b2df49a5
+size 980

4b284b12bc4seed4/latest

@@ -0,0 +1 @@
+global_step80108

4b284b12bc4seed4/sbatch_4b284b12bc4seed4val.sh

@@ -0,0 +1,169 @@
+#!/bin/bash
+#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
+#SBATCH --nodes=32
+#SBATCH --ntasks-per-node=1
+#SBATCH --cpus-per-task=40
+#SBATCH --mem=256G
+#SBATCH -p standard-g
+#SBATCH -t 48:00:00
+#SBATCH --gpus-per-node=mi250:8
+#SBATCH --exclusive=user
+#SBATCH --hint=nomultithread
+#SBATCH --account=project_462000119
+#SBATCH -o logs/%j.out
+#SBATCH -e logs/%j.err
+
+VARIANT=4b284b12bc4seed4val
+VARIANT_CKPT=lm1-4b2-84b-c4seeds/4b284b12bc4seed4
+
+# 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.out and logs/latest.err
+ln -f -s $SLURM_JOB_ID.out logs/latest.out
+ln -f -s $SLURM_JOB_ID.err logs/latest.err
+
+KILL_SWITCH_PATH=kill-switch-$VARIANT
+CHECKPOINT_PATH=$VARIANT_CKPT
+TENSORBOARD_PATH=tensorboard_$VARIANT
+
+# Data
+VOCAB_FILE="gpt2/vocab.json"
+MERGE_FILE="gpt2/merges.txt"
+#DATA_PATH="/scratch/project_462000119/data/pile/megatron_data/meg-gpt2_pile_text_document"
+TRAIN_DATA_PATH=train1b5.txt
+# "train: 1.0 0:1 /scratch/project_462000119/data/c4_subsampled/gpt2tok_c4_en_12B_text_document"
+VALID_DATA_PATH=val.txt
+# "validation: 1.0 0:1 /scratch/project_462000119/data/c4_validation/gpt2tok_c4validation_rerun_text_document"
+
+PP_SIZE=1
+TP_SIZE=2
+
+MICRO_BATCH_SIZE=2
+GRADIENT_ACCUMULATION_STEPS=2
+WORLD_SIZE=$((SLURM_GPUS_ON_NODE*SLURM_JOB_NUM_NODES))
+GLOBAL_BATCH_SIZE=$((MICRO_BATCH_SIZE*WORLD_SIZE*GRADIENT_ACCUMULATION_STEPS))
+
+# Model parameters
+source model_params.sh
+MODEL_PARAM=("${PARAM_4516M[@]}")
+NHIDDEN=${MODEL_PARAM[0]}
+FFN_HIDDEN_SIZE=${MODEL_PARAM[1]}
+KV_SIZE=${MODEL_PARAM[2]}
+NHEADS=${MODEL_PARAM[3]}
+NLAYERS=${MODEL_PARAM[4]}
+SEQ_LEN=2048
+
+echo "Model parameters: d_model $NHIDDEN ffw_size $FFN_HIDDEN_SIZE kv_size $KV_SIZE n_heads $NHEADS n_layers $NLAYERS"
+
+SAVE_INTERVAL=1000
+
+# Tokens: 25611230000
+# -> Samples: 12505484
+TRAIN_SAMPLES=1
+
+OPTIMIZER_ARGS=" \
+    --optimizer adam \
+    --adam-beta1 0.9 \
+    --adam-beta2 0.999 \
+    --adam-eps 1e-8 \
+    --lr 2e-4 \
+    --min-lr 2e-5 \
+    --lr-decay-style cosine \
+    --lr-decay-samples $TRAIN_SAMPLES \
+    --lr-warmup-samples 0 \
+    --clip-grad 1.0 \
+    --weight-decay 1e-1 \
+    --override-lr-scheduler \
+    --reset-progress \
+    --no-load-optim \
+    "
+
+GPT_ARGS=" \
+    --num-layers $NLAYERS \
+    --hidden-size $NHIDDEN \
+    --num-attention-heads $NHEADS \
+    --kv-channels $KV_SIZE \
+    --ffn-hidden-size $FFN_HIDDEN_SIZE \
+    --seq-length $SEQ_LEN \
+    --max-position-embeddings $SEQ_LEN \
+    --micro-batch-size $MICRO_BATCH_SIZE \
+    --global-batch-size $GLOBAL_BATCH_SIZE \
+    --train-samples $TRAIN_SAMPLES \
+    --vocab-file $VOCAB_FILE \
+    --merge-file $MERGE_FILE \
+    --clip-grad 1.0 \
+    --kill-switch-path $KILL_SWITCH_PATH \
+    --bf16 \
+    $OPTIMIZER_ARGS \
+    "
+
+OUTPUT_ARGS=" \
+    --log-interval 10 \
+    --save-interval $SAVE_INTERVAL \
+    --eval-interval 1 \
+    --eval-iters 100 \
+    --eval-only true \
+    --tensorboard-dir $TENSORBOARD_PATH \
+    --tensorboard-queue-size 5 \
+    --log-timers-to-tensorboard \
+    --log-batch-size-to-tensorboard \
+    --log-validation-ppl-to-tensorboard \
+    "
+
+ZERO_STAGE=0
+
+mkdir -p ds_configs
+DS_CONFIG_PATH="ds_configs/$SLURM_JOB_ID.json"
+
+cat <<EOF > $DS_CONFIG_PATH
+{
+    "train_micro_batch_size_per_gpu": $MICRO_BATCH_SIZE,
+    "train_batch_size": $GLOBAL_BATCH_SIZE,
+    "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/pretrain_gpt.py \
+    --tensor-model-parallel-size $TP_SIZE \
+    --pipeline-model-parallel-size $PP_SIZE \
+    $GPT_ARGS \
+    $OUTPUT_ARGS \
+    --save $CHECKPOINT_PATH \
+    --load $CHECKPOINT_PATH \
+    --train-weighted-split-paths-path $TRAIN_DATA_PATH \
+    --valid-weighted-split-paths-path $VALID_DATA_PATH \
+    --data-impl mmap \
+     $DEEPSPEED_ARGS \
+    --seed 4 \
+    "
+
+echo $CMD
+
+echo "START $SLURM_JOBID: $(date)"
+
+# bash launch_srun_32.sh $CMD
+srun --label launch.sh $CMD
+
+echo "END $SLURM_JOBID: $(date)"

4b284b12bc4seed4/tensorboard_4b284b12bc4seed4val/events.out.tfevents.1683577715.nid006413.10980.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:78e76f2fef9cb661a1204967db823e4a32c736fe986b44e66d9ee4687cfc8ab5
+size 980

4b284b17bc4seed1/latest

@@ -0,0 +1 @@
+global_step80108

4b284b17bc4seed1/sbatch_4b284b17bc4seed1val.sh

@@ -0,0 +1,169 @@
+#!/bin/bash
+#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
+#SBATCH --nodes=32
+#SBATCH --ntasks-per-node=1
+#SBATCH --cpus-per-task=40
+#SBATCH --mem=256G
+#SBATCH -p standard-g
+#SBATCH -t 48:00:00
+#SBATCH --gpus-per-node=mi250:8
+#SBATCH --exclusive=user
+#SBATCH --hint=nomultithread
+#SBATCH --account=project_462000119
+#SBATCH -o logs/%j.out
+#SBATCH -e logs/%j.err
+
+VARIANT=4b284b17bc4seed1val
+VARIANT_CKPT=lm1-4b2-84b-c4seeds/4b284b17bc4seed1
+
+# 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.out and logs/latest.err
+ln -f -s $SLURM_JOB_ID.out logs/latest.out
+ln -f -s $SLURM_JOB_ID.err logs/latest.err
+
+KILL_SWITCH_PATH=kill-switch-$VARIANT
+CHECKPOINT_PATH=$VARIANT_CKPT
+TENSORBOARD_PATH=tensorboard_$VARIANT
+
+# Data
+VOCAB_FILE="gpt2/vocab.json"
+MERGE_FILE="gpt2/merges.txt"
+#DATA_PATH="/scratch/project_462000119/data/pile/megatron_data/meg-gpt2_pile_text_document"
+TRAIN_DATA_PATH=train1b5.txt
+# "train: 1.0 0:1 /scratch/project_462000119/data/c4_subsampled/gpt2tok_c4_en_12B_text_document"
+VALID_DATA_PATH=val.txt
+# "validation: 1.0 0:1 /scratch/project_462000119/data/c4_validation/gpt2tok_c4validation_rerun_text_document"
+
+PP_SIZE=1
+TP_SIZE=2
+
+MICRO_BATCH_SIZE=2
+GRADIENT_ACCUMULATION_STEPS=2
+WORLD_SIZE=$((SLURM_GPUS_ON_NODE*SLURM_JOB_NUM_NODES))
+GLOBAL_BATCH_SIZE=$((MICRO_BATCH_SIZE*WORLD_SIZE*GRADIENT_ACCUMULATION_STEPS))
+
+# Model parameters
+source model_params.sh
+MODEL_PARAM=("${PARAM_4516M[@]}")
+NHIDDEN=${MODEL_PARAM[0]}
+FFN_HIDDEN_SIZE=${MODEL_PARAM[1]}
+KV_SIZE=${MODEL_PARAM[2]}
+NHEADS=${MODEL_PARAM[3]}
+NLAYERS=${MODEL_PARAM[4]}
+SEQ_LEN=2048
+
+echo "Model parameters: d_model $NHIDDEN ffw_size $FFN_HIDDEN_SIZE kv_size $KV_SIZE n_heads $NHEADS n_layers $NLAYERS"
+
+SAVE_INTERVAL=1000
+
+# Tokens: 25611230000
+# -> Samples: 12505484
+TRAIN_SAMPLES=1
+
+OPTIMIZER_ARGS=" \
+    --optimizer adam \
+    --adam-beta1 0.9 \
+    --adam-beta2 0.999 \
+    --adam-eps 1e-8 \
+    --lr 2e-4 \
+    --min-lr 2e-5 \
+    --lr-decay-style cosine \
+    --lr-decay-samples $TRAIN_SAMPLES \
+    --lr-warmup-samples 0 \
+    --clip-grad 1.0 \
+    --weight-decay 1e-1 \
+    --override-lr-scheduler \
+    --reset-progress \
+    --no-load-optim \
+    "
+
+GPT_ARGS=" \
+    --num-layers $NLAYERS \
+    --hidden-size $NHIDDEN \
+    --num-attention-heads $NHEADS \
+    --kv-channels $KV_SIZE \
+    --ffn-hidden-size $FFN_HIDDEN_SIZE \
+    --seq-length $SEQ_LEN \
+    --max-position-embeddings $SEQ_LEN \
+    --micro-batch-size $MICRO_BATCH_SIZE \
+    --global-batch-size $GLOBAL_BATCH_SIZE \
+    --train-samples $TRAIN_SAMPLES \
+    --vocab-file $VOCAB_FILE \
+    --merge-file $MERGE_FILE \
+    --clip-grad 1.0 \
+    --kill-switch-path $KILL_SWITCH_PATH \
+    --bf16 \
+    $OPTIMIZER_ARGS \
+    "
+
+OUTPUT_ARGS=" \
+    --log-interval 10 \
+    --save-interval $SAVE_INTERVAL \
+    --eval-interval 1 \
+    --eval-iters 100 \
+    --eval-only true \
+    --tensorboard-dir $TENSORBOARD_PATH \
+    --tensorboard-queue-size 5 \
+    --log-timers-to-tensorboard \
+    --log-batch-size-to-tensorboard \
+    --log-validation-ppl-to-tensorboard \
+    "
+
+ZERO_STAGE=0
+
+mkdir -p ds_configs
+DS_CONFIG_PATH="ds_configs/$SLURM_JOB_ID.json"
+
+cat <<EOF > $DS_CONFIG_PATH
+{
+    "train_micro_batch_size_per_gpu": $MICRO_BATCH_SIZE,
+    "train_batch_size": $GLOBAL_BATCH_SIZE,
+    "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/pretrain_gpt.py \
+    --tensor-model-parallel-size $TP_SIZE \
+    --pipeline-model-parallel-size $PP_SIZE \
+    $GPT_ARGS \
+    $OUTPUT_ARGS \
+    --save $CHECKPOINT_PATH \
+    --load $CHECKPOINT_PATH \
+    --train-weighted-split-paths-path $TRAIN_DATA_PATH \
+    --valid-weighted-split-paths-path $VALID_DATA_PATH \
+    --data-impl mmap \
+     $DEEPSPEED_ARGS \
+    --seed 1 \
+    "
+
+echo $CMD
+
+echo "START $SLURM_JOBID: $(date)"
+
+# bash launch_srun_32.sh $CMD
+srun --label launch.sh $CMD
+
+echo "END $SLURM_JOBID: $(date)"

4b284b17bc4seed1/tensorboard_4b284b17bc4seed1val/events.out.tfevents.1683577811.nid006797.69786.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ae2ce505d4666394d2c563f1f824bc9af9046b95e2887466d9d8cbb434a795db
+size 980

4b284b17bc4seed2/latest

@@ -0,0 +1 @@
+global_step80108

4b284b17bc4seed2/sbatch_4b284b17bc4seed2val.sh

@@ -0,0 +1,169 @@
+#!/bin/bash
+#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
+#SBATCH --nodes=32
+#SBATCH --ntasks-per-node=1
+#SBATCH --cpus-per-task=40
+#SBATCH --mem=256G
+#SBATCH -p standard-g
+#SBATCH -t 48:00:00
+#SBATCH --gpus-per-node=mi250:8
+#SBATCH --exclusive=user
+#SBATCH --hint=nomultithread
+#SBATCH --account=project_462000119
+#SBATCH -o logs/%j.out
+#SBATCH -e logs/%j.err
+
+VARIANT=4b284b17bc4seed2val
+VARIANT_CKPT=lm1-4b2-84b-c4seeds/4b284b17bc4seed2
+
+# 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.out and logs/latest.err
+ln -f -s $SLURM_JOB_ID.out logs/latest.out
+ln -f -s $SLURM_JOB_ID.err logs/latest.err
+
+KILL_SWITCH_PATH=kill-switch-$VARIANT
+CHECKPOINT_PATH=$VARIANT_CKPT
+TENSORBOARD_PATH=tensorboard_$VARIANT
+
+# Data
+VOCAB_FILE="gpt2/vocab.json"
+MERGE_FILE="gpt2/merges.txt"
+#DATA_PATH="/scratch/project_462000119/data/pile/megatron_data/meg-gpt2_pile_text_document"
+TRAIN_DATA_PATH=train1b5.txt
+# "train: 1.0 0:1 /scratch/project_462000119/data/c4_subsampled/gpt2tok_c4_en_12B_text_document"
+VALID_DATA_PATH=val.txt
+# "validation: 1.0 0:1 /scratch/project_462000119/data/c4_validation/gpt2tok_c4validation_rerun_text_document"
+
+PP_SIZE=1
+TP_SIZE=2
+
+MICRO_BATCH_SIZE=2
+GRADIENT_ACCUMULATION_STEPS=2
+WORLD_SIZE=$((SLURM_GPUS_ON_NODE*SLURM_JOB_NUM_NODES))
+GLOBAL_BATCH_SIZE=$((MICRO_BATCH_SIZE*WORLD_SIZE*GRADIENT_ACCUMULATION_STEPS))
+
+# Model parameters
+source model_params.sh
+MODEL_PARAM=("${PARAM_4516M[@]}")
+NHIDDEN=${MODEL_PARAM[0]}
+FFN_HIDDEN_SIZE=${MODEL_PARAM[1]}
+KV_SIZE=${MODEL_PARAM[2]}
+NHEADS=${MODEL_PARAM[3]}
+NLAYERS=${MODEL_PARAM[4]}
+SEQ_LEN=2048
+
+echo "Model parameters: d_model $NHIDDEN ffw_size $FFN_HIDDEN_SIZE kv_size $KV_SIZE n_heads $NHEADS n_layers $NLAYERS"
+
+SAVE_INTERVAL=1000
+
+# Tokens: 25611230000
+# -> Samples: 12505484
+TRAIN_SAMPLES=1
+
+OPTIMIZER_ARGS=" \
+    --optimizer adam \
+    --adam-beta1 0.9 \
+    --adam-beta2 0.999 \
+    --adam-eps 1e-8 \
+    --lr 2e-4 \
+    --min-lr 2e-5 \
+    --lr-decay-style cosine \
+    --lr-decay-samples $TRAIN_SAMPLES \
+    --lr-warmup-samples 0 \
+    --clip-grad 1.0 \
+    --weight-decay 1e-1 \
+    --override-lr-scheduler \
+    --reset-progress \
+    --no-load-optim \
+    "
+
+GPT_ARGS=" \
+    --num-layers $NLAYERS \
+    --hidden-size $NHIDDEN \
+    --num-attention-heads $NHEADS \
+    --kv-channels $KV_SIZE \
+    --ffn-hidden-size $FFN_HIDDEN_SIZE \
+    --seq-length $SEQ_LEN \
+    --max-position-embeddings $SEQ_LEN \
+    --micro-batch-size $MICRO_BATCH_SIZE \
+    --global-batch-size $GLOBAL_BATCH_SIZE \
+    --train-samples $TRAIN_SAMPLES \
+    --vocab-file $VOCAB_FILE \
+    --merge-file $MERGE_FILE \
+    --clip-grad 1.0 \
+    --kill-switch-path $KILL_SWITCH_PATH \
+    --bf16 \
+    $OPTIMIZER_ARGS \
+    "
+
+OUTPUT_ARGS=" \
+    --log-interval 10 \
+    --save-interval $SAVE_INTERVAL \
+    --eval-interval 1 \
+    --eval-iters 100 \
+    --eval-only true \
+    --tensorboard-dir $TENSORBOARD_PATH \
+    --tensorboard-queue-size 5 \
+    --log-timers-to-tensorboard \
+    --log-batch-size-to-tensorboard \
+    --log-validation-ppl-to-tensorboard \
+    "
+
+ZERO_STAGE=0
+
+mkdir -p ds_configs
+DS_CONFIG_PATH="ds_configs/$SLURM_JOB_ID.json"
+
+cat <<EOF > $DS_CONFIG_PATH
+{
+    "train_micro_batch_size_per_gpu": $MICRO_BATCH_SIZE,
+    "train_batch_size": $GLOBAL_BATCH_SIZE,
+    "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/pretrain_gpt.py \
+    --tensor-model-parallel-size $TP_SIZE \
+    --pipeline-model-parallel-size $PP_SIZE \
+    $GPT_ARGS \
+    $OUTPUT_ARGS \
+    --save $CHECKPOINT_PATH \
+    --load $CHECKPOINT_PATH \
+    --train-weighted-split-paths-path $TRAIN_DATA_PATH \
+    --valid-weighted-split-paths-path $VALID_DATA_PATH \
+    --data-impl mmap \
+     $DEEPSPEED_ARGS \
+    --seed 2 \
+    "
+
+echo $CMD
+
+echo "START $SLURM_JOBID: $(date)"
+
+# bash launch_srun_32.sh $CMD
+srun --label launch.sh $CMD
+
+echo "END $SLURM_JOBID: $(date)"

4b284b17bc4seed2/tensorboard_4b284b17bc4seed2val/events.out.tfevents.1683577811.nid005490.71167.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c8d3f565ce06c741250c0bea8064bc9670cd7370d6d32dd25bbe705016c48a32
+size 980

4b284b17bc4seed3/latest

@@ -0,0 +1 @@
+global_step80108

4b284b17bc4seed3/sbatch_4b284b17bc4seed3val.sh

@@ -0,0 +1,169 @@
+#!/bin/bash
+#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
+#SBATCH --nodes=32
+#SBATCH --ntasks-per-node=1
+#SBATCH --cpus-per-task=40
+#SBATCH --mem=256G
+#SBATCH -p standard-g
+#SBATCH -t 48:00:00
+#SBATCH --gpus-per-node=mi250:8
+#SBATCH --exclusive=user
+#SBATCH --hint=nomultithread
+#SBATCH --account=project_462000119
+#SBATCH -o logs/%j.out
+#SBATCH -e logs/%j.err
+
+VARIANT=4b284b17bc4seed3val
+VARIANT_CKPT=lm1-4b2-84b-c4seeds/4b284b17bc4seed3
+
+# 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.out and logs/latest.err
+ln -f -s $SLURM_JOB_ID.out logs/latest.out
+ln -f -s $SLURM_JOB_ID.err logs/latest.err
+
+KILL_SWITCH_PATH=kill-switch-$VARIANT
+CHECKPOINT_PATH=$VARIANT_CKPT
+TENSORBOARD_PATH=tensorboard_$VARIANT
+
+# Data
+VOCAB_FILE="gpt2/vocab.json"
+MERGE_FILE="gpt2/merges.txt"
+#DATA_PATH="/scratch/project_462000119/data/pile/megatron_data/meg-gpt2_pile_text_document"
+TRAIN_DATA_PATH=train1b5.txt
+# "train: 1.0 0:1 /scratch/project_462000119/data/c4_subsampled/gpt2tok_c4_en_12B_text_document"
+VALID_DATA_PATH=val.txt
+# "validation: 1.0 0:1 /scratch/project_462000119/data/c4_validation/gpt2tok_c4validation_rerun_text_document"
+
+PP_SIZE=1
+TP_SIZE=2
+
+MICRO_BATCH_SIZE=2
+GRADIENT_ACCUMULATION_STEPS=2
+WORLD_SIZE=$((SLURM_GPUS_ON_NODE*SLURM_JOB_NUM_NODES))
+GLOBAL_BATCH_SIZE=$((MICRO_BATCH_SIZE*WORLD_SIZE*GRADIENT_ACCUMULATION_STEPS))
+
+# Model parameters
+source model_params.sh
+MODEL_PARAM=("${PARAM_4516M[@]}")
+NHIDDEN=${MODEL_PARAM[0]}
+FFN_HIDDEN_SIZE=${MODEL_PARAM[1]}
+KV_SIZE=${MODEL_PARAM[2]}
+NHEADS=${MODEL_PARAM[3]}
+NLAYERS=${MODEL_PARAM[4]}
+SEQ_LEN=2048
+
+echo "Model parameters: d_model $NHIDDEN ffw_size $FFN_HIDDEN_SIZE kv_size $KV_SIZE n_heads $NHEADS n_layers $NLAYERS"
+
+SAVE_INTERVAL=1000
+
+# Tokens: 25611230000
+# -> Samples: 12505484
+TRAIN_SAMPLES=1
+
+OPTIMIZER_ARGS=" \
+    --optimizer adam \
+    --adam-beta1 0.9 \
+    --adam-beta2 0.999 \
+    --adam-eps 1e-8 \
+    --lr 2e-4 \
+    --min-lr 2e-5 \
+    --lr-decay-style cosine \
+    --lr-decay-samples $TRAIN_SAMPLES \
+    --lr-warmup-samples 0 \
+    --clip-grad 1.0 \
+    --weight-decay 1e-1 \
+    --override-lr-scheduler \
+    --reset-progress \
+    --no-load-optim \
+    "
+
+GPT_ARGS=" \
+    --num-layers $NLAYERS \
+    --hidden-size $NHIDDEN \
+    --num-attention-heads $NHEADS \
+    --kv-channels $KV_SIZE \
+    --ffn-hidden-size $FFN_HIDDEN_SIZE \
+    --seq-length $SEQ_LEN \
+    --max-position-embeddings $SEQ_LEN \
+    --micro-batch-size $MICRO_BATCH_SIZE \
+    --global-batch-size $GLOBAL_BATCH_SIZE \
+    --train-samples $TRAIN_SAMPLES \
+    --vocab-file $VOCAB_FILE \
+    --merge-file $MERGE_FILE \
+    --clip-grad 1.0 \
+    --kill-switch-path $KILL_SWITCH_PATH \
+    --bf16 \
+    $OPTIMIZER_ARGS \
+    "
+
+OUTPUT_ARGS=" \
+    --log-interval 10 \
+    --save-interval $SAVE_INTERVAL \
+    --eval-interval 1 \
+    --eval-iters 100 \
+    --eval-only true \
+    --tensorboard-dir $TENSORBOARD_PATH \
+    --tensorboard-queue-size 5 \
+    --log-timers-to-tensorboard \
+    --log-batch-size-to-tensorboard \
+    --log-validation-ppl-to-tensorboard \
+    "
+
+ZERO_STAGE=0
+
+mkdir -p ds_configs
+DS_CONFIG_PATH="ds_configs/$SLURM_JOB_ID.json"
+
+cat <<EOF > $DS_CONFIG_PATH
+{
+    "train_micro_batch_size_per_gpu": $MICRO_BATCH_SIZE,
+    "train_batch_size": $GLOBAL_BATCH_SIZE,
+    "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/pretrain_gpt.py \
+    --tensor-model-parallel-size $TP_SIZE \
+    --pipeline-model-parallel-size $PP_SIZE \
+    $GPT_ARGS \
+    $OUTPUT_ARGS \
+    --save $CHECKPOINT_PATH \
+    --load $CHECKPOINT_PATH \
+    --train-weighted-split-paths-path $TRAIN_DATA_PATH \
+    --valid-weighted-split-paths-path $VALID_DATA_PATH \
+    --data-impl mmap \
+     $DEEPSPEED_ARGS \
+    --seed 3 \
+    "
+
+echo $CMD
+
+echo "START $SLURM_JOBID: $(date)"
+
+# bash launch_srun_32.sh $CMD
+srun --label launch.sh $CMD
+
+echo "END $SLURM_JOBID: $(date)"

4b284b17bc4seed3/tensorboard_4b284b17bc4seed3val/events.out.tfevents.1683577786.nid005175.50318.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8f2c88716d43138146c4709fb26cc758c4ffdf1be37add9d81e2a6a5b64c4ed3
+size 980

4b284b17bc4seed4/latest

@@ -0,0 +1 @@
+global_step80108

4b284b17bc4seed4/sbatch_4b284b17bc4seed4val.sh

@@ -0,0 +1,169 @@
+#!/bin/bash
+#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
+#SBATCH --nodes=32
+#SBATCH --ntasks-per-node=1
+#SBATCH --cpus-per-task=40
+#SBATCH --mem=256G
+#SBATCH -p standard-g
+#SBATCH -t 48:00:00
+#SBATCH --gpus-per-node=mi250:8
+#SBATCH --exclusive=user
+#SBATCH --hint=nomultithread
+#SBATCH --account=project_462000119
+#SBATCH -o logs/%j.out
+#SBATCH -e logs/%j.err
+
+VARIANT=4b284b17bc4seed4val
+VARIANT_CKPT=lm1-4b2-84b-c4seeds/4b284b17bc4seed4
+
+# 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.out and logs/latest.err
+ln -f -s $SLURM_JOB_ID.out logs/latest.out
+ln -f -s $SLURM_JOB_ID.err logs/latest.err
+
+KILL_SWITCH_PATH=kill-switch-$VARIANT
+CHECKPOINT_PATH=$VARIANT_CKPT
+TENSORBOARD_PATH=tensorboard_$VARIANT
+
+# Data
+VOCAB_FILE="gpt2/vocab.json"
+MERGE_FILE="gpt2/merges.txt"
+#DATA_PATH="/scratch/project_462000119/data/pile/megatron_data/meg-gpt2_pile_text_document"
+TRAIN_DATA_PATH=train1b5.txt
+# "train: 1.0 0:1 /scratch/project_462000119/data/c4_subsampled/gpt2tok_c4_en_12B_text_document"
+VALID_DATA_PATH=val.txt
+# "validation: 1.0 0:1 /scratch/project_462000119/data/c4_validation/gpt2tok_c4validation_rerun_text_document"
+
+PP_SIZE=1
+TP_SIZE=2
+
+MICRO_BATCH_SIZE=2
+GRADIENT_ACCUMULATION_STEPS=2
+WORLD_SIZE=$((SLURM_GPUS_ON_NODE*SLURM_JOB_NUM_NODES))
+GLOBAL_BATCH_SIZE=$((MICRO_BATCH_SIZE*WORLD_SIZE*GRADIENT_ACCUMULATION_STEPS))
+
+# Model parameters
+source model_params.sh
+MODEL_PARAM=("${PARAM_4516M[@]}")
+NHIDDEN=${MODEL_PARAM[0]}
+FFN_HIDDEN_SIZE=${MODEL_PARAM[1]}
+KV_SIZE=${MODEL_PARAM[2]}
+NHEADS=${MODEL_PARAM[3]}
+NLAYERS=${MODEL_PARAM[4]}
+SEQ_LEN=2048
+
+echo "Model parameters: d_model $NHIDDEN ffw_size $FFN_HIDDEN_SIZE kv_size $KV_SIZE n_heads $NHEADS n_layers $NLAYERS"
+
+SAVE_INTERVAL=1000
+
+# Tokens: 25611230000
+# -> Samples: 12505484
+TRAIN_SAMPLES=1
+
+OPTIMIZER_ARGS=" \
+    --optimizer adam \
+    --adam-beta1 0.9 \
+    --adam-beta2 0.999 \
+    --adam-eps 1e-8 \
+    --lr 2e-4 \
+    --min-lr 2e-5 \
+    --lr-decay-style cosine \
+    --lr-decay-samples $TRAIN_SAMPLES \
+    --lr-warmup-samples 0 \
+    --clip-grad 1.0 \
+    --weight-decay 1e-1 \
+    --override-lr-scheduler \
+    --reset-progress \
+    --no-load-optim \
+    "
+
+GPT_ARGS=" \
+    --num-layers $NLAYERS \
+    --hidden-size $NHIDDEN \
+    --num-attention-heads $NHEADS \
+    --kv-channels $KV_SIZE \
+    --ffn-hidden-size $FFN_HIDDEN_SIZE \
+    --seq-length $SEQ_LEN \
+    --max-position-embeddings $SEQ_LEN \
+    --micro-batch-size $MICRO_BATCH_SIZE \
+    --global-batch-size $GLOBAL_BATCH_SIZE \
+    --train-samples $TRAIN_SAMPLES \
+    --vocab-file $VOCAB_FILE \
+    --merge-file $MERGE_FILE \
+    --clip-grad 1.0 \
+    --kill-switch-path $KILL_SWITCH_PATH \
+    --bf16 \
+    $OPTIMIZER_ARGS \
+    "
+
+OUTPUT_ARGS=" \
+    --log-interval 10 \
+    --save-interval $SAVE_INTERVAL \
+    --eval-interval 1 \
+    --eval-iters 100 \
+    --eval-only true \
+    --tensorboard-dir $TENSORBOARD_PATH \
+    --tensorboard-queue-size 5 \
+    --log-timers-to-tensorboard \
+    --log-batch-size-to-tensorboard \
+    --log-validation-ppl-to-tensorboard \
+    "
+
+ZERO_STAGE=0
+
+mkdir -p ds_configs
+DS_CONFIG_PATH="ds_configs/$SLURM_JOB_ID.json"
+
+cat <<EOF > $DS_CONFIG_PATH
+{
+    "train_micro_batch_size_per_gpu": $MICRO_BATCH_SIZE,
+    "train_batch_size": $GLOBAL_BATCH_SIZE,
+    "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/pretrain_gpt.py \
+    --tensor-model-parallel-size $TP_SIZE \
+    --pipeline-model-parallel-size $PP_SIZE \
+    $GPT_ARGS \
+    $OUTPUT_ARGS \
+    --save $CHECKPOINT_PATH \
+    --load $CHECKPOINT_PATH \
+    --train-weighted-split-paths-path $TRAIN_DATA_PATH \
+    --valid-weighted-split-paths-path $VALID_DATA_PATH \
+    --data-impl mmap \
+     $DEEPSPEED_ARGS \
+    --seed 4 \
+    "
+
+echo $CMD
+
+echo "START $SLURM_JOBID: $(date)"
+
+# bash launch_srun_32.sh $CMD
+srun --label launch.sh $CMD
+
+echo "END $SLURM_JOBID: $(date)"

4b284b17bc4seed4/tensorboard_4b284b17bc4seed4val/events.out.tfevents.1683577786.nid007191.52326.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e682cde3f6246d721792cbb49e50f94bccbc5185ebc000b84ba5d1a9fdcf5721
+size 980

4b284b1b9c4seed1/latest

@@ -0,0 +1 @@
+global_step80108

4b284b1b9c4seed1/sbatch_4b284b1b9c4seed1val.sh

@@ -0,0 +1,169 @@
+#!/bin/bash
+#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
+#SBATCH --nodes=32
+#SBATCH --ntasks-per-node=1
+#SBATCH --cpus-per-task=40
+#SBATCH --mem=256G
+#SBATCH -p standard-g
+#SBATCH -t 48:00:00
+#SBATCH --gpus-per-node=mi250:8
+#SBATCH --exclusive=user
+#SBATCH --hint=nomultithread
+#SBATCH --account=project_462000119
+#SBATCH -o logs/%j.out
+#SBATCH -e logs/%j.err
+
+VARIANT=4b284b1b9c4seed1val
+VARIANT_CKPT=lm1-4b2-84b-c4seeds/4b284b1b9c4seed1
+
+# 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.out and logs/latest.err
+ln -f -s $SLURM_JOB_ID.out logs/latest.out
+ln -f -s $SLURM_JOB_ID.err logs/latest.err
+
+KILL_SWITCH_PATH=kill-switch-$VARIANT
+CHECKPOINT_PATH=$VARIANT_CKPT
+TENSORBOARD_PATH=tensorboard_$VARIANT
+
+# Data
+VOCAB_FILE="gpt2/vocab.json"
+MERGE_FILE="gpt2/merges.txt"
+#DATA_PATH="/scratch/project_462000119/data/pile/megatron_data/meg-gpt2_pile_text_document"
+TRAIN_DATA_PATH=train1b5.txt
+# "train: 1.0 0:1 /scratch/project_462000119/data/c4_subsampled/gpt2tok_c4_en_12B_text_document"
+VALID_DATA_PATH=val.txt
+# "validation: 1.0 0:1 /scratch/project_462000119/data/c4_validation/gpt2tok_c4validation_rerun_text_document"
+
+PP_SIZE=1
+TP_SIZE=2
+
+MICRO_BATCH_SIZE=2
+GRADIENT_ACCUMULATION_STEPS=2
+WORLD_SIZE=$((SLURM_GPUS_ON_NODE*SLURM_JOB_NUM_NODES))
+GLOBAL_BATCH_SIZE=$((MICRO_BATCH_SIZE*WORLD_SIZE*GRADIENT_ACCUMULATION_STEPS))
+
+# Model parameters
+source model_params.sh
+MODEL_PARAM=("${PARAM_4516M[@]}")
+NHIDDEN=${MODEL_PARAM[0]}
+FFN_HIDDEN_SIZE=${MODEL_PARAM[1]}
+KV_SIZE=${MODEL_PARAM[2]}
+NHEADS=${MODEL_PARAM[3]}
+NLAYERS=${MODEL_PARAM[4]}
+SEQ_LEN=2048
+
+echo "Model parameters: d_model $NHIDDEN ffw_size $FFN_HIDDEN_SIZE kv_size $KV_SIZE n_heads $NHEADS n_layers $NLAYERS"
+
+SAVE_INTERVAL=1000
+
+# Tokens: 25611230000
+# -> Samples: 12505484
+TRAIN_SAMPLES=1
+
+OPTIMIZER_ARGS=" \
+    --optimizer adam \
+    --adam-beta1 0.9 \
+    --adam-beta2 0.999 \
+    --adam-eps 1e-8 \
+    --lr 2e-4 \
+    --min-lr 2e-5 \
+    --lr-decay-style cosine \
+    --lr-decay-samples $TRAIN_SAMPLES \
+    --lr-warmup-samples 0 \
+    --clip-grad 1.0 \
+    --weight-decay 1e-1 \
+    --override-lr-scheduler \
+    --reset-progress \
+    --no-load-optim \
+    "
+
+GPT_ARGS=" \
+    --num-layers $NLAYERS \
+    --hidden-size $NHIDDEN \
+    --num-attention-heads $NHEADS \
+    --kv-channels $KV_SIZE \
+    --ffn-hidden-size $FFN_HIDDEN_SIZE \
+    --seq-length $SEQ_LEN \
+    --max-position-embeddings $SEQ_LEN \
+    --micro-batch-size $MICRO_BATCH_SIZE \
+    --global-batch-size $GLOBAL_BATCH_SIZE \
+    --train-samples $TRAIN_SAMPLES \
+    --vocab-file $VOCAB_FILE \
+    --merge-file $MERGE_FILE \
+    --clip-grad 1.0 \
+    --kill-switch-path $KILL_SWITCH_PATH \
+    --bf16 \
+    $OPTIMIZER_ARGS \
+    "
+
+OUTPUT_ARGS=" \
+    --log-interval 10 \
+    --save-interval $SAVE_INTERVAL \
+    --eval-interval 1 \
+    --eval-iters 100 \
+    --eval-only true \
+    --tensorboard-dir $TENSORBOARD_PATH \
+    --tensorboard-queue-size 5 \
+    --log-timers-to-tensorboard \
+    --log-batch-size-to-tensorboard \
+    --log-validation-ppl-to-tensorboard \
+    "
+
+ZERO_STAGE=0
+
+mkdir -p ds_configs
+DS_CONFIG_PATH="ds_configs/$SLURM_JOB_ID.json"
+
+cat <<EOF > $DS_CONFIG_PATH
+{
+    "train_micro_batch_size_per_gpu": $MICRO_BATCH_SIZE,
+    "train_batch_size": $GLOBAL_BATCH_SIZE,
+    "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/pretrain_gpt.py \
+    --tensor-model-parallel-size $TP_SIZE \
+    --pipeline-model-parallel-size $PP_SIZE \
+    $GPT_ARGS \
+    $OUTPUT_ARGS \
+    --save $CHECKPOINT_PATH \
+    --load $CHECKPOINT_PATH \
+    --train-weighted-split-paths-path $TRAIN_DATA_PATH \
+    --valid-weighted-split-paths-path $VALID_DATA_PATH \
+    --data-impl mmap \
+     $DEEPSPEED_ARGS \
+    --seed 1 \
+    "
+
+echo $CMD
+
+echo "START $SLURM_JOBID: $(date)"
+
+# bash launch_srun_32.sh $CMD
+srun --label launch.sh $CMD
+
+echo "END $SLURM_JOBID: $(date)"

4b284b1b9c4seed1/tensorboard_4b284b1b9c4seed1val/events.out.tfevents.1683577914.nid005595.130240.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5c9586d7b9a849c71bfeab6a4e57133d8bf1c742cacdf79849fb3b906394e0ca
+size 40

4b284b1b9c4seed1/tensorboard_4b284b1b9c4seed1val/events.out.tfevents.1683579446.nid007109.42748.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:677b7daa12cd5c3acf2d4f0125d81c11a158c47819aaf2e875f0fe35bf8bd14b
+size 980

4b284b1b9c4seed2/latest

@@ -0,0 +1 @@
+global_step80108