#!/bin/bash
set -e
source /usr/local/Ascend/ascend-toolkit/set_env.sh
# 该变量只用于规避megatron对其校验，对npu无效
export CUDA_DEVICE_MAX_CONNECTIONS=1
MASTER_ADDR=localhost
MASTER_PORT=6100
NNODES=1
NODE_RANK=0
NPUS_PER_NODE=1
WORLD_SIZE=$(($NPUS_PER_NODE * $NNODES))

TP=1
PP=1
CP=1
MBS=1
GBS=$(($WORLD_SIZE*$MBS/$CP/$TP))
BASEPATH=$(cd `dirname $0`; cd ../../../; pwd)

echo "LOCATION: $LOCATION"
echo "BASEPATH: $BASEPATH"

PROMPT_PATH="$BASEPATH/tests/st/run_configs/inference_cogvideox_t2v_1_5/samples_prompts.txt"
MM_MODEL="$BASEPATH/tests/st/run_configs/inference_cogvideox_t2v_1_5/inference_model_t2v_1.5.json"
LOAD_PATH="./"

sed -i "s^{{prompt_path}}^$PROMPT_PATH^g" $MM_MODEL

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

GPT_ARGS="
    --tensor-model-parallel-size ${TP} \
    --pipeline-model-parallel-size ${PP} \
    --context-parallel-size ${CP} \
    --micro-batch-size ${MBS} \
    --global-batch-size ${GBS} \
    --lr 2e-5 \
    --min-lr 2e-5 \
    --train-iters 5010 \
    --weight-decay 0 \
    --clip-grad 1 \
    --adam-beta1 0.9 \
    --adam-beta2 0.999 \
    --no-gradient-accumulation-fusion \
    --no-load-optim \
    --no-load-rng \
    --no-save-optim \
    --no-save-rng \
    --fp16 \
    --load $LOAD_PATH \
"

torchrun $DISTRIBUTED_ARGS $BASEPATH/inference_sora.py $MM_ARGS $GPT_ARGS