#!/bin/bash

#=============================================
# Author: LinSHua
# Date: 2026-06-02
# Description：This use case is used to care for TP/PP/VPP pipeline split, distributed optimizer, O2 gradient/optimizer optimization, layer recompute, FA_TND and fused rotary position embedding features.
# Remarks:
#   - Checkpoint "/data/ci/models/Qwen3-8B/mg/qwen3-8b-layer16-tp2pp4vpp2-generated" is generated by hf weights and the num-layers is reduced to 16
#=============================================
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=8
MASTER_ADDR=localhost
MASTER_PORT=6012
NNODES=1
NODE_RANK=0
WORLD_SIZE=$(($NPUS_PER_NODE*$NNODES))

DATA_PATH="/data/ci/datasets/processed/qwen3-8b/alpaca_text_document"
TOKENIZER_PATH="/data/ci/models/Qwen3-8B/hf/Qwen3-8B"
CKPT_LOAD_DIR="/data/ci/models/Qwen3-8B/mg/qwen3-8b-layer16-tp2pp4vpp2-generated"

TP=2
PP=4
CP=1
VPP=2
MBS=1
GBS=16
SEQ_LENGTH=2048
TRAIN_ITERS=15

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 \
    --o2-gradient \
    --o2-optimizer \
    --reuse-fp32-param \
    --overlap-grad-reduce \
    --overlap-param-gather \
    --use-ascend-coc
    --enable-recompute-layers-per-pp-rank
    --recompute-granularity full
    --recompute-method block
    --recompute-num-layers 1
"

TRAIN_ARGS="
    --transformer-impl local \
    --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} \
    --num-layers-per-virtual-pipeline-stage ${VPP} \
    --context-parallel-size ${CP} \
    --context-parallel-algo megatron_cp_algo \
    --use-fused-ring-attention-update \
    --attention-mask-type general \
    --sequence-parallel
"

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 16 \
    --hidden-size 2048 \
    --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 \
    --ckpt-format torch
"

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} \
    --distributed-backend nccl