#!/bin/bash

export HCCL_CONNECT_TIMEOUT=1800
export CUDA_DEVICE_MAX_CONNECTIONS=1
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export NPU_ASD_ENABLE=0
export TASK_QUEUE_ENABLE=2

NPUS_PER_NODE=2
MASTER_ADDR=localhost
MASTER_PORT=6000
NNODES=1
NODE_RANK=0
WORLD_SIZE=$(($NPUS_PER_NODE*$NNODES))

# please fill these path configurations
CKPT_LOAD_DIR="/data/mindspeed_test/model_weights/Qwen3-0.6B_convert/"
CKPT_SAVE_DIR="/data/mindspeed_test/model_weights/Qwen3-0.6B_train_260528_practice/"
DATA_PATH="/data/mindspeed_test/datasets/alpaca_text_document.bin"     # �~U��~M��~[~F路�~D�~L填�~E��~U��~M��~D�~D�~P~F�~W��~]�~X�~Z~D�~U��~M�路�~D�~L注�~D~O�~\~@覦
TOKENIZER_PATH="/data/mindspeed_test/model_weights/Qwen3-0.6B/"     # �~M表路�~D�~L填�~E��~K载�~Z~D�~@�~P�~]~C�~G~M�~M表路�~D

TP=1                # �~]~C�~G~M转�~M�设置--target-tensor-parallel-size 1�~L修�~T�为1
PP=2                # �~]~C�~G~M转�~M�设置--target-pipeline-parallel-size 2�~L修�~T�为2�~L�~N�~]~C�~G~M转�~M��~W��~@�~G�
SEQ_LENGTH=4096     # 设置seq_length为4096
MBS=1               # 设置micro-batch-size为1
GBS=64              # 设置global-batch-size为64
TRAIN_ITERS=30    # 设置训�~C迭代步�~U�步�~U��

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

OPTIMIZE_ARGS="
    --use-flash-attn \
    --use-fused-rotary-pos-emb \
    --use-rotary-position-embeddings \
    --use-fused-swiglu \
    --use-fused-rmsnorm \
    --no-masked-softmax-fusion \
    --use-distributed-optimizer \
    --reuse-fp32-param \
    --overlap-grad-reduce \
    --overlap-param-gather \
    --use-ascend-coc
"

TRAIN_ARGS="
    --micro-batch-size ${MBS} \
    --global-batch-size ${GBS} \
    --lr 1.25e-6 \
    --lr-decay-style cosine \
    --min-lr 1.25e-7 \
    --weight-decay 1e-1 \
    --lr-warmup-fraction 0.01 \
    --attention-dropout 0.0 \
    --init-method-std 0.01 \
    --hidden-dropout 0.0 \
    --clip-grad 1.0 \
    --adam-beta1 0.9 \
    --adam-beta2 0.95 \
    --initial-loss-scale 4096 \
    --seed 42 \
    --bf16 \
    --train-iters ${TRAIN_ITERS} \
    --seq-length ${SEQ_LENGTH}
"

MODEL_PARALLEL_ARGS="
    --tensor-model-parallel-size ${TP} \
    --pipeline-model-parallel-size ${PP} \
"

GPT_ARGS="
    --use-mcore-models \
    --spec mindspeed_llm.tasks.models.spec.qwen3_spec layer_spec \
    --qk-layernorm \
    --tokenizer-name-or-path ${TOKENIZER_PATH} \
    --max-position-embeddings ${SEQ_LENGTH} \
    --num-layers 36 \
    --hidden-size 4096 \
    --ffn-hidden-size 12288 \
    --num-attention-heads 32 \
    --tokenizer-type PretrainedFromHF \
    --make-vocab-size-divisible-by 1 \
    --padded-vocab-size 151936 \
    --rotary-base 1000000 \
    --untie-embeddings-and-output-weights \
    --disable-bias-linear \
    --position-embedding-type rope \
    --normalization RMSNorm \
    --swiglu \
    --attention-softmax-in-fp32 \
    --no-gradient-accumulation-fusion \
    --group-query-attention \
    --num-query-groups 8 \
    --norm-epsilon 1e-6
"

DATA_ARGS="
    --data-path $DATA_PATH \
    --split 100,0,0
"

OUTPUT_ARGS="
    --log-interval 1 \
    --save-interval ${TRAIN_ITERS} \
    --eval-interval ${TRAIN_ITERS} \
    --eval-iters 0 \
    --no-load-optim \
    --no-load-rng
"

torchrun $DISTRIBUTED_ARGS pretrain_gpt.py \
    $GPT_ARGS \
    $DATA_ARGS \
    $MOE_ARGS \
    $OUTPUT_ARGS \
    $OPTIMIZE_ARGS \
    $TRAIN_ARGS \
    $MODEL_PARALLEL_ARGS \
    --load ${CKPT_LOAD_DIR} \
    --save ${CKPT_SAVE_DIR} \
    --distributed-backend nccl \
    --transformer-impl local \
    | tee logs/train_mcore_qwen3_8b.log