#!/bin/bash

source /opt/dtk/env.sh
# Runs Mixtral 8x7B model
export HIP_DIRECT_DISPATCH=0
export HSA_FORCE_FINE_GRAIN_PCIE=1
export OMP_NUM_THREADS=1
export GPU_MAX_HW_QUEUES=10
export NVTE_FORCE_BLASLT=1

export NCCL_ALGO=Ring
export NCCL_NCHANNELS_PER_PEER=2
export NCCL_MIN_NCHANNELS=16
export NCCL_IB_TIMEOUT=22
export CUDA_DEVICE_MAX_CONNECTIONS=1
#export NCCL_IB_HCA=mlx5_0
#export NCCL_SOCKET_IFNAME=enp145s0f0
export NCCL_NET_GDR_LEVEL=SYS
export NCCL_NET_GDR_READ=0
export LD_LIBRARY_PATH=/opt/hipblaslt-install/lib/:$LD_LIBRARY_PATH

RANK=$OMPI_COMM_WORLD_RANK
LOCAL_RANK=$OMPI_COMM_WORLD_LOCAL_RANK
WORLD_SIZE=$OMPI_COMM_WORLD_SIZE
DIST_URL=${1}
DIST_PORT=25900

CHECKPOINT_PATH=./CKPT 
TOKENIZER_MODEL=../Mixtral8x7B/mixtral_dataset/tokenizer.model
DATA_PATH=../Mixtral8x7B/mixtral_dataset/my-mixtral_text_document

DISTRIBUTED_ARGS=(
    --rank ${RANK}
    --world-size ${WORLD_SIZE}
    --local-rank ${LOCAL_RANK}
    --dist-url tcp://${DIST_URL}:${DIST_PORT}
)

MODEL_ARGS=(
    --use-mcore-models
    --disable-bias-linear
    --seq-length 4096
    --max-position-embeddings 32768
    --num-layers 2
    --hidden-size 1024
    --ffn-hidden-size 14336
    --num-attention-heads 32
    --init-method-std 0.01
    --attention-dropout 0.0
    --hidden-dropout 0.0
    --normalization RMSNorm
    --position-embedding-type rope
    --swiglu
    --untie-embeddings-and-output-weights
    --group-query-attention
    --num-query-groups 8
    --no-masked-softmax-fusion
    --no-position-embedding
    --rotary-base 1000000
)

MOE_ARGS=(
    --num-experts 8
    --moe-router-topk 2
    --moe-router-load-balancing-type aux_loss
    --moe-aux-loss-coeff 1e-2
    --moe-token-dispatcher-type alltoall
    --overlap-param-gather
    --overlap-grad-reduce
    --moe-expert-capacity-factor 0.5
    --moe-pad-expert-input-to-capacity
    --moe-grouped-gemm
)

DATA_ARGS=(
    --tokenizer-type Llama2Tokenizer
    --tokenizer-model ${TOKENIZER_MODEL}
    --data-path $DATA_PATH
    --split 99990,8,2
)

TRAINING_ARGS=(
    --micro-batch-size 1
    --global-batch-size 16
    --lr 1e-4
    --train-iters 20
    --lr-decay-iters 320000
    --lr-decay-style cosine
    --min-lr 1.0e-5
    --weight-decay 0.1
    --lr-warmup-iters 500
    --clip-grad 1.0
    --bf16
)

MODEL_PARALLEL_ARGS=(
    --tensor-model-parallel-size 2
    --pipeline-model-parallel-size 1
    --expert-model-parallel-size 2
    --expert-tensor-parallel-size 1
    --use-distributed-optimizer
    --sequence-parallel
)

LOGGING_ARGS=(
    --log-throughput \
    --log-interval 1 \
    --save-interval 10000 \
    --eval-interval 1000 \
    --eval-iters -1 \
    #--save $CHECKPOINT_PATH \
    #--load $CHECKPOINT_PATH \
    --tensorboard-dir "${CHECKPOINT_PATH}/tensorboard" \
    --no-load-optim \
    --no-load-rng
)

if [ -n "${WANDB_API_KEY}" ]; then
    LOGGING_ARGS+=(
        --wandb-project ${WANDB_PROJECT:-"Mixtral"}
        --wandb-exp-name ${WANDB_NAME:-"Mixtral_8x7B"}
    )
fi

APP="python3 -u pretrain_gpt.py \
    ${DISTRIBUTED_ARGS[@]} \
    ${MODEL_ARGS[@]} \
    ${MOE_ARGS[@]} \
    ${DATA_ARGS[@]} \
    ${TRAINING_ARGS[@]} \
    ${MODEL_PARALLEL_ARGS[@]} \
    ${LOGGING_ARGS[@]} \
    "

#for hygon cpu
case ${LOCAL_RANK} 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