#!/bin/bash

# Runs the "70B" parameter model
export HSA_FORCE_FINE_GRAIN_PCIE=1
export OMP_NUM_THREADS=1
export NCCL_P2P_LEVEL=SYS

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=xx #based on your environment
export NCCL_NET_GDR_LEVEL=SYS
export NCCL_NET_GDR_READ=0

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="/path_to_dataset/my-llama_text_document" #<Specify path and file prefix>_text_document
TOKENIZER_PATH="/path_to_tokenizer.model"

GPT_MODEL_ARGS=(
    --num-layers 80 
    --hidden-size 8192
    --num-attention-heads 64
    --ffn-hidden-size 28672
    --seq-length 4096 
    --max-position-embeddings 4096
    --num-query-groups 8
    --group-query-attention
)

TRAINING_ARGS=(
    --log-throughput
    --transformer-impl transformer_engine
    --use-mcore-models
    --micro-batch-size 1
    --global-batch-size 512
    --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
    --use-distributed-optimizer
    --use-flash-attn-triton
    --disable-bias-linear
    --attention-dropout 0
    --hidden-dropout 0
    --no-gradient-accumulation-fusion
    --overlap-grad-reduce
    --swiglu
    --lr 5.0e-4 
    --lr-decay-style cosine 
    --min-lr 1.0e-4
    --lr-warmup-iters 1
    --use-fast-cross-entropy-loss
)

MODEL_PARALLEL_ARGS=(
	--sequence-parallel
	--tensor-model-parallel-size 4
	--pipeline-model-parallel-size 8 
)
DATA_ARGS=(
    --data-path $DATA_PATH 
    --split 949,50,1
    --untie-embeddings-and-output-weights
    --use-rotary-position-embeddings 
    --normalization RMSNorm 
    --no-position-embedding 
    --tokenizer-model $TOKENIZER_PATH 
    --tokenizer-type Llama2Tokenizer
)

EVAL_AND_LOGGING_ARGS=(
    --log-interval 1
    --log-throughput
    --save-interval 200 
    --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 \
    "

#for intel cpu
case ${lrank} in
[0])
  export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
  numactl --cpunodebind=0 --membind=0 ${APP}
  ;;
[1])
  export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
  numactl --cpunodebind=1 --membind=1 ${APP}
  ;;
[2])
  export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
  numactl --cpunodebind=2 --membind=2 ${APP}
  ;;
[3])
  export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
  numactl --cpunodebind=3 --membind=3 ${APP}
  ;;
[4])
  export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
  numactl --cpunodebind=4 --membind=4 ${APP}
  ;;
[5])
  export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
  numactl --cpunodebind=5 --membind=5 ${APP}
  ;;
[6])
  export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
  numactl --cpunodebind=6 --membind=6 ${APP}
  ;;
[7])
  export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
  numactl --cpunodebind=7 --membind=7 ${APP}
  ;;
esac