#!/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=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=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} \
    --num-layers-per-virtual-pipeline-stage 1 \
    --use-flash-attn \
    --num-layers 4 \
    --hidden-size 2048 \
    --ffn-hidden-size 5504 \
    --num-attention-heads 16 \
    --tokenizer-type Llama2Tokenizer \
    --tokenizer-model ${TOKENIZER_MODEL} \
    --seq-length 4096 \
    --max-position-embeddings 4096 \
    --micro-batch-size 1 \
    --global-batch-size 4 \
    --make-vocab-size-divisible-by 1 \
    --lr 1.0e-6 \
    --train-iters 500 \
    --lr-decay-style cosine \
    --untie-embeddings-and-output-weights \
    --attention-dropout 0.0 \
    --init-method-std 0.01 \
    --hidden-dropout 0.0 \
    --normalization RMSNorm \
    --norm-epsilon 1e-6 \
    --swiglu \
    --no-masked-softmax-fusion \
    --attention-softmax-in-fp32 \
    --min-lr 1.0e-7 \
    --weight-decay 0.1 \
    --clip-grad 1.0 \
    --adam-beta1 0.9 \
    --initial-loss-scale 4096.0 \
    --adam-beta2 0.95 \
    --adam-eps 1e-5 \
    --disable-bias-linear \
    --num-query-groups 8 \
    --lr-warmup-fraction 0.01 \
    --bf16 \
    --seed 42
"

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

OUTPUT_ARGS="
    --log-throughput \
    --log-interval 1 \
    --log-params-norm \
    --log-progress \
    --timing-log-level 0 \
    --logging-level 20 \
    --log-num-zeros-in-grad \
    --save-interval 10000 \
    --eval-interval 50 \
    --eval-iters 10 \
"

export NVTE_APPLY_QK_LAYER_SCALING=1
ATTN_ARGS="
    --q-lora-rank 1536 \
    --kv-lora-rank 512 \
    --qk-head-dim 128 \
    --qk-pos-emb-head-dim 64 \
    --v-head-dim 128 \
    --attention-backend auto \
    --apply-query-key-layer-scaling \
    --qk-layernorm \
    --relative-attention-num-buckets 8 \
    --relative-attention-max-distance 128 \
"

GQA_ARGS="
    --group-query-attention \

"

MLA_ARGS="
    --multi-latent-attention \
"

DP_ARGS="
    --data-parallel-sharding-strategy no_shard \
    --data-parallel-random-init \
    --ddp-num-buckets 8 \
    --ddp-pad-buckets-for-high-nccl-busbw \
    --ddp-average-in-collective \
"

EMBEDDINGS_ARGS="
    --position-embedding-type rope \
    --use-rotary-position-embeddings \
    --rotary-base 10000 \
    --rotary-percent 1.0 \
    --rotary-seq-len-interpolation-factor 1 \
    --use-rope-scaling \
    --rope-scaling-factor 8.0 \
    --no-rope-fusion \
    --rotary-scaling-factor 1.0 \
    --mscale 1.0 \
    --mscale-all-dim 1.0 \
"

COMM_ARGS="
    --use-tp-pp-dp-mapping \
    --overlap-grad-reduce \
    --overlap-param-gather \
    --overlap-param-gather-with-optimizer-step \
"

DIST_ARGS="
    --use-distributed-optimizer \
    --distribute-saved-activations \
    --distributed-backend nccl \
    --distributed-timeout-minutes 3 \
    --ckpt-format torch \
    --use-sharp \
    --sharp-enabled-group dp \
    --create-all-gather-group \
"

RECOMPUTE_ARGS="
    --recompute-granularity full \
    --recompute-method block \
    --recompute-num-layers 4 \
"

DECAY_ARGS="
    --wgrad-deferral-limit 1 \
    --weight-decay-incr-style cosine \
    --start-weight-decay 1 \
    --end-weight-decay 5 \
"

STRAGGLER_ARGS="
    --straggler-ctrlr-port 6012 \
    --straggler-minmax-count 1 \
    --disable-straggler-on-startup \
"

FUSION_ARGS="
    --no-gradient-reduce-div-fusion \
    --no-bias-gelu-fusion \
    --no-bias-swiglu-fusion \
    --cross-entropy-loss-fusion \
    --cross-entropy-fusion-impl native \
    --use-fused-weighted-squared-relu \
"

# GQA
torchrun $DISTRIBUTED_ARGS pretrain_gpt.py \
    $GPT_ARGS \
    $DATA_ARGS \
    $OUTPUT_ARGS \
    $ATTN_ARGS \
    $GQA_ARGS \
    $DP_ARGS \
    $EMBEDDINGS_ARGS \
    $COMM_ARGS \
    $DIST_ARGS \
    $RECOMPUTE_ARGS \
    $DECAY_ARGS \
    $STRAGGLER_ARGS \
    $FUSION_ARGS \
    --exit-interval 10 \

# MLA
torchrun $DISTRIBUTED_ARGS pretrain_gpt.py \
    $GPT_ARGS \
    $DATA_ARGS \
    $OUTPUT_ARGS \
    $ATTN_ARGS \
    $MLA_ARGS \
    $DP_ARGS \
    $EMBEDDINGS_ARGS \
    $COMM_ARGS \
    $DIST_ARGS \
    $RECOMPUTE_ARGS \
    $DECAY_ARGS \
    $STRAGGLER_ARGS \
    $FUSION_ARGS \
    --exit-interval 10 \

set +x