# 环境配置
1. 创建容器，安装基础依赖包
<pre>
docker run --name=libo_mlperf_test --hostname=node72 --user=root --volume=/mnt/fs/user/llama:/mnt/fs/user/llama --volume=/root/.ssh:/root/.ssh --volume=/opt/hyhal:/opt/hyhal:ro --cap-add=SYS_PTRACE  --shm-size 80g --network=host --privileged  --device /dev/kfd:/dev/kfd --device /sys/kernel/debug/dri:/sys/kernel/debug/dri --device /dev/dri:/dev/dri --runtime=runc --detach=true -t b8d662003405  /bin/bash

docker exec -it libo_mlperf_test bash

cd /mnt/fs/user/llama/custom_model/libo_test/mlperf_test
git clone https://github.com/mlperf/logging.git mlperf-logging
cd mlperf-logging/
pip install -e .

git clone http://10.6.10.68/libo11/mlperf_transformer_v0.7.git



</pre>

2. 数据集下载及处理

<pre>

#（进入目录）
cd /mnt/fs/user/llama/custom_model/libo_test/mlperf_test/transformer/implementations/pytorch
#更改run_preprocessing.sh ：                      
vim run_preprocessing.sh 
#更改为以下内容 ：

#!/bin/bash

set -e

SEED=$1
mkdir -p /workspace/translation/examples/translation/wmt14_en_de
mkdir -p /workspace/translation/examples/translation/wmt14_en_de/utf8

# TODO: Add SEED to process_data.py since this uses a random generator (future PR)
#export PYTHONPATH=/research/transformer/transformer:${PYTHONPATH}
# Add compliance to PYTHONPATH
# export PYTHONPATH=/mlperf/training/compliance:${PYTHONPATH}

cp ./reference_dictionary.ende.txt /workspace/translation/

cd /workspace/translation

cp /workspace/translation/reference_dictionary.ende.txt /workspace/translation/examples/translation/wmt14_en_de/dict.en.txt
cp /workspace/translation/reference_dictionary.ende.txt /workspace/translation/examples/translation/wmt14_en_de/dict.de.txt

sed -i "1s/^/\'<lua_index_compat>\'\n/" /workspace/translation/examples/translation/wmt14_en_de/dict.en.txt
sed -i "1s/^/\'<lua_index_compat>\'\n/" /workspace/translation/examples/translation/wmt14_en_de/dict.de.txt

# TODO: make code consistent to not look in two places (allows temporary hack above for preprocessing-vs-training)
cp /workspace/translation/reference_dictionary.ende.txt /workspace/translation/examples/translation/wmt14_en_de/utf8/dict.en.txt
cp /workspace/translation/reference_dictionary.ende.txt /workspace/translation/examples/translation/wmt14_en_de/utf8/dict.de.txt

#wget https://raw.githubusercontent.com/tensorflow/models/master/official/transformer/test_data/newstest2014.en -O /workspace/translation/examples/translation/wmt14_en_de/newstest2014.en
#wget https://raw.githubusercontent.com/tensorflow/models/master/official/transformer/test_data/newstest2014.de -O /workspace/translation/examples/translation/wmt14_en_de/newstest2014.de

cd /mnt/fs/user/llama/custom_model/libo_test/mlperf_test/transformer/implementations/pytorch
cp ./newstest2014.en /workspace/translation
cp ./newstest2014.de /workspace/translation

cp /workspace/translation/newstest2014.en /workspace/translation/examples/translation/wmt14_en_de/newstest2014.en
cp /workspace/translation/newstest2014.de /workspace/translation/examples/translation/wmt14_en_de/newstest2014.de

python3 preprocess.py --raw_dir /raw_data/ --data_dir /workspace/translation/examples/translation/wmt14_en_de


#数据集处理
chmod +x run_preprocessing.sh 
chmod +x run_conversion.sh
bash run_preprocessing.sh
bash run_conversion.sh#该过程巨慢
#完成后会在/workspace/translation/examples/translation/wmt14_en_de/utf8里生成几个.bin文件，数据集路径按需修改即可

</pre>
# 训练

1. 编译apex
<pre>
#项目中apex适配的版本较老，直接将老版本apex中的optimizers 拷贝到环境中，替换掉原有的包就行

cd /mnt/fs/user/llama/custom_model/libo_test/mlperf_test/transformer
cp ./optimizers.zip /usr/local/lib/python3.10/site-packages/apex/contrib/
cd /usr/local/lib/python3.10/site-packages/apex/contrib
rm -rf optimizers
unzip optimizers.zip
</pre>

2. 训练指令
<pre>
source config_DGX1.sh

bash  ./run_and_time.sh
</pre>


# 截图
训练
![训练](训练成功.png)


验证
![验证](验证成功.png)


# 详细的bug处理及配置过程可参阅：
https://r0ddbu55vzx.feishu.cn/wiki/ETtMwanKhi77M8kZHmhcdD6Anqc

https://r0ddbu55vzx.feishu.cn/wiki/COKQwN0qci5UrpkMdsMckn7In7d