Commit 8a6d7a00 authored by Gounley, John's avatar Gounley, John
Browse files

Update adam submission script for 1.3B and 3.9B megatron-style berts

parent 6594122f
Loading
Loading
Loading
Loading
Loading
+9 −1
Original line number Diff line number Diff line
@@ -52,7 +52,15 @@ export CHECKPOINT_PATH=/gpfs/alpine/world-shared/med106/gounley1/crusher2/Megatr
# source export_DDP_vars.sh
# python pretrain_bert.py --tensor-model-parallel-size 1 --pipeline-model-parallel-size 1 --num-layers 12 --hidden-size 768 --num-attention-heads 12 --micro-batch-size 256 --global-batch-size 6144 --seq-length 64 --max-position-embeddings 512 --train-iters 6000 --save $CHECKPOINT_PATH --data-path $DATASET --vocab-file $VOCAB_FILE --data-impl mmap --split 949,50,1 --distributed-backend nccl --lr 0.0001 --min-lr 0.00001 --lr-decay-style linear --lr-warmup-fraction .01 --weight-decay 1e-2 --clip-grad 1.0 --log-interval 100 --save-interval 10000 --eval-interval 100 --eval-iters 10 --bert-no-binary-head --num-workers 2 --HIP 1 --fp16"

# srun -u -n $ranks_total -c 8 --gpus-per-task=$gpus_per_rank --gpu-bind=closest bash -c "
# source export_DDP_vars.sh
# python pretrain_bert.py --tensor-model-parallel-size 1 --pipeline-model-parallel-size 1 --num-layers 24 --hidden-size 1024 --num-attention-heads 16 --micro-batch-size 96 --global-batch-size 2304 --seq-length 64 --max-position-embeddings 512 --train-iters 6000 --save $CHECKPOINT_PATH --data-path $DATASET --vocab-file $VOCAB_FILE --data-impl mmap --split 949,50,1 --distributed-backend nccl --lr 0.0001 --min-lr 0.00001 --lr-decay-style linear --lr-warmup-fraction .01 --weight-decay 1e-2 --clip-grad 1.0 --log-interval 100 --save-interval 10000 --eval-interval 100 --eval-iters 100 --bert-no-binary-head --num-workers 2 --HIP 1 --fp16"

srun -u -n $ranks_total -c 8 --gpus-per-task=$gpus_per_rank --gpu-bind=closest bash -c "
source export_DDP_vars.sh
python pretrain_bert.py --tensor-model-parallel-size 1 --pipeline-model-parallel-size 1 --num-layers 24 --hidden-size 1024 --num-attention-heads 16 --micro-batch-size 96 --global-batch-size 2304 --seq-length 64 --max-position-embeddings 512 --train-iters 6000 --save $CHECKPOINT_PATH --data-path $DATASET --vocab-file $VOCAB_FILE --data-impl mmap --split 949,50,1 --distributed-backend nccl --lr 0.0001 --min-lr 0.00001 --lr-decay-style linear --lr-warmup-fraction .01 --weight-decay 1e-2 --clip-grad 1.0 --log-interval 100 --save-interval 10000 --eval-interval 100 --eval-iters 100 --bert-no-binary-head --num-workers 2 --HIP 1 --fp16"
python pretrain_bert.py --tensor-model-parallel-size 1 --pipeline-model-parallel-size 1 --num-layers 24 --hidden-size 2048 --num-attention-heads 32 --micro-batch-size 48 --global-batch-size 1152 --seq-length 64 --max-position-embeddings 512 --train-iters 6000 --save $CHECKPOINT_PATH --data-path $DATASET --vocab-file $VOCAB_FILE --data-impl mmap --split 949,50,1 --distributed-backend nccl --lr 0.0001 --min-lr 0.00001 --lr-decay-style linear --lr-warmup-fraction .01 --weight-decay 1e-2 --clip-grad 1.0 --log-interval 100 --save-interval 10000 --eval-interval 100 --eval-iters 100 --bert-no-binary-head --num-workers 2 --HIP 1 --fp16"

# srun -u -n $ranks_total -c 8 --gpus-per-task=$gpus_per_rank --gpu-bind=closest bash -c "
# source export_DDP_vars.sh
# python pretrain_bert.py --tensor-model-parallel-size 1 --pipeline-model-parallel-size 2 --num-layers 48 --hidden-size 2560 --num-attention-heads 40 --micro-batch-size 16 --global-batch-size 192 --seq-length 64 --max-position-embeddings 512 --train-iters 500 --save $CHECKPOINT_PATH --data-path $DATASET --vocab-file $VOCAB_FILE --data-impl mmap --split 949,50,1 --distributed-backend nccl --lr 0.0001 --min-lr 0.00001 --lr-decay-style linear --lr-warmup-fraction .01 --weight-decay 1e-2 --clip-grad 1.0 --log-interval 100 --save-interval 10000 --eval-interval 100 --eval-iters 100 --bert-no-binary-head --num-workers 2 --HIP 1 --fp16"