#!/bin/bash

export CUDA_DEVICE_MAX_CONNECTIONS=1
source "tests_extend/system_tests/env_npu.sh"

NPUS_PER_NODE=8
MASTER_ADDR=localhost
MASTER_PORT=6001
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=1
CP=4
EP=1

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

RECOMPUTE_ARGS="
    --enable-recompute-layers-per-pp-rank \
    --recompute-activation-function \
    --swap-attention \
    --recompute-num-layers 0 \
"

GPT_ARGS="
    --yaml-cfg "./tests_extend/system_tests/yaml_args_example/example.yaml"\
    --use-mcore-models
"

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

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

torchrun $DISTRIBUTED_ARGS pretrain_gpt.py \
    $GPT_ARGS \
    $RECOMPUTE_ARGS \
    $DATA_ARGS \
    $OUTPUT_ARGS \
    --distributed-backend nccl \

set +x