#!/bin/bash

export CUDA_DEVICE_MAX_CONNECTIONS=1
source "test/system_tests/env_npu.sh"
export STREAMS_PER_DEVICE=32

NPUS_PER_NODE=8
MASTER_ADDR=localhost
MASTER_PORT=6000
NNODES=1
NODE_RANK=0
WORLD_SIZE=$(($NPUS_PER_NODE*$NNODES))

CKPT_DIR=./ckpt_llama
DATA_PATH="/home/dataset/llama2/alpaca_text_document"
TOKENIZER_MODEL="/home/dataset/model/llama-2-7b-hf/tokenizer.model"

TP=2
PP=2
CP=1
EP=1

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

# Activation function
ACT_ARGS="
    --openai-gelu
"

# Mixed precision
MIX_PRC_ARGS="
    --fp16 \
    --accumulate-allreduce-grads-in-fp32 \
    --attention-softmax-in-fp32 \
    --fp16-lm-cross-entropy \
"

# Normalization
NORM_ARGS="
    --normalization RMSNorm \
"

# Network size
NET_SIZE_ARGS="
    --hidden-size 1024 \
    --ffn-hidden-size 4000 \
    --num-layers 4 \
    --num-attention-heads 16 \
    --num-query-groups 8 \
    --kv-channels 64
"

# Loss
LOSS_ARGS="
    --initial-loss-scale 4096 \
    --loss-scale-window 200 \
    --min-loss-scale 1.0 \
    --hysteresis 2
"

# Learning rate schedule
LR_SCHED_ARGS="
    --lr 1.0e-2 \
    --min-lr 1.0e-7 \
    --lr-decay-style cosine \
    --lr-warmup-fraction 0.01
"

# Parallelism
PARAL_ARGS="
    --seq-length 2048 \
    --sequence-parallel \
    --context-parallel-size ${CP} \
    --tensor-model-parallel-size ${TP} \
    --pipeline-model-parallel-size ${PP} \
    --num-layers-per-virtual-pipeline-stage 1 \
"

BASE_ARGS="
    --transformer-impl transformer_engine \
    --q-lora-rank 1536 \
    --kv-lora-rank 512 \
    --qk-head-dim 128 \
    --qk-pos-emb-head-dim 64 \
    --v-head-dim 128 \
    --use-flash-attn \
    --tokenizer-type Llama2Tokenizer \
    --tokenizer-model ${TOKENIZER_MODEL} \
    --max-position-embeddings 4096 \
    --micro-batch-size 1 \
    --global-batch-size 4 \
    --make-vocab-size-divisible-by 1 \
    --train-iters 1000 \
    --untie-embeddings-and-output-weights \
    --attention-dropout 0.0 \
    --init-method-std 0.01 \
    --hidden-dropout 0.0 \
    --position-embedding-type rope \
    --no-masked-softmax-fusion \
    --weight-decay 0.1 \
    --clip-grad 1.0 \
    --adam-beta1 0.9 \
    --adam-beta2 0.95 \
    --adam-eps 1e-5 \
    --group-query-attention \
    --seed 42
"

DATA_ARGS="
    --data-path ${DATA_PATH} \
    --split 949,50,1
"

OUTPUT_ARGS="
    --log-throughput \
    --log-interval 1 \
    --save-interval 10000 \
    --eval-interval 50 \
    --eval-iters 10 \
"

torchrun $DISTRIBUTED_ARGS pretrain_gpt.py \
    $ACT_ARGS \
    $MIX_PRC_ARGS \
    $NET_SIZE_ARGS \
    $NORM_ARGS \
    $LOSS_ARGS \
    $LR_SCHED_ARGS \
    $PARAL_ARGS \
    $BASE_ARGS \
    $DATA_ARGS \
    $OUTPUT_ARGS \
    --exit-interval 10

set +x