#!/bin/bash
source /usr/local/Ascend/cann/set_env.sh
# 该变量只用于规避megatron对其校验，对npu无效
export CUDA_DEVICE_MAX_CONNECTIONS=1
export ASCEND_SLOG_PRINT_TO_STDOUT=0
export ASCEND_GLOBAL_LOG_LEVEL=3
export TASK_QUEUE_ENABLE=2
export COMBINED_ENABLE=1
export CPU_AFFINITY_CONF=1
export HCCL_CONNECT_TIMEOUT=1200

GPUS_PER_NODE=8
MASTER_ADDR=localhost
MASTER_PORT=29501
NNODES=1
NODE_RANK=0
WORLD_SIZE=$(($GPUS_PER_NODE*$NNODES))

TP=1
PP=1
CP=1
MBS=8
GBS=$(($WORLD_SIZE*$MBS/$CP))

MM_DATA="./examples/whisper/data.json"
MM_MODEL="./examples/whisper/model.json"
MM_TOOL="./mindspeed_mm/tools/tools.json"

DISTRIBUTED_ARGS="
    --nproc_per_node $GPUS_PER_NODE \
    --nnodes $NNODES \
    --node_rank $NODE_RANK \
    --master_addr $MASTER_ADDR \
    --master_port $MASTER_PORT
"

GPT_ARGS="
    --tensor-model-parallel-size ${TP} \
    --pipeline-model-parallel-size ${PP} \
    --context-parallel-size ${CP} \
    --micro-batch-size ${MBS} \
    --global-batch-size ${GBS} \
    --num-layers 32 \
    --hidden-size 1280 \
    --num-attention-heads 20 \
    --seq-length 448 \
    --max-position-embeddings 1500 \
    --attention-dropout 0.0 \
    --hidden-dropout 0.0 \
    --tokenizer-type NullTokenizer \
    --vocab-size 0 \
    --position-embedding-type rope \
    --rotary-base 500000 \
    --swiglu \
    --no-masked-softmax-fusion \
    --lr 1e-5 \
    --train-iters 300 \
    --weight-decay 0.0 \
    --no-gradient-accumulation-fusion \
    --no-load-optim \
    --no-load-rng \
    --no-save-optim \
    --no-save-rng \
    --bf16 \
    --use-distributed-optimizer \
    --overlap-grad-reduce \
    --overlap-param-gather \
    --num-workers 4 \
"
# To ensure code security, configure trust_remote_code to default to False.
# Users need to add the following parameter and ensure the security of the models and data they download.
# --trust-remote-code \
MM_ARGS="
    --mm-data $MM_DATA \
    --mm-model $MM_MODEL \
    --mm-tool $MM_TOOL
"

OUTPUT_ARGS="
    --log-interval 1 \
    --save-interval 10000 \
    --eval-interval 10000 \
    --eval-iters 10 \
"
logfile=$(date +%Y%m%d)_$(date +%H%M%S)
mkdir -p logs
torchrun $DISTRIBUTED_ARGS pretrain_whisper.py \
    $GPT_ARGS \
    $MM_ARGS \
    $OUTPUT_ARGS \
    --distributed-backend nccl \
    2>&1 | tee logs/train_${logfile}.log

chmod 440 logs/train_${logfile}.log
STEP_TIME=`grep "elapsed time per iteration" logs/train_${logfile}.log | awk -F ':' '{print$5}' | awk -F '|' '{print$1}' | head -n 200 | tail -n 50 | awk '{sum+=$1} END {if (NR != 0) printf("%.1f",sum/NR)}'`
PERF=`awk 'BEGIN{printf "%.3f\n", '${GBS}'*1000/'${STEP_TIME}'}'`
echo "Elapsed Time Per iteration: $STEP_TIME, Average Samples per Second: $PERF"