Merge branch 'hepj-test' into 'main'

修改README，增加训练脚本，完善模型转换代码

See merge request dcutoolkit/deeplearing/dlexamples_new!38

PyTorch/NLP/BERT/single_pre2_1.sh

+#!/bin/bash
+#export MIOPEN_DEBUG_DISABLE_FIND_DB=1
+#export NCCL_SOCKET_IFNAME=ib0
+#export HSA_USERPTR_FOR_PAGED_MEM=0
+#export MIOPEN_DEBUG_DISABLE_FIND_DB=1
+#export NCCL_SOCKET_IFNAME=eno1
+export HSA_FORCE_FINE_GRAIN_PCIE=1
+#source /public/software/apps/DeepLearning/PyTorch/pytorch-env.sh
+export MIOPEN_FIND_MODE=1
+#export MIOPEN_ENABLE_LOGGING_CMD=1
+#export ROCBLAS_LAYER=3
+module unload compiler/rocm/2.9
+#export MIOPEN_DEBUG_CONV_DIRECT_ASM_3X3U=0
+#export MIOPEN_DEBUG_CONV_DIRECT_ASM_WRW3X3=0
+echo "MIOPEN_FIND_MODE=$MIOPEN_FIND_MODE"
+#module load apps/PyTorch/1.5.0a0/hpcx-2.4.1-gcc-7.3.1-rocm3.3
+#source /public/home/aiss/Pytorch/env_rocm3.3_torch1.5.sh
+lrank=$OMPI_COMM_WORLD_LOCAL_RANK
+comm_rank=$OMPI_COMM_WORLD_RANK
+comm_size=$OMPI_COMM_WORLD_SIZE
+
+#下边是修改的
+
+
+#export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+#HSA_FORCE_FINE_GRAIN_PCIE=1 numactl --cpunodebind=4,5,6,7 --membind=4,5,6,7
+
+#module load compiler/rocm/3.9.1
+
+export PATH_PHRASE2=/public/software/apps/DeepLearning/Data/wikicorpus_en/lower_case_1_seq_len_512_max_pred_80_masked_lm_prob_0.15_random_seed_12345_dupe_factor_5_shard_1472_test_split_10/wikicorpus_en/training
+
+APP="python3 run_pretraining_v1.py  \
+    --input_dir=${PATH_PHRASE2}    \
+    --output_dir=/public/home/hepj/outdir/torch/pre_wiki/phrase2/fp32 \
+    --config_file=/public/home/hepj/model_source/pytorch_bert/bert_config.json \
+    --bert_model=bert-large-uncased \
+    --train_batch_size=2 \
+    --max_seq_length=512 \
+    --max_predictions_per_seq=80 \
+    --max_steps=400000 \
+    --warmup_proportion=0128. \
+    --num_steps_per_checkpoint=20000 \
+    --learning_rate=4.0e-3 \
+    --seed=12439 \
+    --gradient_accumulation_steps=1 \
+    --allreduce_post_accumulation \
+    --do_train \
+    --phase2 \
+    --phase1_end_step=0 \
+    --gpus_per_node 1 \
+    --local_rank -1 \
+    --json-summary /public/home/hepj/outdir/torch/pre_wiki/phrase2/fp32/dllogger_dtk22.04.json
+
+ "
+#--fp16  \
+# --amp \
+
+
+#--json-summary  /public/home/hepj/out_dir/tourch/SQuAD/results.json 
+
+
+case ${lrank} in
+[0])
+  export HIP_VISIBLE_DEVICES=0
+  export UCX_NET_DEVICES=mlx5_0:1
+  export UCX_IB_PCI_BW=mlx5_0:50Gbs
+  echo numactl --cpunodebind=0 --membind=0 ${APP}
+  numactl --cpunodebind=0 --membind=0 ${APP}
+
+  #echo GLOO_SOCKET_IFNAME=ib0 numactl --cpunodebind=0 --membind=0 ${APP} 
+  #GLOO_SOCKET_IFNAME=ib0 numactl --cpunodebind=0 --membind=0 ${APP}
+  ;;
+[1])
+  export HIP_VISIBLE_DEVICES=1
+  export UCX_NET_DEVICES=mlx5_1:1
+  export UCX_IB_PCI_BW=mlx5_1:50Gbs
+  echo numactl --cpunodebind=1 --membind=1 ${APP}
+  numactl --cpunodebind=1 --membind=1 ${APP}
+  ;;
+[2])
+  export HIP_VISIBLE_DEVICES=2
+  export UCX_NET_DEVICES=mlx5_2:1
+  export UCX_IB_PCI_BW=mlx5_2:50Gbs
+  echo numactl --cpunodebind=2 --membind=2 ${APP} 
+  numactl --cpunodebind=2 --membind=2 ${APP}
+  ;;
+[3])
+  export HIP_VISIBLE_DEVICES=3
+  export UCX_NET_DEVICES=mlx5_3:1
+  export UCX_IB_PCI_BW=mlx5_3:50Gbs
+  echo numactl --cpunodebind=3 --membind=3 ${APP}
+  numactl --cpunodebind=3 --membind=3 ${APP}
+  ;;
+esac

PyTorch/NLP/BERT/single_pre2_1_fp16.sh

+#!/bin/bash
+export HSA_FORCE_FINE_GRAIN_PCIE=1
+export MIOPEN_FIND_MODE=1
+#export MIOPEN_ENABLE_LOGGING_CMD=1
+#export ROCBLAS_LAYER=3
+module unload compiler/rocm/2.9
+echo "MIOPEN_FIND_MODE=$MIOPEN_FIND_MODE"
+lrank=$OMPI_COMM_WORLD_LOCAL_RANK
+comm_rank=$OMPI_COMM_WORLD_RANK
+comm_size=$OMPI_COMM_WORLD_SIZE
+
+#下边是修改的
+
+export PATH_PHRASE2=/public/software/apps/DeepLearning/Data/wikicorpus_en/lower_case_1_seq_len_512_max_pred_80_masked_lm_prob_0.15_random_seed_12345_dupe_factor_5_shard_1472_test_split_10/wikicorpus_en/training
+
+APP="python3 run_pretraining_v1.py  \
+    --input_dir=${PATH_PHRASE2}    \
+    --output_dir=/public/home/hepj/outdir/torch/pre_wiki/phrase2/fp16 \
+    --config_file=/public/home/hepj/model_source/pytorch_bert/bert_config.json \
+    --bert_model=bert-large-uncased \
+    --train_batch_size=2 \
+    --max_seq_length=512 \
+    --max_predictions_per_seq=80 \
+    --max_steps=400000 \
+    --warmup_proportion=0128. \
+    --num_steps_per_checkpoint=20000 \
+    --learning_rate=4.0e-3 \
+    --seed=12439 \
+    --gradient_accumulation_steps=1 \
+    --allreduce_post_accumulation \
+    --do_train \
+    --fp16 \
+    --amp \
+    --phase2 \
+    --phase1_end_step=0 \
+    --gpus_per_node 1 \
+    --local_rank -1 \
+    --json-summary /public/home/hepj/outdir/torch/pre_wiki/phrase2/fp16/dllogger_dtk22.04.json
+
+ "
+
+
+#--json-summary  /public/home/hepj/out_dir/tourch/SQuAD/results.json 
+
+
+case ${lrank} in
+[0])
+  export HIP_VISIBLE_DEVICES=0
+  export UCX_NET_DEVICES=mlx5_0:1
+  export UCX_IB_PCI_BW=mlx5_0:50Gbs
+  echo numactl --cpunodebind=0 --membind=0 ${APP}
+  numactl --cpunodebind=0 --membind=0 ${APP}
+
+  #echo GLOO_SOCKET_IFNAME=ib0 numactl --cpunodebind=0 --membind=0 ${APP} 
+  #GLOO_SOCKET_IFNAME=ib0 numactl --cpunodebind=0 --membind=0 ${APP}
+  ;;
+[1])
+  export HIP_VISIBLE_DEVICES=1
+  export UCX_NET_DEVICES=mlx5_1:1
+  export UCX_IB_PCI_BW=mlx5_1:50Gbs
+  echo numactl --cpunodebind=1 --membind=1 ${APP}
+  numactl --cpunodebind=1 --membind=1 ${APP}
+  ;;
+[2])
+  export HIP_VISIBLE_DEVICES=2
+  export UCX_NET_DEVICES=mlx5_2:1
+  export UCX_IB_PCI_BW=mlx5_2:50Gbs
+  echo numactl --cpunodebind=2 --membind=2 ${APP} 
+  numactl --cpunodebind=2 --membind=2 ${APP}
+  ;;
+[3])
+  export HIP_VISIBLE_DEVICES=3
+  export UCX_NET_DEVICES=mlx5_3:1
+  export UCX_IB_PCI_BW=mlx5_3:50Gbs
+  echo numactl --cpunodebind=3 --membind=3 ${APP}
+  numactl --cpunodebind=3 --membind=3 ${APP}
+  ;;
+esac

PyTorch/NLP/BERT/single_pre2_4.sh

+#!/bin/bash
+export HSA_FORCE_FINE_GRAIN_PCIE=1
+export MIOPEN_FIND_MODE=1
+module unload compiler/rocm/2.9
+echo "MIOPEN_FIND_MODE=$MIOPEN_FIND_MODE"
+lrank=$OMPI_COMM_WORLD_LOCAL_RANK
+comm_rank=$OMPI_COMM_WORLD_RANK
+comm_size=$OMPI_COMM_WORLD_SIZE
+export HIP_VISIBLE_DEVICES=0,1,2,3
+
+export PATH_PHRASE2=/public/software/apps/DeepLearning/Data/wikicorpus_en/lower_case_1_seq_len_512_max_pred_80_masked_lm_prob_0.15_random_seed_12345_dupe_factor_5_shard_1472_test_split_10/wikicorpus_en/training
+
+APP="python3 run_pretraining_v4.py  \
+    --input_dir=${PATH_PHRASE2}    \
+    --output_dir=/public/home/hepj/outdir/tourch/pre_wiki4/phrase2/fp32 \
+    --config_file=/public/home/hepj/model_source/pytorch_bert/bert_config.json \
+    --bert_model=bert-large-uncased \
+    --train_batch_size=2 \
+    --max_seq_length=512 \
+    --max_predictions_per_seq=80 \
+    --max_steps=400000 \
+    --warmup_proportion=0128. \
+    --num_steps_per_checkpoint=200000 \
+    --learning_rate=4.0e-3 \
+    --seed=12439 \
+    --gradient_accumulation_steps=1 \
+    --allreduce_post_accumulation \
+    --do_train \
+    --phase2 \
+    --phase1_end_step 0 \
+    --local_rank ${comm_rank} \
+    --world_size 4 \
+    --gpus_per_node 1 \
+    --dist_url tcp://localhost:34567 \
+    --json-summary /public/home/hepj/outdir/torch/pre_wiki4/phrase2/fp32/dllogger.json
+ "
+#--fp16  \
+# --amp \
+
+
+#--json-summary  /public/home/hepj/out_dir/tourch/SQuAD/results.json 
+
+
+case ${lrank} in
+[0])
+  export HIP_VISIBLE_DEVICES=0
+  export UCX_NET_DEVICES=mlx5_0:1
+  export UCX_IB_PCI_BW=mlx5_0:50Gbs
+  echo numactl --cpunodebind=0 --membind=0 ${APP}
+  numactl --cpunodebind=0 --membind=0 ${APP}
+
+  #echo GLOO_SOCKET_IFNAME=ib0 numactl --cpunodebind=0 --membind=0 ${APP} 
+  #GLOO_SOCKET_IFNAME=ib0 numactl --cpunodebind=0 --membind=0 ${APP}
+  ;;
+[1])
+  export HIP_VISIBLE_DEVICES=1
+  export UCX_NET_DEVICES=mlx5_1:1
+  export UCX_IB_PCI_BW=mlx5_1:50Gbs
+  echo numactl --cpunodebind=1 --membind=1 ${APP}
+  numactl --cpunodebind=1 --membind=1 ${APP}
+  ;;
+[2])
+  export HIP_VISIBLE_DEVICES=2
+  export UCX_NET_DEVICES=mlx5_2:1
+  export UCX_IB_PCI_BW=mlx5_2:50Gbs
+  echo numactl --cpunodebind=2 --membind=2 ${APP} 
+  numactl --cpunodebind=2 --membind=2 ${APP}
+  ;;
+[3])
+  export HIP_VISIBLE_DEVICES=3
+  export UCX_NET_DEVICES=mlx5_3:1
+  export UCX_IB_PCI_BW=mlx5_3:50Gbs
+  echo numactl --cpunodebind=3 --membind=3 ${APP}
+  numactl --cpunodebind=3 --membind=3 ${APP}
+  ;;
+esac

PyTorch/NLP/BERT/single_pre2_4_fp16.sh

+#!/bin/bash
+export HSA_FORCE_FINE_GRAIN_PCIE=1
+export MIOPEN_FIND_MODE=1
+module unload compiler/rocm/2.9
+echo "MIOPEN_FIND_MODE=$MIOPEN_FIND_MODE"
+lrank=$OMPI_COMM_WORLD_LOCAL_RANK
+comm_rank=$OMPI_COMM_WORLD_RANK
+comm_size=$OMPI_COMM_WORLD_SIZE
+
+#下边是修改的
+export HIP_LAUNCH_BLOCKING=1
+
+export HIP_VISIBLE_DEVICES=0,1,2,3
+
+export PATH_PHRASE2=/public/software/apps/DeepLearning/Data/wikicorpus_en/lower_case_1_seq_len_512_max_pred_80_masked_lm_prob_0.15_random_seed_12345_dupe_factor_5_shard_1472_test_split_10/wikicorpus_en/training
+
+APP="python3 run_pretraining_v4.py  \
+    --input_dir=${PATH_PHRASE2}    \
+    --output_dir=/public/home/hepj/outdir/tourch/pre_wiki4/phrase2/fp16 \
+    --config_file=/public/home/hepj/model_source/pytorch_bert/bert_config.json \
+    --bert_model=bert-large-uncased \
+    --train_batch_size=2 \
+    --max_seq_length=512 \
+    --max_predictions_per_seq=80 \
+    --max_steps=40000 \
+    --warmup_proportion=0128. \
+    --num_steps_per_checkpoint=10000 \
+    --learning_rate=4.0e-3 \
+    --seed=12439 \
+    --gradient_accumulation_steps=1 \
+    --allreduce_post_accumulation \
+    --do_train \
+    --phase2 \
+    --phase1_end_step 0 \
+    --local_rank ${comm_rank} \
+    --world_size 4 \
+    --gpus_per_node 1 \
+    --fp16 \
+    --amp \
+    --dist_url tcp://localhost:34567 \
+    --json-summary /public/home/hepj/outdir/torch/pre_wiki4/phrase2/fp16/dllogger.json
+ "
+
+case ${lrank} in
+[0])
+  export HIP_VISIBLE_DEVICES=0
+  export UCX_NET_DEVICES=mlx5_0:1
+  export UCX_IB_PCI_BW=mlx5_0:50Gbs
+  echo numactl --cpunodebind=0 --membind=0 ${APP}
+  numactl --cpunodebind=0 --membind=0 ${APP}
+
+  #echo GLOO_SOCKET_IFNAME=ib0 numactl --cpunodebind=0 --membind=0 ${APP} 
+  #GLOO_SOCKET_IFNAME=ib0 numactl --cpunodebind=0 --membind=0 ${APP}
+  ;;
+[1])
+  export HIP_VISIBLE_DEVICES=1
+  export UCX_NET_DEVICES=mlx5_1:1
+  export UCX_IB_PCI_BW=mlx5_1:50Gbs
+  echo numactl --cpunodebind=1 --membind=1 ${APP}
+  numactl --cpunodebind=1 --membind=1 ${APP}
+  ;;
+[2])
+  export HIP_VISIBLE_DEVICES=2
+  export UCX_NET_DEVICES=mlx5_2:1
+  export UCX_IB_PCI_BW=mlx5_2:50Gbs
+  echo numactl --cpunodebind=2 --membind=2 ${APP} 
+  numactl --cpunodebind=2 --membind=2 ${APP}
+  ;;
+[3])
+  export HIP_VISIBLE_DEVICES=3
+  export UCX_NET_DEVICES=mlx5_3:1
+  export UCX_IB_PCI_BW=mlx5_3:50Gbs
+  echo numactl --cpunodebind=3 --membind=3 ${APP}
+  numactl --cpunodebind=3 --membind=3 ${APP}
+  ;;
+esac

PyTorch/NLP/BERT/single_squad.sh

+#!/bin/bash
+#export MIOPEN_DEBUG_DISABLE_FIND_DB=1
+#export NCCL_SOCKET_IFNAME=ib0
+#export HSA_USERPTR_FOR_PAGED_MEM=0
+#export MIOPEN_DEBUG_DISABLE_FIND_DB=1
+#export NCCL_SOCKET_IFNAME=eno1
+export HSA_FORCE_FINE_GRAIN_PCIE=1
+#source /public/software/apps/DeepLearning/PyTorch/pytorch-env.sh
+export MIOPEN_FIND_MODE=3
+#export MIOPEN_ENABLE_LOGGING_CMD=1
+#export ROCBLAS_LAYER=3
+module unload compiler/rocm/2.9
+#export MIOPEN_DEBUG_CONV_DIRECT_ASM_3X3U=0
+#export MIOPEN_DEBUG_CONV_DIRECT_ASM_WRW3X3=0
+echo "MIOPEN_FIND_MODE=$MIOPEN_FIND_MODE"
+#module load apps/PyTorch/1.5.0a0/hpcx-2.4.1-gcc-7.3.1-rocm3.3
+#source /public/home/aiss/Pytorch/env_rocm3.3_torch1.5.sh
+lrank=$OMPI_COMM_WORLD_LOCAL_RANK
+comm_rank=$OMPI_COMM_WORLD_RANK
+comm_size=$OMPI_COMM_WORLD_SIZE
+
+
+
+#module load compiler/rocm/3.9.1
+
+
+
+
+APP="python3 run_squad_v1.py \
+  --train_file  /public/home/hepj/data/sq1.1/train-v1.1.json \
+  --predict_file  /public/home/hepj/data/sq1.1/dev-v1.1.json \
+  --init_checkpoint  /public/home/hepj/model_source/pytorch_bert/model.ckpt-28252.pt \
+  --vocab_file  /public/home/hepj/model_source/pytorch_bert/vocab.txt \
+  --output_dir  /public/home/hepj/outdir/torch/SQuAD \
+  --config_file  /public/home/hepj/model_source/pytorch_bert/bert_config.json \
+  --json-summary  /public/home/hepj/outdir/torch/SQuAD/results.json \
+  --bert_model bert-large-uncased \
+  --do_train \
+  --do_predict \
+  --train_batch_size  4 \
+  --predict_batch_size 4 \
+  --gpus_per_node  1 \
+  --local_rank -1 \
+  --eval_script /public/home/hepj/squad_eval/evaluate-v1.1.py
+ "
+#--json-summary  /public/home/hepj/out_dir/tourch/SQuAD/results.json 
+
+
+case ${lrank} in
+[0])
+  export HIP_VISIBLE_DEVICES=0
+  export UCX_NET_DEVICES=mlx5_0:1
+  export UCX_IB_PCI_BW=mlx5_0:50Gbs
+  echo numactl --cpunodebind=0 --membind=0 ${APP}
+  numactl --cpunodebind=0 --membind=0 ${APP}
+
+  #echo GLOO_SOCKET_IFNAME=ib0 numactl --cpunodebind=0 --membind=0 ${APP} 
+  #GLOO_SOCKET_IFNAME=ib0 numactl --cpunodebind=0 --membind=0 ${APP}
+  ;;
+[1])
+  export HIP_VISIBLE_DEVICES=1
+  export UCX_NET_DEVICES=mlx5_1:1
+  export UCX_IB_PCI_BW=mlx5_1:50Gbs
+  echo numactl --cpunodebind=1 --membind=1 ${APP}
+  numactl --cpunodebind=1 --membind=1 ${APP}
+  ;;
+[2])
+  export HIP_VISIBLE_DEVICES=2
+  export UCX_NET_DEVICES=mlx5_2:1
+  export UCX_IB_PCI_BW=mlx5_2:50Gbs
+  echo numactl --cpunodebind=2 --membind=2 ${APP} 
+  numactl --cpunodebind=2 --membind=2 ${APP}
+  ;;
+[3])
+  export HIP_VISIBLE_DEVICES=3
+  export UCX_NET_DEVICES=mlx5_3:1
+  export UCX_IB_PCI_BW=mlx5_3:50Gbs
+  echo numactl --cpunodebind=3 --membind=3 ${APP}
+  numactl --cpunodebind=3 --membind=3 ${APP}
+  ;;
+esac

PyTorch/NLP/BERT/single_squad4.sh

+#!/bin/bash
+#export MIOPEN_DEBUG_DISABLE_FIND_DB=1
+#export NCCL_SOCKET_IFNAME=ib0
+#export HSA_USERPTR_FOR_PAGED_MEM=0
+#export MIOPEN_DEBUG_DISABLE_FIND_DB=1
+#export NCCL_SOCKET_IFNAME=eno1
+export HSA_FORCE_FINE_GRAIN_PCIE=1
+#source /public/software/apps/DeepLearning/PyTorch/pytorch-env.sh
+export MIOPEN_FIND_MODE=3
+#export MIOPEN_ENABLE_LOGGING_CMD=1
+#export export ROCBLAS_LAYER=3
+module unload compiler/rocm/2.9
+#export MIOPEN_DEBUG_CONV_DIRECT_ASM_3X3U=0
+#export MIOPEN_DEBUG_CONV_DIRECT_ASM_WRW3X3=0
+echo "MIOPEN_FIND_MODE=$MIOPEN_FIND_MODE"
+#module load apps/PyTorch/1.5.0a0/hpcx-2.4.1-gcc-7.3.1-rocm3.3
+#source /public/home/aiss/Pytorch/env_rocm3.3_torch1.5.sh
+lrank=$OMPI_COMM_WORLD_LOCAL_RANK
+comm_rank=$OMPI_COMM_WORLD_RANK
+comm_size=$OMPI_COMM_WORLD_SIZE
+
+#下边是修改的
+
+
+export HIP_VISIBLE_DEVICES=0,1,2,3
+#HSA_FORCE_FINE_GRAIN_PCIE=1 numactl --cpunodebind=4,5,6,7 --membind=4,5,6,7
+
+#module load compiler/rocm/3.9.1
+
+
+
+
+APP="python3 run_squad_v4.py \
+  --train_file  /public/home/hepj/data/sq1.1/train-v1.1.json \
+  --predict_file  /public/home/hepj/data/sq1.1/dev-v1.1.json \
+  --init_checkpoint  /public/home/hepj/model_source/pytorch_bert/model.ckpt-28252.pt \
+  --vocab_file  /public/home/hepj/model_source/pytorch_bert/vocab.txt \
+  --output_dir  /public/home/hepj/outdir/torch/SQuAD4 \
+  --config_file  /public/home/hepj/model_source/pytorch_bert/bert_config.json \
+  --json-summary  /public/home/hepj/outdir/torch/SQuAD4/results.json \
+  --bert_model bert-large-uncased \
+  --do_train \
+  --do_predict \
+  --do_eval \
+  --train_batch_size  4 \
+  --predict_batch_size 4 \
+  --gpus_per_node  1 \
+  --local_rank ${comm_rank} \
+  --world_size 4 \
+  --use_env  \
+  --dist_url tcp://localhost:34567 \
+  --eval_script /public/home/hepj/squad_eval/evaluate-v1.1.py
+ "
+#  --fp16  \
+#  --amp \
+
+case ${lrank} in
+[0])
+  export HIP_VISIBLE_DEVICES=0
+  export UCX_NET_DEVICES=mlx5_0:1
+  export UCX_IB_PCI_BW=mlx5_0:50Gbs
+  echo numactl --cpunodebind=0 --membind=0 ${APP}
+  numactl --cpunodebind=0 --membind=0 ${APP}
+
+  #echo GLOO_SOCKET_IFNAME=ib0 numactl --cpunodebind=0 --membind=0 ${APP} 
+  #GLOO_SOCKET_IFNAME=ib0 numactl --cpunodebind=0 --membind=0 ${APP}
+  ;;
+[1])
+  export HIP_VISIBLE_DEVICES=1
+  export UCX_NET_DEVICES=mlx5_1:1
+  export UCX_IB_PCI_BW=mlx5_1:50Gbs
+  echo numactl --cpunodebind=1 --membind=1 ${APP}
+  numactl --cpunodebind=1 --membind=1 ${APP}
+  ;;
+[2])
+  export HIP_VISIBLE_DEVICES=2
+  export UCX_NET_DEVICES=mlx5_2:1
+  export UCX_IB_PCI_BW=mlx5_2:50Gbs
+  echo numactl --cpunodebind=2 --membind=2 ${APP} 
+  numactl --cpunodebind=2 --membind=2 ${APP}
+  ;;
+[3])
+  export HIP_VISIBLE_DEVICES=3
+  export UCX_NET_DEVICES=mlx5_3:1
+  export UCX_IB_PCI_BW=mlx5_3:50Gbs
+  echo numactl --cpunodebind=3 --membind=3 ${APP}
+  numactl --cpunodebind=3 --membind=3 ${APP}
+  ;;
+esac

PyTorch/NLP/BERT/single_squad4_fp16.sh

+#!/bin/bash
+#export MIOPEN_DEBUG_DISABLE_FIND_DB=1
+#export NCCL_SOCKET_IFNAME=ib0
+#export HSA_USERPTR_FOR_PAGED_MEM=0
+#export MIOPEN_DEBUG_DISABLE_FIND_DB=1
+#export NCCL_SOCKET_IFNAME=eno1
+export HSA_FORCE_FINE_GRAIN_PCIE=1
+#source /public/software/apps/DeepLearning/PyTorch/pytorch-env.sh
+export MIOPEN_FIND_MODE=3
+#export MIOPEN_ENABLE_LOGGING_CMD=1
+#export export ROCBLAS_LAYER=3
+module unload compiler/rocm/2.9
+#export MIOPEN_DEBUG_CONV_DIRECT_ASM_3X3U=0
+#export MIOPEN_DEBUG_CONV_DIRECT_ASM_WRW3X3=0
+echo "MIOPEN_FIND_MODE=$MIOPEN_FIND_MODE"
+#module load apps/PyTorch/1.5.0a0/hpcx-2.4.1-gcc-7.3.1-rocm3.3
+#source /public/home/aiss/Pytorch/env_rocm3.3_torch1.5.sh
+lrank=$OMPI_COMM_WORLD_LOCAL_RANK
+comm_rank=$OMPI_COMM_WORLD_RANK
+comm_size=$OMPI_COMM_WORLD_SIZE
+
+#下边是修改的
+
+
+export HIP_VISIBLE_DEVICES=0,1,2,3
+#HSA_FORCE_FINE_GRAIN_PCIE=1 numactl --cpunodebind=4,5,6,7 --membind=4,5,6,7
+
+#module load compiler/rocm/3.9.1
+
+
+
+
+APP="python3 run_squad_v4.py \
+  --train_file  /public/home/hepj/data/sq1.1/train-v1.1.json \
+  --predict_file  /public/home/hepj/data/sq1.1/dev-v1.1.json \
+  --init_checkpoint  /public/home/hepj/model_source/pytorch_bert/model.ckpt-28252.pt \
+  --vocab_file  /public/home/hepj/model_source/pytorch_bert/vocab.txt \
+  --output_dir  /public/home/hepj/outdir/torch/SQuAD4 \
+  --config_file  /public/home/hepj/model_source/pytorch_bert/bert_config.json \
+  --json-summary  /public/home/hepj/outdir/torch/SQuAD4/results.json \
+  --bert_model bert-large-uncased \
+  --do_train \
+  --do_predict \
+  --do_eval \
+  --train_batch_size  4 \
+  --predict_batch_size 4 \
+  --gpus_per_node  1 \
+  --local_rank ${comm_rank} \
+  --world_size 4 \
+  --use_env  \
+  --fp16 \
+  --amp \
+  --dist_url tcp://localhost:34567 \
+  --eval_script /public/home/hepj/squad_eval/evaluate-v1.1.py
+ "
+
+case ${lrank} in
+[0])
+  export HIP_VISIBLE_DEVICES=0
+  export UCX_NET_DEVICES=mlx5_0:1
+  export UCX_IB_PCI_BW=mlx5_0:50Gbs
+  echo numactl --cpunodebind=0 --membind=0 ${APP}
+  numactl --cpunodebind=0 --membind=0 ${APP}
+
+  #echo GLOO_SOCKET_IFNAME=ib0 numactl --cpunodebind=0 --membind=0 ${APP} 
+  #GLOO_SOCKET_IFNAME=ib0 numactl --cpunodebind=0 --membind=0 ${APP}
+  ;;
+[1])
+  export HIP_VISIBLE_DEVICES=1
+  export UCX_NET_DEVICES=mlx5_1:1
+  export UCX_IB_PCI_BW=mlx5_1:50Gbs
+  echo numactl --cpunodebind=1 --membind=1 ${APP}
+  numactl --cpunodebind=1 --membind=1 ${APP}
+  ;;
+[2])
+  export HIP_VISIBLE_DEVICES=2
+  export UCX_NET_DEVICES=mlx5_2:1
+  export UCX_IB_PCI_BW=mlx5_2:50Gbs
+  echo numactl --cpunodebind=2 --membind=2 ${APP} 
+  numactl --cpunodebind=2 --membind=2 ${APP}
+  ;;
+[3])
+  export HIP_VISIBLE_DEVICES=3
+  export UCX_NET_DEVICES=mlx5_3:1
+  export UCX_IB_PCI_BW=mlx5_3:50Gbs
+  echo numactl --cpunodebind=3 --membind=3 ${APP}
+  numactl --cpunodebind=3 --membind=3 ${APP}
+  ;;
+esac

PyTorch/NLP/BERT/single_squad_fp16.sh

+#!/bin/bash
+#export MIOPEN_DEBUG_DISABLE_FIND_DB=1
+#export NCCL_SOCKET_IFNAME=ib0
+#export HSA_USERPTR_FOR_PAGED_MEM=0
+#export MIOPEN_DEBUG_DISABLE_FIND_DB=1
+#export NCCL_SOCKET_IFNAME=eno1
+export HSA_FORCE_FINE_GRAIN_PCIE=1
+#source /public/software/apps/DeepLearning/PyTorch/pytorch-env.sh
+export MIOPEN_FIND_MODE=3
+#export MIOPEN_ENABLE_LOGGING_CMD=1
+#export ROCBLAS_LAYER=3
+module unload compiler/rocm/2.9
+#export MIOPEN_DEBUG_CONV_DIRECT_ASM_3X3U=0
+#export MIOPEN_DEBUG_CONV_DIRECT_ASM_WRW3X3=0
+echo "MIOPEN_FIND_MODE=$MIOPEN_FIND_MODE"
+#module load apps/PyTorch/1.5.0a0/hpcx-2.4.1-gcc-7.3.1-rocm3.3
+#source /public/home/aiss/Pytorch/env_rocm3.3_torch1.5.sh
+lrank=$OMPI_COMM_WORLD_LOCAL_RANK
+comm_rank=$OMPI_COMM_WORLD_RANK
+comm_size=$OMPI_COMM_WORLD_SIZE
+
+#下边是修改的
+
+
+#export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+#HSA_FORCE_FINE_GRAIN_PCIE=1 numactl --cpunodebind=4,5,6,7 --membind=4,5,6,7
+
+#module load compiler/rocm/3.9.1
+
+
+
+
+APP="python3 run_squad_v1.py \
+  --train_file  /public/home/hepj/model/model_source/sq1.1/train-v1.1.json \
+  --predict_file  /public/home/hepj/model/model_source/sq1.1/dev-v1.1.json \
+  --init_checkpoint  /public/home/hepj/model/model_source/pytorch_bert/model.ckpt-28252.pt \
+  --vocab_file  /public/home/hepj/model/model_source/pytorch_bert/vocab.txt \
+  --output_dir  /public/home/hepj/outdir/tourch/SQuAD \
+  --config_file  /public/home/hepj/model/model_source/pytorch_bert/bert_config.json \
+  --json-summary  /public/home/hepj/outdir/torch/SQuAD/results.json \
+  --bert_model bert-large-uncased \
+  --do_train \
+  --do_predict \
+  --do_eval \
+  --train_batch_size  4 \
+  --predict_batch_size 4 \
+  --gpus_per_node  1 \
+  --local_rank -1 \
+  --fp16  \
+  --amp  \
+  --eval_script /public/home/hepj/squad_eval/evaluate-v1.1.py
+ "
+#--json-summary  /public/home/hepj/out_dir/tourch/SQuAD/results.json 
+
+
+case ${lrank} in
+[0])
+  export HIP_VISIBLE_DEVICES=0
+  export UCX_NET_DEVICES=mlx5_0:1
+  export UCX_IB_PCI_BW=mlx5_0:50Gbs
+  echo numactl --cpunodebind=0 --membind=0 ${APP}
+  numactl --cpunodebind=0 --membind=0 ${APP}
+
+  #echo GLOO_SOCKET_IFNAME=ib0 numactl --cpunodebind=0 --membind=0 ${APP} 
+  #GLOO_SOCKET_IFNAME=ib0 numactl --cpunodebind=0 --membind=0 ${APP}
+  ;;
+[1])
+  export HIP_VISIBLE_DEVICES=1
+  export UCX_NET_DEVICES=mlx5_1:1
+  export UCX_IB_PCI_BW=mlx5_1:50Gbs
+  echo numactl --cpunodebind=1 --membind=1 ${APP}
+  numactl --cpunodebind=1 --membind=1 ${APP}
+  ;;
+[2])
+  export HIP_VISIBLE_DEVICES=2
+  export UCX_NET_DEVICES=mlx5_2:1
+  export UCX_IB_PCI_BW=mlx5_2:50Gbs
+  echo numactl --cpunodebind=2 --membind=2 ${APP} 
+  numactl --cpunodebind=2 --membind=2 ${APP}
+  ;;
+[3])
+  export HIP_VISIBLE_DEVICES=3
+  export UCX_NET_DEVICES=mlx5_3:1
+  export UCX_IB_PCI_BW=mlx5_3:50Gbs
+  echo numactl --cpunodebind=3 --membind=3 ${APP}
+  numactl --cpunodebind=3 --membind=3 ${APP}
+  ;;
+esac

PyTorch/NLP/BERT/tf_to_torch/Dockerfile

+# Copyright (c) 2019-2020, NVIDIA CORPORATION. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+ARG FROM_IMAGE_NAME=nvcr.io/nvidia/pytorch:20.06-py3
+FROM ${FROM_IMAGE_NAME}
+
+# Install dependencies
+RUN apt-get update \
+ && apt-get install -y --no-install-recommends \
+        bzip2 \
+        cabextract \
+        iputils-ping \
+        pbzip2 \
+        pv \
+ && rm -rf /var/lib/apt/lists/*
+
+WORKDIR /workspace/bert
+COPY requirements.txt .
+RUN pip install --no-cache-dir https://github.com/mlperf/logging/archive/9ea0afa.zip \
+ && pip install --no-cache-dir -r requirements.txt
+
+# Preprocessing
+#WORKDIR /workspace
+#RUN git clone https://github.com/attardi/wikiextractor.git
+
+# Install BERT
+ENV BERT_PREP_WORKING_DIR /workspace/bert/data
+WORKDIR /workspace/bert
+COPY . .
+
+RUN cd /workspace/bert/mhalib && python setup.py build && cp build/lib*/mhalib* ../

PyTorch/NLP/BERT/tf_to_torch/LICENSE

+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
\ No newline at end of file

PyTorch/NLP/BERT/tf_to_torch/NOTICE

+BERT PyTorch
+
+This repository includes software from https://github.com/huggingface/pytorch-pretrained-BERT
+licensed under the Apache License 2.0.
+

PyTorch/NLP/BERT/tf_to_torch/README.md

+# Location of the input files 
+
+This [GCS location](https://console.cloud.google.com/storage/browser/pkanwar-bert) contains the following.
+* TensorFlow checkpoint (bert_model.ckpt) containing the pre-trained weights (which is actually 3 files).
+* Vocab file (vocab.txt) to map WordPiece to word id.
+* Config file (bert_config.json) which specifies the hyperparameters of the model.
+
+# Download and preprocess datasets
+
+Download the [wikipedia dump](https://dumps.wikimedia.org/enwiki/20200101/enwiki-20200101-pages-articles-multistream.xml.bz2) and extract the pages
+The wikipedia dump can be downloaded from this link in this directory, and should contain the following file:
+enwiki-20200101-pages-articles-multistream.xml.bz2
+
+Run [WikiExtractor.py](https://github.com/attardi/wikiextractor) to extract the wiki pages from the XML
+The generated wiki pages file will be stored as <data dir>/LL/wiki_nn; for example <data dir>/AA/wiki_00. Each file is ~1MB, and each sub directory has 100 files from wiki_00 to wiki_99, except the last sub directory. For the 20200101 dump, the last file is FE/wiki_17.
+
+Clean up
+The clean up scripts (some references here) are in the scripts directory.
+The following command will run the clean up steps, and put the results in ./results
+./process_wiki.sh '<data dir>/*/wiki_??'
+
+```shell
+cd cleanup_scripts  
+mkdir -p wiki  
+cd wiki  
+wget https://dumps.wikimedia.org/enwiki/20200101/enwiki-20200101-pages-articles-multistream.xml.bz2    # Optionally use curl instead  
+bzip2 -d enwiki-20200101-pages-articles-multistream.xml.bz2  
+cd ..    # back to bert/cleanup_scripts  
+git clone https://github.com/attardi/wikiextractor.git  
+python3 wikiextractor/WikiExtractor.py wiki/enwiki-20200101-pages-articles-multistream.xml    # Results are placed in bert/cleanup_scripts/text  
+./process_wiki.sh '<text/*/wiki_??'  
+```
+
+After running the process_wiki.sh script, for the 20200101 wiki dump, there will be 500 files, named part-00xxx-of-00500 in the ./results directory.
+
+# Checkpoint conversion
+python convert_tf_checkpoint.py --tf_checkpoint /cks/model.ckpt-28252 --bert_config_path /cks/bert_config.json --output_checkpoint model.ckpt-28252.pt
+
+# Generate the BERT input dataset
+
+The create_pretraining_data.py script duplicates the input plain text, replaces different sets of words with masks for each duplication, and serializes the output into the TFRecord file format. 
+
+```shell
+python3 create_pretraining_data.py \
+   --input_file=<path to ./results of previous step>/part-XX-of-00500 \
+   --output_file=<tfrecord dir>/part-XX-of-00500 \
+   --vocab_file=<path to vocab.txt> \
+   --do_lower_case=True \
+   --max_seq_length=512 \
+   --max_predictions_per_seq=76 \
+   --masked_lm_prob=0.15 \
+   --random_seed=12345 \
+   --dupe_factor=10
+```
+
+The generated tfrecord has 500 parts, totalling to ~365GB.
+
+# Running the model
+
+Building the Docker container
+```shell
+docker build --pull -t <docker/registry>/mlperf-nvidia:language_model .
+docker push <docker/registry>/mlperf-nvidia:language_model
+```
+
+To run this model, use the following command. Replace the configuration script to match the system being used (e.g., config_DGX1.sh, config_DGX2.sh, config_DGXA100.sh).
+
+```shell
+source config_DGXA100.sh
+sbatch -N${DGXNNODES} --ntasks-per-node=${DGXNGPU} --time=${WALLTIME} run.sub
+```
+
+Alternative launch with nvidia-docker. Replace the configuration script to match the system being used (e.g., config_DGX1.sh, config_DGX2.sh, config_DGXA100.sh).
+
+```bash
+docker build --pull -t mlperf-nvidia:language_model .
+source config_DGXA100.sh
+CONT=mlperf-nvidia:language_model DATADIR=<path/to/datadir> DATADIR_PHASE2=<path/to/datadir_phase2> EVALDIR=<path/to/evaldir> CHECKPOINTDIR=<path/to/checkpointdir> CHECKPOINTDIR_PHASE1=<path/to/checkpointdir_phase1> ./run_with_docker.sh
+```
+
+For multi-node training, we use Slurm for scheduling and Pyxis to run our container.
+
+## Configuration File Naming Convention
+
+All configuration files follow the format `config_<SYSTEM_NAME>_<NODES>x<GPUS/NODE>x<BATCH/GPU>x<GRADIENT_ACCUMULATION_STEPS>.sh`.
+
+### Example 1
+A DGX1 system with 1 node, 8 GPUs per node, batch size of 6 per GPU, and 6 gradient accumulation steps would use `config_DGX1_1x8x6x6.sh`.
+
+### Example 2
+A DGX A100 system with 32 nodes, 8 GPUs per node, batch size of 20 per GPU, and no gradient accumulation would use `config_DGXA100_32x8x20x1.sh`
+
+# Acknowledgements
+
+We'd like to thank members of the ONNX Runtime team at Microsoft for their suggested performance optimization to reduce the size of the last linear layer to only output the fraction of tokens that participate in the MLM loss calculation.
+
+
+

PyTorch/NLP/BERT/tf_to_torch/bind.sh

+#! /bin/bash
+set -euo pipefail
+
+print_usage() {
+    cat << EOF
+${0} [options] [--] COMMAND [ARG...]
+
+Control binding policy for each task. Assumes one rank will be launched for each GPU.
+
+Options:
+    --cpu=MODE
+        * exclusive -- bind each rank to an exclusive set of cores near its GPU
+        * exclusive,nosmt -- bind each rank to an exclusive set of cores near its GPU, without hyperthreading
+        * node -- bind each rank to all cores in the NUMA node nearest its GPU [default]
+	* *.sh -- bind each rank using the bash associative array bind_cpu_cores or bind_cpu_nodes from a file
+        * off -- don't bind
+    --mem=MODE
+        * node -- bind each rank to the nearest NUMA node [default]
+	* *.sh -- bind each rank using the bash associative array bind_mem from a file
+        * off -- don't bind
+    --ib=MODE
+        * single -- bind each rank to a single IB device near its GPU
+        * off -- don't bind [default]
+    --cluster=CLUSTER
+        Select which cluster is being used. May be required if system params cannot be detected.
+EOF
+}
+
+################################################################################
+# Argument parsing
+################################################################################
+
+cpu_mode='node'
+mem_mode='node'
+ib_mode='off'
+cluster=''
+while [ $# -gt 0 ]; do
+    case "$1" in
+        -h|--help) print_usage ; exit 0 ;;
+        --cpu=*) cpu_mode="${1/*=/}"; shift ;;
+        --cpu)   cpu_mode="$2"; shift 2 ;;
+        --mem=*) mem_mode="${1/*=/}"; shift ;;
+        --mem)   mem_mode="$2"; shift 2 ;;
+        --ib=*) ib_mode="${1/*=/}"; shift ;;
+        --ib)   ib_mode="$2"; shift 2 ;;
+        --cluster=*) cluster="${1/*=/}"; shift ;;
+        --cluster)   cluster="$2"; shift 2 ;;
+        --) shift; break ;;
+        *) break ;;
+    esac
+done
+if [ $# -lt 1 ]; then
+    echo 'ERROR: no command given' 2>&1
+    print_usage
+    exit 1
+fi
+
+################################################################################
+# Get system params
+################################################################################
+
+# LOCAL_RANK is set with an enroot hook for Pytorch containers
+# SLURM_LOCALID is set by Slurm
+# OMPI_COMM_WORLD_LOCAL_RANK is set by mpirun
+readonly local_rank="${LOCAL_RANK:=${SLURM_LOCALID:=${OMPI_COMM_WORLD_LOCAL_RANK:-}}}"
+if [ -z "${local_rank}" ]; then
+    echo 'ERROR: cannot read LOCAL_RANK from env' >&2
+    exit 1
+fi
+
+num_gpus=$(nvidia-smi -i 0 --query-gpu=count --format=csv,noheader,nounits)
+if [ "${local_rank}" -ge "${num_gpus}" ]; then
+    echo "ERROR: local rank is ${local_rank}, but there are only ${num_gpus} gpus available" >&2
+    exit 1
+fi
+
+get_lscpu_value() {
+    awk -F: "(\$1 == \"${1}\"){gsub(/ /, \"\", \$2); print \$2; found=1} END{exit found!=1}"
+}
+lscpu_out=$(lscpu)
+num_sockets=$(get_lscpu_value 'Socket(s)' <<< "${lscpu_out}")
+num_nodes=$(get_lscpu_value 'NUMA node(s)' <<< "${lscpu_out}")
+cores_per_socket=$(get_lscpu_value 'Core(s) per socket' <<< "${lscpu_out}")
+
+echo "num_sockets = ${num_sockets} num_nodes=${num_nodes} cores_per_socket=${cores_per_socket}"
+
+readonly cores_per_node=$(( (num_sockets * cores_per_socket) / num_nodes ))
+if [ ${num_gpus} -gt 1 ]; then
+    readonly gpus_per_node=$(( num_gpus / num_nodes ))
+else
+    readonly gpus_per_node=1
+fi
+readonly cores_per_gpu=$(( cores_per_node / gpus_per_node ))
+readonly local_node=$(( local_rank / gpus_per_node ))
+
+
+declare -a ibdevs=()
+case "${cluster}" in
+    circe)
+        # Need to specialize for circe because IB detection is hard
+        ibdevs=(mlx5_1 mlx5_2 mlx5_3 mlx5_4 mlx5_7 mlx5_8 mlx5_9 mlx5_10)
+        ;;
+   selene)
+        # Need to specialize for selene because IB detection is hard
+        ibdevs=(mlx5_0 mlx5_1 mlx5_2 mlx5_3 mlx5_6 mlx5_7 mlx5_8 mlx5_9)
+        ;;
+    '')
+        if ibstat_out="$(ibstat -l 2>/dev/null | sort -V)" ; then
+            mapfile -t ibdevs <<< "${ibstat_out}"
+        fi
+        ;;
+    *)
+        echo "ERROR: Unknown cluster '${cluster}'" >&2
+        exit 1
+        ;;
+esac
+readonly num_ibdevs="${#ibdevs[@]}"
+
+################################################################################
+# Setup for exec
+################################################################################
+
+declare -a numactl_args=()
+
+case "${cpu_mode}" in
+    exclusive)
+        numactl_args+=( "$(printf -- "--physcpubind=%u-%u,%u-%u" \
+            $(( local_rank * cores_per_gpu )) \
+            $(( (local_rank + 1) * cores_per_gpu - 1 )) \
+            $(( local_rank * cores_per_gpu + (cores_per_gpu * gpus_per_node * num_nodes) )) \
+            $(( (local_rank + 1) * cores_per_gpu + (cores_per_gpu * gpus_per_node * num_nodes) - 1 )) \
+        )" )
+        ;;
+    exclusive,nosmt)
+        numactl_args+=( "$(printf -- "--physcpubind=%u-%u" \
+            $(( local_rank * cores_per_gpu )) \
+            $(( (local_rank + 1) * cores_per_gpu - 1 )) \
+        )" )
+        ;;
+    node)
+        numactl_args+=( "--cpunodebind=${local_node}" )
+        ;;
+    *.sh)
+	source "${cpu_mode}"
+	if [ -n "${bind_cpu_cores:-}" ]; then
+	    numactl_args+=( "--physcpubind=${bind_cpu_cores[${local_rank}]}" )
+	elif [ -n "${bind_cpu_nodes:-}" ]; then
+	    numactl_args+=( "--cpunodebind=${bind_cpu_nodes[${local_rank}]}" )
+	else
+	    echo "ERROR: invalid CPU affinity file ${cpu_mode}." >&2
+	    exit 1
+	fi
+	;;
+    off|'')
+        ;;
+    *)
+        echo "ERROR: invalid cpu mode '${cpu_mode}'" 2>&1
+        print_usage
+        exit 1
+        ;;
+esac
+
+case "${mem_mode}" in
+    node)
+        numactl_args+=( "--membind=${local_node}" )
+        ;;
+    *.sh)
+	source "${mem_mode}"
+	if [ -z "${bind_mem:-}" ]; then
+	    echo "ERROR: invalid memory affinity file ${mem_mode}." >&2
+	    exit 1
+	fi
+	numactl_args+=( "--membind=${bind_mem[${local_rank}]}" )
+	;;
+    off|'')
+        ;;
+    *)
+        echo "ERROR: invalid mem mode '${mem_mode}'" 2>&1
+        print_usage
+        exit 1
+        ;;
+esac
+
+case "${ib_mode}" in
+    single)
+        if [ "${num_ibdevs}" -eq 0 ]; then
+            echo "WARNING: used '$0 --ib=single', but there are 0 IB devices available; skipping IB binding." 2>&1
+        else
+            readonly ibdev="${ibdevs[$(( local_rank * num_ibdevs / num_gpus ))]}"
+            export OMPI_MCA_btl_openib_if_include="${OMPI_MCA_btl_openib_if_include-$ibdev}"
+            export UCX_NET_DEVICES="${UCX_NET_DEVICES-$ibdev:1}"
+        fi
+        ;;
+    off|'')
+        ;;
+    *)
+        echo "ERROR: invalid ib mode '${ib_mode}'" 2>&1
+        print_usage
+        exit 1
+        ;;
+esac
+
+################################################################################
+# Exec
+################################################################################
+
+if [ "${#numactl_args[@]}" -gt 0 ] ; then
+    set -x
+    exec numactl "${numactl_args[@]}" -- "${@}"
+else
+    exec "${@}"
+fi

PyTorch/NLP/BERT/tf_to_torch/bind_pyt.py

+# Copyright (c) 2019 NVIDIA CORPORATION. All rights reserved.
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import sys
+import subprocess
+import os
+import socket
+from argparse import ArgumentParser, REMAINDER
+
+import torch
+
+
+def parse_args():
+    """
+    Helper function parsing the command line options
+    @retval ArgumentParser
+    """
+    parser = ArgumentParser(description="PyTorch distributed training launch "
+                                        "helper utilty that will spawn up "
+                                        "multiple distributed processes")
+
+    # Optional arguments for the launch helper
+    parser.add_argument("--nnodes", type=int, default=1,
+                        help="The number of nodes to use for distributed "
+                             "training")
+    parser.add_argument("--node_rank", type=int, default=0,
+                        help="The rank of the node for multi-node distributed "
+                             "training")
+    parser.add_argument("--nproc_per_node", type=int, default=1,
+                        help="The number of processes to launch on each node, "
+                             "for GPU training, this is recommended to be set "
+                             "to the number of GPUs in your system so that "
+                             "each process can be bound to a single GPU.")
+    parser.add_argument("--master_addr", default="127.0.0.1", type=str,
+                        help="Master node (rank 0)'s address, should be either "
+                             "the IP address or the hostname of node 0, for "
+                             "single node multi-proc training, the "
+                             "--master_addr can simply be 127.0.0.1")
+    parser.add_argument("--master_port", default=29500, type=int,
+                        help="Master node (rank 0)'s free port that needs to "
+                             "be used for communciation during distributed "
+                             "training")
+    parser.add_argument('--no_hyperthreads', action='store_true',
+                        help='Flag to disable binding to hyperthreads')
+    parser.add_argument('--no_membind', action='store_true',
+                        help='Flag to disable memory binding')
+
+    # non-optional arguments for binding
+    parser.add_argument("--nsockets_per_node", type=int, required=True,
+                        help="Number of CPU sockets on a node")
+    parser.add_argument("--ncores_per_socket", type=int, required=True,
+                        help="Number of CPU cores per socket")
+
+    # positional
+    parser.add_argument("training_script", type=str,
+                        help="The full path to the single GPU training "
+                             "program/script to be launched in parallel, "
+                             "followed by all the arguments for the "
+                             "training script")
+
+    # rest from the training program
+    parser.add_argument('training_script_args', nargs=REMAINDER)
+    return parser.parse_args()
+
+def main():
+    args = parse_args()
+
+    # variables for numactrl binding
+    
+    
+    NSOCKETS = args.nsockets_per_node
+    NGPUS_PER_SOCKET = (args.nproc_per_node // args.nsockets_per_node) + (1 if (args.nproc_per_node % args.nsockets_per_node) else 0)
+    NCORES_PER_GPU = args.ncores_per_socket // NGPUS_PER_SOCKET
+
+    # world size in terms of number of processes
+    dist_world_size = args.nproc_per_node * args.nnodes
+
+    # set PyTorch distributed related environmental variables
+    current_env = os.environ.copy()
+    current_env["MASTER_ADDR"] = args.master_addr
+    current_env["MASTER_PORT"] = str(args.master_port)
+    current_env["WORLD_SIZE"] = str(dist_world_size)
+
+    processes = []
+
+    for local_rank in range(0, args.nproc_per_node):
+        # each process's rank
+        dist_rank = args.nproc_per_node * args.node_rank + local_rank
+        current_env["RANK"] = str(dist_rank)
+
+        # form numactrl binding command
+        cpu_ranges = [local_rank * NCORES_PER_GPU,
+                     (local_rank + 1) * NCORES_PER_GPU - 1,
+                     local_rank * NCORES_PER_GPU + (NCORES_PER_GPU * NGPUS_PER_SOCKET * NSOCKETS),
+                     (local_rank + 1) * NCORES_PER_GPU + (NCORES_PER_GPU * NGPUS_PER_SOCKET * NSOCKETS) - 1]
+
+        numactlargs = []
+        if args.no_hyperthreads:
+            numactlargs += [ "--physcpubind={}-{}".format(*cpu_ranges[0:2]) ]
+        else:
+            numactlargs += [ "--physcpubind={}-{},{}-{}".format(*cpu_ranges) ]
+
+        if not args.no_membind:
+            memnode = local_rank // NGPUS_PER_SOCKET
+            numactlargs += [ "--membind={}".format(memnode) ]
+
+        # spawn the processes
+        cmd = [ "/usr/bin/numactl" ] \
+            + numactlargs \
+            + [ sys.executable,
+                "-u",
+                args.training_script,
+                "--local_rank={}".format(local_rank)
+              ] \
+            + args.training_script_args
+
+        process = subprocess.Popen(cmd, env=current_env)
+        processes.append(process)
+
+    for process in processes:
+        process.wait()
+
+
+if __name__ == "__main__":
+    main()
+
+

PyTorch/NLP/BERT/tf_to_torch/bmm1.py

+import torch
+import mhalib
+
+###########################################################################################
+
+class Bmm1Function(torch.autograd.Function):
+
+    @staticmethod
+    def forward(ctx, batch1, batch2, seqlen, batch, maxseqlen, heads, embed, scale, stream, sync):
+        ctx.save_for_backward(batch1, batch2, seqlen)
+        ctx.batch = batch
+        ctx.maxseqlen = maxseqlen
+        ctx.heads = heads
+        ctx.embed = embed
+        ctx.scale = scale
+        ctx.sync = sync
+        ctx.stream = stream
+        ntokens = seqlen.sum().item()
+        ctx.ntokens = ntokens
+        ntokens2 = 0
+        for i in range(batch):
+            ntokens2 += seqlen[i]*seqlen[i]
+
+        output = torch.empty(ntokens2*heads, device="cuda", dtype=torch.float16)
+        mhalib.FastBmm1Fprop(batch2.flatten().contiguous(), batch1.flatten().contiguous(), output.flatten().contiguous(), batch, seqlen, heads, embed, scale, False, stream, sync)
+
+        return output[:ntokens2*heads]
+
+    @staticmethod
+    def backward(ctx, grad_output):
+
+        batch1, batch2, seqlen = ctx.saved_tensors
+        batch = ctx.batch
+        maxseqlen = ctx.maxseqlen
+        heads = ctx.heads
+        embed = ctx.embed
+        ntokens = ctx.ntokens
+
+        grad_batch1 = torch.empty(ntokens,heads*embed, device="cuda", dtype=torch.float16)
+        grad_batch2 = torch.empty(ntokens,heads*embed, device="cuda", dtype=torch.float16)
+
+        mhalib.FastBmm1Dgrad2(batch2.flatten().contiguous(), grad_output.flatten().contiguous(), grad_batch1.flatten().contiguous(), batch, seqlen, heads, embed, ctx.scale, False, ctx.stream, ctx.sync)
+        mhalib.FastBmm1Dgrad1(batch1.flatten().contiguous(), grad_output.flatten().contiguous(), grad_batch2.flatten().contiguous(), batch, seqlen, heads, embed, ctx.scale, False, ctx.stream, ctx.sync)
+
+        return grad_batch1[:ntokens], grad_batch2[:ntokens], None, None, None, None, None, None, None, None
+
+class Bmm1(torch.nn.Module):
+    def __init__(self, batch, seqlen, heads, embed, scale=False, stream=True, sync=True):
+        super(Bmm1, self).__init__()
+
+        self.heads = heads
+        self.embed = embed
+        self.maxseqlen = seqlen
+        self.scale = scale
+        self.sync = sync
+        self.stream = stream
+
+    def forward(self, batch1, batch2, batch, seqlen):
+        return Bmm1Function.apply(batch1, batch2, seqlen, batch, self.maxseqlen, self.heads, self.embed, self.scale, self.stream, self.sync)
+
+##########################################################################################
+
+class Bmm1StridedFunction(torch.autograd.Function):
+
+    @staticmethod
+    def forward(ctx, mixed, seqlen, batch, maxseqlen, heads, embed, scale, stream, sync, timers):
+        ctx.save_for_backward(mixed, seqlen)
+        ctx.batch = batch
+        ctx.maxseqlen = maxseqlen
+        ctx.heads = heads
+        ctx.embed = embed
+        ctx.scale = scale
+        ctx.sync = sync
+        ctx.stream = stream
+        ctx.timers = timers
+        ntokens = seqlen.sum().item()
+        ctx.ntokens = ntokens
+        ntokens2 = 0
+        for i in range(batch):
+            ntokens2 += seqlen[i]*seqlen[i]
+
+        output = torch.empty(ntokens2*heads, device="cuda", dtype=torch.float16)
+
+        if timers: timers['start_fprop'].record()
+        mhalib.FastBmm1Fprop(mixed, mixed, output, batch, seqlen, heads, embed, scale, True, stream, sync)
+
+        if timers: timers['stop_fprop'].record()
+
+        return output[:ntokens2*heads], mixed
+
+    @staticmethod
+    #def backward(ctx, grad_output):
+    def backward(ctx, grad_output, grad_mixed):
+
+        mixed, seqlen = ctx.saved_tensors
+        batch = ctx.batch
+        maxseqlen = ctx.maxseqlen
+        heads = ctx.heads
+        embed = ctx.embed
+        ntokens = ctx.ntokens
+
+        #grad_mixed = torch.empty([ntokens,heads*3*embed], device="cuda", dtype=torch.float16)
+
+        if ctx.timers: ctx.timers['start_dgrad'].record()
+        mhalib.FastBmm1Dgrad2(mixed, grad_output, grad_mixed, batch, seqlen, heads, embed, ctx.scale, True, ctx.stream, ctx.sync)
+        if ctx.timers: ctx.timers['stop_dgrad'].record()
+        if ctx.timers: ctx.timers['start_wgrad'].record()
+        mhalib.FastBmm1Dgrad1(mixed, grad_output, grad_mixed, batch, seqlen, heads, embed, ctx.scale, True, ctx.stream, ctx.sync)
+        if ctx.timers: ctx.timers['stop_wgrad'].record()
+        #return grad_mixed[:ntokens], None, None, None, None, None, None, None, None, None
+        return grad_mixed[:ntokens], grad_mixed, None, None, None, None, None, None, None, None, None
+
+class Bmm1Strided(torch.nn.Module):
+    def __init__(self, batch, seqlen, heads, embed, scale=True, stream=True, sync=True, timer=False):
+        super(Bmm1Strided, self).__init__()
+
+        self.heads = heads
+        self.embed = embed
+        self.maxseqlen = seqlen
+        self.scale = scale
+        self.sync = sync
+        self.stream = stream
+        if timer:
+            self.timers = {'start_fprop':torch.cuda.Event(enable_timing=True),
+                           'start_dgrad':torch.cuda.Event(enable_timing=True),
+                           'start_wgrad':torch.cuda.Event(enable_timing=True),
+                           'stop_fprop':torch.cuda.Event(enable_timing=True),
+                           'stop_dgrad':torch.cuda.Event(enable_timing=True),
+                           'stop_wgrad':torch.cuda.Event(enable_timing=True)}
+        else:
+            self.timers = None
+
+    def forward(self, mixed, batch, seqlen):
+        return Bmm1StridedFunction.apply(mixed, seqlen, batch, self.maxseqlen, self.heads, self.embed, self.scale, self.stream, self.sync, self.timers)
+
+###########################################################################################

PyTorch/NLP/BERT/tf_to_torch/bmm2.py

+import torch
+import mhalib
+
+###########################################################################################
+
+class Bmm2Function(torch.autograd.Function):
+
+    @staticmethod
+    def forward(ctx, batch1, batch2, seqlen, batch, maxseqlen, heads, embed, sync, stream):
+        ctx.save_for_backward(batch1, batch2, seqlen)
+        ctx.batch = batch
+        ctx.maxseqlen = maxseqlen
+        ctx.heads = heads
+        ctx.embed = embed
+        ctx.stream = stream
+        ctx.sync = sync
+        ntokens = seqlen.sum().item()
+        ctx.ntokens = ntokens
+
+        output = torch.empty([ntokens,heads,embed], device="cuda", dtype=torch.float16)
+        mhalib.FastBmm2Fprop(batch2.flatten().contiguous(), batch1.flatten().contiguous(), output, batch, seqlen, heads, embed, False, False, stream, sync)
+
+        return output[:ntokens]
+
+    @staticmethod
+    def backward(ctx, grad_output):
+
+        batch1, batch2, seqlen = ctx.saved_tensors
+        batch = ctx.batch
+        maxseqlen = ctx.maxseqlen
+        heads = ctx.heads
+        embed = ctx.embed
+        ntokens = ctx.ntokens
+        ntokens2 = 0
+        for i in range(batch):
+            ntokens2 += seqlen[i]*seqlen[i]
+
+        grad_batch1 = torch.empty([ntokens2*heads], device="cuda", dtype=torch.float16)
+        grad_batch2 = torch.empty([ntokens,heads*embed], device="cuda", dtype=torch.float16)
+
+        mhalib.FastBmm2Dgrad1(batch2.flatten().contiguous(), grad_output, grad_batch1, batch, seqlen, heads, embed, False, False, ctx.stream, ctx.sync)
+        mhalib.FastBmm2Dgrad2(grad_output, batch1, grad_batch2, batch, seqlen, heads, embed, False, False, ctx.stream, ctx.sync)
+
+        return grad_batch1[:ntokens2*heads], grad_batch2[:ntokens], None, None, None, None, None, None, None
+
+class Bmm2(torch.nn.Module):
+    def __init__(self, batch, seqlen, heads, embed, stream=True, sync=True):
+        super(Bmm2, self).__init__()
+
+        self.heads = heads
+        self.embed = embed
+        self.maxseqlen = seqlen
+        self.stream = stream
+        self.sync = sync
+
+    def forward(self, batch1, batch2, batch, seqlen):
+        return Bmm2Function.apply(batch1, batch2, seqlen, batch, self.maxseqlen, self.heads, self.embed, self.stream, self.sync)
+
+###########################################################################################
+
+class Bmm2StridedFunction(torch.autograd.Function):
+
+    @staticmethod
+    def forward(ctx, batch1, mixed, seqlen, batch, maxseqlen, heads, embed, stream, sync, timers):
+        ctx.save_for_backward(batch1, mixed, seqlen)
+        ctx.batch = batch
+        ctx.maxseqlen = maxseqlen
+        ctx.heads = heads
+        ctx.embed = embed
+        ctx.stream = stream
+        ctx.sync = sync
+        ctx.timers = timers
+        ntokens = seqlen.sum().item()
+        ctx.ntokens = ntokens
+
+        output = torch.empty([ntokens,heads,embed], device="cuda", dtype=torch.float16)
+
+        if timers: timers['start_fprop'].record()
+        mhalib.FastBmm2Fprop(mixed, batch1, output, batch, seqlen, heads, embed, False, True, stream, sync)
+        if timers: timers['stop_fprop'].record()
+
+        return output[:ntokens]
+
+    @staticmethod
+    def backward(ctx, grad_output):
+
+        batch1, mixed, seqlen = ctx.saved_tensors
+        batch = ctx.batch
+        maxseqlen = ctx.maxseqlen
+        heads = ctx.heads
+        embed = ctx.embed
+        ntokens = ctx.ntokens
+        ntokens2 = 0
+        for i in range(batch):
+            ntokens2 += seqlen[i]*seqlen[i]
+
+        grad_batch1 = torch.empty(ntokens2*heads, device="cuda", dtype=torch.float16)
+        grad_mixed = torch.empty([ntokens,heads*3*embed], device="cuda", dtype=torch.float16)
+
+        if ctx.timers: ctx.timers['start_dgrad'].record()
+        mhalib.FastBmm2Dgrad1(mixed, grad_output, grad_batch1, batch, seqlen, heads, embed, False, True, ctx.stream, ctx.sync)
+        if ctx.timers: ctx.timers['stop_dgrad'].record()
+        if ctx.timers: ctx.timers['start_wgrad'].record()
+        mhalib.FastBmm2Dgrad2(grad_output, batch1, grad_mixed, batch, seqlen, heads, embed, False, True, ctx.stream, ctx.sync)
+        if ctx.timers: ctx.timers['stop_wgrad'].record()
+        return grad_batch1[:ntokens2*heads], grad_mixed[:ntokens], None, None, None, None, None, None, None, None
+
+class Bmm2Strided(torch.nn.Module):
+    def __init__(self, batch, seqlen, heads, embed, stream=True, sync=True, timer=False):
+        super(Bmm2Strided, self).__init__()
+
+        self.heads = heads
+        self.embed = embed
+        self.maxseqlen = seqlen
+        self.stream = stream
+        self.sync = sync
+        if timer:
+            self.timers = {'start_fprop':torch.cuda.Event(enable_timing=True),
+                           'start_dgrad':torch.cuda.Event(enable_timing=True),
+                           'start_wgrad':torch.cuda.Event(enable_timing=True),
+                           'stop_fprop':torch.cuda.Event(enable_timing=True),
+                           'stop_dgrad':torch.cuda.Event(enable_timing=True),
+                           'stop_wgrad':torch.cuda.Event(enable_timing=True)}
+        else:
+            self.timers = None
+
+    def forward(self, batch1, mixed, batch, seqlen):
+        return Bmm2StridedFunction.apply(batch1, mixed, seqlen, batch, self.maxseqlen, self.heads, self.embed, self.stream, self.sync, self.timers)
+
+###########################################################################################

PyTorch/NLP/BERT/tf_to_torch/cleanup_scripts/chop_hdf5_files.py

+# Copyright (c) 2019 NVIDIA CORPORATION. All rights reserved.
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import glob
+import h5py
+import multiprocessing
+import numpy as np
+
+hdf5_compression_method = None
+
+input_path = 'hdf5'
+input_files = sorted(glob.glob(input_path + '/part*', recursive=False))
+
+print('n_input_shards =', len(input_files))
+
+max_pred_per_seq = 76
+seq_length = 512
+n_output_shards = 2048
+
+ofile_prefix = 'chopped_2048_balanced/part_'
+ofile_suffix = '_of_' + str(n_output_shards) + '.hdf5'
+
+print('n_output_shards =', n_output_shards)
+
+# First pass over data to get sample count (read only the smallest array to get count)
+n_samples = 0
+for idx, ifile in enumerate(input_files):
+  print("Scanning:", ifile, " --  Progress:", idx+1, '/', len(input_files))
+  h5_ifile = h5py.File(ifile, 'r')
+
+  f_next_sentence_labels = h5_ifile['next_sentence_labels'][:]
+
+  h5_ifile.close()
+  n_samples += f_next_sentence_labels.shape[0]
+  
+ 
+# Find a "nominal" number of samples per shard (calculated to always go over by one shard size)
+# Find excess samples in last shard and distribute removal of excess over first "N" shards (could be done over last, but it doesn't matter and math is easier this way)
+#  (since 0 <= excess < nominal_shard_size, the max imbalance will be 1 sample to minimize the straggler effect)
+n_sample_per_ofile_nominal = (n_samples + n_output_shards - 1) // n_output_shards
+n_excess = n_output_shards * n_sample_per_ofile_nominal - n_samples  # Always a positive number
+
+ofile_handles = [h5py.File(ofile_prefix + str(x) + ofile_suffix, 'w') for x in range(n_output_shards)]
+
+ofile_idx = 0  # which output file
+ofile_entry_idx = 0  # index into an individual data element of an output file
+ifile_entry_idx = 0
+
+n_samples_in_this_shard = n_sample_per_ofile_nominal - 1 
+o_input_ids = np.ndarray((n_samples_in_this_shard, seq_length))
+o_input_masks = np.ndarray((n_samples_in_this_shard, seq_length))
+o_segment_ids = np.ndarray((n_samples_in_this_shard, seq_length))
+o_masked_lm_positions = np.ndarray((n_samples_in_this_shard, max_pred_per_seq))
+o_masked_lm_ids = np.ndarray((n_samples_in_this_shard, max_pred_per_seq))
+o_next_sentence_labels = np.ndarray((n_samples_in_this_shard))
+
+for ifile in input_files:
+  print("Processing:", ifile, " --  Progress:", idx+1, '/', len(input_files))
+  h5_ifile = h5py.File(ifile, 'r')
+  
+  ifile_entry_idx = 0
+  f_input_ids = h5_ifile['input_ids'][:]
+  f_input_masks = h5_ifile['input_mask'][:]
+  f_segment_ids = h5_ifile['segment_ids'][:]
+  f_masked_lm_positions = h5_ifile['masked_lm_positions'][:]
+  f_masked_lm_ids = h5_ifile['masked_lm_ids'][:]
+  f_next_sentence_labels = h5_ifile['next_sentence_labels'][:]
+
+  h5_ifile.close()
+
+  # This could be vectorized but keeping it simple due to lack of time
+  while ifile_entry_idx < f_input_ids.shape[0]:
+    if ofile_entry_idx == n_samples_in_this_shard:
+      ofile_handles[ofile_idx].create_dataset("input_ids",            data=o_input_ids,            dtype='i4', compression=hdf5_compression_method)
+      ofile_handles[ofile_idx].create_dataset("input_mask",           data=o_input_masks,          dtype='i1', compression=hdf5_compression_method)
+      ofile_handles[ofile_idx].create_dataset("segment_ids",          data=o_segment_ids,          dtype='i1', compression=hdf5_compression_method)
+      ofile_handles[ofile_idx].create_dataset("masked_lm_positions",  data=o_masked_lm_positions,  dtype='i4', compression=hdf5_compression_method)
+      ofile_handles[ofile_idx].create_dataset("masked_lm_ids",        data=o_masked_lm_ids,        dtype='i4', compression=hdf5_compression_method)
+      ofile_handles[ofile_idx].create_dataset("next_sentence_labels", data=o_next_sentence_labels, dtype='i1', compression=hdf5_compression_method)
+      ofile_handles[ofile_idx].flush()
+      ofile_handles[ofile_idx].close()
+
+      ofile_entry_idx = 0
+      ofile_idx += 1
+      print("Opening output idx:", ofile_idx)
+
+      n_samples_in_this_shard = n_sample_per_ofile_nominal
+      if ofile_entry_idx < n_excess:
+        n_samples_in_this_shard -= 1
+
+      o_input_ids = np.ndarray((n_samples_in_this_shard, seq_length))
+      o_input_masks = np.ndarray((n_samples_in_this_shard, seq_length))
+      o_segment_ids = np.ndarray((n_samples_in_this_shard, seq_length))
+      o_masked_lm_positions = np.ndarray((n_samples_in_this_shard, max_pred_per_seq))
+      o_masked_lm_ids = np.ndarray((n_samples_in_this_shard, max_pred_per_seq))
+      o_next_sentence_labels = np.ndarray((n_samples_in_this_shard))
+
+    o_input_ids[ofile_entry_idx] = f_input_ids[ifile_entry_idx]
+    o_input_masks[ofile_entry_idx] = f_input_masks[ifile_entry_idx]
+    o_segment_ids[ofile_entry_idx] = f_segment_ids[ifile_entry_idx]
+    o_masked_lm_positions[ofile_entry_idx] = f_masked_lm_positions[ifile_entry_idx]
+    o_masked_lm_ids[ofile_entry_idx] = f_masked_lm_ids[ifile_entry_idx]
+    o_next_sentence_labels[ofile_entry_idx] = f_next_sentence_labels[ifile_entry_idx]
+    ofile_entry_idx += 1
+
+    ifile_entry_idx += 1  

PyTorch/NLP/BERT/tf_to_torch/cleanup_scripts/clean.sh

+#!/bin/bash
+# Copyright 2018 MLBenchmark Group. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+
+filein=$1
+fileout=$2
+
+echo "Further clean up: $filein => $fileout"
+
+cmd="cat $filein "
+cmd+="| grep -v '^<doc [^>]*>$' "
+cmd+="| grep -vE '\[\[Category:[^][]*\]\]' "
+cmd+="| sed 's/\[\[\([^]|[]*\)\]\]/\1/g' "
+cmd+="| sed 's/\[\[\([^]|[]*\)\]\]/\1/g' "
+cmd+="| sed 's/\[\[[^]|[]*|\([^]|[]*\)\]\]/\1/g' "
+cmd+="| sed 's/\[\[[^]|[]*|\([^]|[]*\)\]\]/\1/g' "
+cmd+="| sed 's/\[\[[:]*[Ff]ile:[^][]*\]\]//g' "
+cmd+="| sed 's/\[\[[Mm]edia:[^][]*\]\]//g' "
+cmd+="| sed 's/\[\[[Ii]mage:[^][]*\]\]//g' "
+cmd+="| sed 's/\[\([^]|[]*\)\]/\1/g' "
+cmd+="| sed 's/\[\[\([^][]*\)\]\]//g' "
+cmd+="| sed 's/alt=//g' "
+cmd+="| sed 's/<\/doc>/\r/g' "
+cmd+="| sed 's/<chem\([^<]*\)<\/chem>/\1/g' "
+cmd+="| sed 's/<ins\([^<]*\)<\/ins>/\1/g' "
+cmd+="| sed 's/<\, ref \([^<]*\)<\/ref>//g' "
+cmd+="| sed 's/<includeonly\([^<]*\)<\/includeonly>//g' "
+cmd+="| sed 's/<graph\([^<]*\)<\/graph>//g' "
+cmd+="| sed 's/<section\([^\\]*\)\/>//g' "
+cmd+="| sed 's/<meta\([^\\]*\)\/>//g' "
+cmd+="| sed 's/<hr\([^\\]*\)\/>//g' "
+cmd+="| sed 's/<gallery\([^>]*\)>//g' "
+cmd+="| sed 's/<ref\([^<]*\)<\/ref>//g' "
+cmd+="| sed 's/<ref\([^>]*\)>//g' "
+cmd+="| sed 's/<http\([^>]*\)>//g' "
+cmd+="| sed 's/<Ref\([^>]*\)>//g' "
+cmd+="| sed 's/<mapframe \([^\/]*\)\/>//g' "
+cmd+="| sed 's/<mapframe\([^>]*\)>//g' "
+cmd+="| sed 's/<\/mapframe>//g' "
+cmd+="| sed 's/<poem>//g' "
+cmd+="| sed 's/<\/poem>//g' "
+cmd+="| sed 's/<math>//g' "
+cmd+="| sed 's/<\/math>//g' "
+cmd+="| sed 's/<ref>//g' "
+cmd+="| sed 's/<\/ref>//g' "
+cmd+="| sed 's/<div\([^>]*\)>//g' "
+cmd+="| sed 's/<\/div\([^>]*\)>//g' "
+cmd+="| sed 's/<\/div style>//g' "
+cmd+="| sed 's/<\/div>//g' "
+cmd+="| sed 's/<sup>//g' "
+cmd+="| sed 's/<\/sup>//g' "
+cmd+="| sed 's/<br>//g' "
+cmd+="| sed 's/<\/br>//g' "
+cmd+="| sed 's/<BR>//g' "
+cmd+="| sed 's/<\/BR>//g' "
+cmd+="| sed 's/<Br>//g' "
+cmd+="| sed 's/<\/Br>//g' "
+cmd+="| sed 's/<del>//g' "
+cmd+="| sed 's/<\/del>//g' "
+cmd+="| sed 's/<nowiki>//g' "
+cmd+="| sed 's/<\/nowiki>//g' "
+cmd+="| sed 's/<NOWIKI>//g' "
+cmd+="| sed 's/<\/NOWIKI>//g' "
+cmd+="| sed 's/<onlyinclude>//g' "
+cmd+="| sed 's/<\/onlyinclude>//g' "
+cmd+="| sed 's/<includeonly>//g' "
+cmd+="| sed 's/<\/includeonly>//g' "
+cmd+="| sed 's/<small>//g' "
+cmd+="| sed 's/<\/small>//g' "
+cmd+="| sed 's/<chem>//g' "
+cmd+="| sed 's/<\/chem>//g' "
+cmd+="| sed 's/<noinclude>//g' "
+cmd+="| sed 's/<\/noinclude>//g' "
+cmd+="| sed 's/<gallery>//g' "
+cmd+="| sed 's/<\/gallery>//g' "
+cmd+="| sed 's/<graph>{//g' "
+cmd+="| sed 's/<graph>//g' "
+cmd+="| sed 's/}<\/graph>//g' "
+cmd+="| sed 's/<\/graph>//g' "
+cmd+="| sed 's/<\/references>//g' "
+cmd+="| sed 's/<poem \([^>]*\)>//g' "
+# cmd+="| grep -v '^[ \t]*$' "
+cmd+="> $fileout"
+
+bash -c "${cmd[@]}"

PyTorch/NLP/BERT/tf_to_torch/cleanup_scripts/cleanup_file.py

+# Copyright 2018 MLBenchmark Group. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+
+# Lint as: python3
+"""Script to clean up input wiki dump for BERT input."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import argparse
+import glob
+import io
+import logging
+import multiprocessing
+import os
+import time
+
+
+parser = argparse.ArgumentParser(description='Wiki clean up for BERT.')
+parser.add_argument(
+    '--data',
+    type=str,
+    default='./wiki_??',
+    help='Input files. Default is "./wiki_??"')
+parser.add_argument(
+    '--input_suffix',
+    type=str,
+    default='',
+    help='Suffix for input files. Default is ""')
+parser.add_argument(
+    '--output_suffix',
+    type=str,
+    default='.1',
+    help='Suffix for output files. Default is ".1"')
+parser.add_argument(
+    '--nworker',
+    type=int,
+    default=72,
+    help='Number of workers for parallel processing.')
+args = parser.parse_args()
+
+
+def process_one_file(one_input):
+  """Remove <doc> tag and title of pages, for one file."""
+  input_filename = one_input + args.input_suffix
+  output_filename = one_input + args.output_suffix
+  logging.info('Processing %s => %s', input_filename, output_filename)
+
+  with io.open(input_filename, 'r', encoding='utf-8') as fin:
+    with io.open(output_filename, 'w', encoding='utf-8') as fout:
+
+      keep_next_line = True
+      for line in fin:
+        if not keep_next_line:
+          keep_next_line = True
+          continue
+
+        if '<doc' in line:
+          keep_next_line = False
+          fout.write(u'\n')
+          continue
+
+        if '</doc>' in line:
+          continue
+
+        if len(line) == 1:
+          continue
+
+        fout.write(line)
+
+
+if __name__ == '__main__':
+  input_files = sorted(glob.glob(os.path.expanduser(args.data)))
+  num_files = len(input_files)
+  num_workers = args.nworker
+  logging.basicConfig(level=logging.INFO)
+  logging.info('Number of input files to process = %d', num_files)
+
+  tic = time.time()
+  p = multiprocessing.Pool(num_workers)
+  p.map(process_one_file, input_files)
+  toc = time.time()
+  logging.info('Processed %s in %.2f sec', args.data, toc - tic)

PyTorch/NLP/BERT/tf_to_torch/cleanup_scripts/create_pretraining_data_wrapper.sh

+#!/bin/bash
+# Copyright (c) 2019 NVIDIA CORPORATION. All rights reserved.
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+input_path=${1}
+output_dir="hdf5"
+
+input_file=$(basename $input_path)
+
+mkdir -p ${output_dir}
+
+python3 ../create_pretraining_data.py \
+   --input_file=${input_path} \
+   --output_file="${output_dir}/${input_file}" \
+   --vocab_file=vocab.txt \
+   --do_lower_case=True \
+   --max_seq_length=512 \
+   --max_predictions_per_seq=76 \
+   --masked_lm_prob=0.15 \
+   --random_seed=12345 \
+   --dupe_factor=10