remove param --use_boxing_v2 for all cnns

Classification/cnns/README.md

@@ -601,7 +601,6 @@ python3 cnn_benchmark/of_cnn_train_val.py \
     --val_batch_size_per_device=512 \
     --num_epoch=90 \
     --use_fp16=false \
-    --use_boxing_v2=false \
     --model="alexnet" \
 ```
 
@@ -630,7 +629,6 @@ python3 cnn_benchmark/of_cnn_train_val.py \
     --val_batch_size_per_device=128 \
     --num_epoch=90 \
     --use_fp16=false \
-    --use_boxing_v2=false \
     --model="vgg" \
 ```
 

Classification/cnns/job_function_util.py

@@ -22,8 +22,6 @@ def get_train_config(args):
     train_config.all_reduce_group_min_mbyte(8)
     train_config.all_reduce_group_num(128)
 
-    if args.use_boxing_v2:
-        train_config.use_boxing_v2(True)
 
     train_config.prune_parallel_cast_ops(True)
     train_config.train.model_update_conf(gen_model_update_conf(args))

ClickThroughRate/WideDeepLearning/wdl_train_eval.py

@@ -111,7 +111,6 @@ def _get_train_conf():
         'lazy_adam_conf': {
         }
     })
-    train_conf.use_boxing_v2(True)
     train_conf.default_distribute_strategy(flow.distribute.consistent_strategy())
     train_conf.indexed_slices_optimizer_conf(dict(include_op_names=dict(op_name=['wide_embedding', 'deep_embedding'])))
     return train_conf

ClickThroughRate/WideDeepLearning/wdl_train_eval_test.py

@@ -118,7 +118,6 @@ def _get_train_conf():
         'lazy_adam_conf': {
         }
     })
-    train_conf.use_boxing_v2(True)
     train_conf.default_distribute_strategy(flow.distribute.consistent_strategy())
     train_conf.indexed_slices_optimizer_conf(dict(include_op_names=dict(op_name=['wide_embedding', 'deep_embedding'])))
     return train_conf

LanguageModeling/BERT/run_pretraining.py

@@ -37,8 +37,6 @@ parser.add_argument("--log_every_n_iter", type=int, default=1, help="print loss
 parser.add_argument("--data_dir", type=str, default=None)
 parser.add_argument("--data_part_num", type=int, default=32, help="data part number in dataset")
 parser.add_argument('--use_fp16', type=str2bool, nargs='?', const=True, help='use use fp16 or not')
-parser.add_argument('--use_boxing_v2', type=str2bool, nargs='?', const=True,
-    help='use boxing v2 or not')
 
 # log and resore/save
 parser.add_argument("--loss_print_every_n_iter", type=int, default=10, required=False,
@@ -164,8 +162,6 @@ config.train.model_update_conf(_BERT_MODEL_UPDATE_CONF)
 
 if args.use_fp16:
     config.enable_auto_mixed_precision(True)
-if args.use_boxing_v2:
-    config.use_boxing_v2(True)
 
 
 @flow.global_function(config)
@@ -206,9 +202,6 @@ def main():
 
     flow.config.gpu_device_num(args.gpu_num_per_node)
     flow.env.log_dir(args.log_dir)
-    if args.use_boxing_v2:
-        flow.config.collective_boxing.nccl_fusion_threshold_mb(8)
-        flow.config.collective_boxing.nccl_fusion_all_reduce_use_buffer(False)
 
 
     if args.node_num > 1:

LanguageModeling/BERT/run_squad.py

@@ -37,8 +37,6 @@ parser.add_argument("--log_every_n_iter", type=int, default=1, help="print loss
 parser.add_argument("--data_dir", type=str, default=None)
 parser.add_argument("--data_part_num", type=int, default=32, help="data part number in dataset")
 parser.add_argument('--use_fp16', type=str2bool, nargs='?', const=True, help='use use fp16 or not')
-parser.add_argument('--use_boxing_v2', type=str2bool, nargs='?', const=True,
-    help='use boxing v2 or not')
 
 # log and resore/save
 parser.add_argument("--loss_print_every_n_iter", type=int, default=10, required=False,
@@ -160,8 +158,6 @@ config.train.model_update_conf(_BERT_MODEL_UPDATE_CONF)
 
 if args.use_fp16:
     config.enable_auto_mixed_precision(True)
-if args.use_boxing_v2:
-    config.use_boxing_v2(True)
 
 
 @flow.global_function(config)
@@ -202,9 +198,6 @@ def main():
 
     flow.config.gpu_device_num(args.gpu_num_per_node)
     flow.env.log_dir(args.log_dir)
-    if args.use_boxing_v2:
-        flow.config.collective_boxing.nccl_fusion_threshold_mb(8)
-        flow.config.collective_boxing.nccl_fusion_all_reduce_use_buffer(False)
 
 
     if args.node_num > 1:

LanguageModeling/BERT/run_squad_predict.py

@@ -38,8 +38,6 @@ parser.add_argument("--log_every_n_iter", type=int, default=1, help="print loss
 parser.add_argument("--data_dir", type=str, default=None)
 parser.add_argument("--data_part_num", type=int, default=32, help="data part number in dataset")
 parser.add_argument('--use_fp16', type=str2bool, nargs='?', const=True, help='use use fp16 or not')
-parser.add_argument('--use_boxing_v2', type=str2bool, nargs='?', const=True,
-    help='use boxing v2 or not')
 
 # log and resore/save
 parser.add_argument("--loss_print_every_n_iter", type=int, default=10, required=False,
@@ -156,8 +154,6 @@ config.default_distribute_strategy(flow.distribute.consistent_strategy())
 
 if args.use_fp16:
     config.enable_auto_mixed_precision(True)
-if args.use_boxing_v2:
-    config.use_boxing_v2(True)
 
 
 @flow.global_function(config)
@@ -198,9 +194,6 @@ def main():
 
     flow.config.gpu_device_num(args.gpu_num_per_node)
     flow.env.log_dir(args.log_dir)
-    if args.use_boxing_v2:
-        flow.config.collective_boxing.nccl_fusion_threshold_mb(8)
-        flow.config.collective_boxing.nccl_fusion_all_reduce_use_buffer(False)
 
 
     if args.node_num > 1: