#!/bin/bash

# Runs the "7B" parameter model
export HSA_FORCE_FINE_GRAIN_PCIE=1
export OMP_NUM_THREADS=1
export NCCL_P2P_LEVEL=PXB

source /opt/dtk/env.sh
export LD_LIBRARY_PATH=/mnt/fs/user/llama/dcu/dtk-24.04/rocblas/lib:$LD_LIBRARY_PATH
#export LD_LIBRARY_PATH=/mnt/fs/user/llama/dcu/rocblas-install_padding/lib:$LD_LIBRARY_PATH

#export HIP_ALLOC_INITIALIZE=0
#export GPU_MAX_HW_QUEUES=20

export NCCL_ALGO=Ring
export NCCL_NCHANNELS_PER_PEER=16
export NCCL_MIN_NCHANNELS=20
export NCCL_IB_TIMEOUT=22
export CUDA_DEVICE_MAX_CONNECTIONS=1

export NCCL_IB_HCA=mlx5_1,mlx5_2
#export NCCL_SOCKET_IFNAME=ibs8
export NCCL_NET_GDR_LEVEL=SYS
export NCCL_NET_GDR_READ=0
#export NCCL_DEBUG=info

lrank=$OMPI_COMM_WORLD_LOCAL_RANK
RANK=$OMPI_COMM_WORLD_RANK
WORLD_SIZE=$OMPI_COMM_WORLD_SIZE

CHECKPOINT_PATH=./tmp #$1 #<Specify path>
TENSORBOARD_LOGS_PATH=./tmp  #$2 #<Specify path>
DATA_PATH="/mnt/fs/user/llama/dcu/Megatron-LM-main/oscar/my-gpt2_text_document" #<Specify path and file prefix>_text_document

GPT_MODEL_ARGS=(
    --num-layers 96 
    --hidden-size 12288
    --num-attention-heads 96
    --seq-length 2048 
    --max-position-embeddings 2048
)

TRAINING_ARGS=(
    --transformer-impl local
    --use-legacy-models 
    --micro-batch-size 1 
    --global-batch-size 512 #256 1024
    --train-iters 100 
    --weight-decay 0.1 
    --adam-beta1 0.9 
    --adam-beta2 0.95 
    --init-method-std 0.006 
    --clip-grad 1.0 
    --bf16
    --disable-bias-linear
    --no-bias-gelu-fusion
    --no-gradient-accumulation-fusion
    --use-distributed-optimizer 
    --use-flash-attn-triton
    --lr 1.5e-04
    --use-tp-pp-dp-mapping
    --lr-decay-style cosine 
    --min-lr 1.5e-05
    --lr-warmup-iters 1
)
#--initial-loss-scale 65536
#--loss-scale 1024
#--recompute-granularity full
#--recompute-method uniform
#--recompute-num-layers 1
#--recompute-activations
#--use-flash-attn
MODEL_PARALLEL_ARGS=(
        --sequence-parallel
	--tensor-model-parallel-size 4   #4
	--pipeline-model-parallel-size 8 #4   8
)

DATA_ARGS=(
    --data-path $DATA_PATH 
    --split 949,50,1
    --tokenizer-type GPT2BPETokenizer
    --vocab-file gpt2-vocab.json
    --merge-file gpt2-merges.txt
)

EVAL_AND_LOGGING_ARGS=(
    --log-interval 1
    --log-throughput
    --save-interval 10000 
    --eval-interval 1000 
    --save $CHECKPOINT_PATH 
    --load $CHECKPOINT_PATH 
    --eval-iters 10
    --tensorboard-dir $TENSORBOARD_LOGS_PATH 
)

APP="python3  -u pretrain_gpt.py \
     ${GPT_MODEL_ARGS[@]} \
     ${TRAINING_ARGS[@]} \
     ${MODEL_PARALLEL_ARGS[@]} \
     ${DATA_ARGS[@]} \
     ${EVAL_AND_LOGGING_ARGS[@]}
     --rank ${RANK} \
     --world_size ${WORLD_SIZE} \
     --dist_url tcp://${1}:34566 \
    "


case ${lrank} in
[0])
  export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
  ${APP}
  ;;
[1])
  export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
  ${APP}
  ;;
[2])
  export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
  ${APP}
  ;;
[3])
  export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
  ${APP}
  ;;
[4])
  export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
  ${APP}
  ;;
[5])
  export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
  ${APP}
  ;;
[6])
  export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
  ${APP}
  ;;
[7])
  export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
  ${APP}
  ;;
esac