#!/bin/bash
# test_checkpoint_advanced.sh
# Tests 1/2: torch-format synchronous save and load
# Tests 3/4: torch_dist-format asynchronous save and load

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_adv
CKPT_DIR_ASYNC=./ckpt_adv_async
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
"

GPT_ARGS="
    --transformer-impl transformer_engine \
    --tensor-model-parallel-size ${TP} \
    --pipeline-model-parallel-size ${PP} \
    --sequence-parallel \
    --num-layers 4 \
    --train-iters 1000 \
    --hidden-size 1024 \
    --num-attention-heads 16 \
    --seq-length 1024 \
    --max-position-embeddings 1024 \
    --tokenizer-type Llama2Tokenizer \
    --tokenizer-model ${TOKENIZER_MODEL} \
    --micro-batch-size 1 \
    --global-batch-size 4 \
    --lr 1.0e-4 \
    --lr-decay-style cosine \
    --clip-grad 1.0 \
    --weight-decay 0.1 \
    --init-method-std 0.01 \
    --disable-bias-linear \
    --position-embedding-type rope \
    --no-bias-dropout-fusion \
    --attention-dropout 0.0 \
    --hidden-dropout 0.0 \
    --bf16
"

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

OUTPUT_ARGS="
    --log-throughput \
    --log-interval 1 \
    --eval-interval 1000 \
    --eval-iters 10
"

CKPT_SAVE_ARGS="
    --save ${CKPT_DIR} \
    --ckpt-format torch \
    --save-interval 10 \
    --no-save-optim \
    --no-save-rng \
    --ckpt-fully-parallel-save \
"

CKPT_LOAD_ARGS="
    --load ${CKPT_DIR} \
    --ckpt-format torch \
    --no-load-optim \
    --no-load-rng \
"

CKPT_ASYNC_SAVE_ARGS="
    --save ${CKPT_DIR_ASYNC} \
    --ckpt-format torch_dist \
    --save-interval 5 \
    --async-save \
    --use-distributed-optimizer \
    --ckpt-fully-parallel-save \
"

CKPT_ASYNC_LOAD_ARGS="
    --load ${CKPT_DIR_ASYNC} \
    --ckpt-format torch_dist \
    --use-distributed-optimizer \
"

# Test 1: save checkpoint
echo "========================================"
echo "Running checkpoint save test..."
echo "========================================"
torchrun $DISTRIBUTED_ARGS pretrain_gpt.py \
    $GPT_ARGS \
    $DATA_ARGS \
    $OUTPUT_ARGS \
    $CKPT_SAVE_ARGS \
    --exit-interval 10

# Test 2: load checkpoint
echo "========================================"
echo "Running checkpoint load test..."
echo "========================================"
torchrun $DISTRIBUTED_ARGS pretrain_gpt.py \
    $GPT_ARGS \
    $DATA_ARGS \
    $OUTPUT_ARGS \
    $CKPT_LOAD_ARGS \
    --exit-interval 10

# Test 3: torch_dist async checkpoint save
echo "========================================"
echo "Running checkpoint async (torch_dist) save test..."
echo "========================================"
rm -rf ${CKPT_DIR_ASYNC}
torchrun $DISTRIBUTED_ARGS pretrain_gpt.py \
    $GPT_ARGS \
    $DATA_ARGS \
    $OUTPUT_ARGS \
    $CKPT_ASYNC_SAVE_ARGS \
    --exit-interval 15

# Test 4: load the torch_dist async checkpoint and resume training
echo "========================================"
echo "Running checkpoint async (torch_dist) load test..."
echo "========================================"
torchrun $DISTRIBUTED_ARGS pretrain_gpt.py \
    $GPT_ARGS \
    $DATA_ARGS \
    $OUTPUT_ARGS \
    $CKPT_ASYNC_LOAD_ARGS \
    --exit-interval 20

set +x