提交 4e3210d6 编写于 作者: S ShawnXuan

open all reduce group params

上级 56cfb728
......@@ -16,8 +16,8 @@ def get_train_config(args):
train_config.train.primary_lr(args.learning_rate)
train_config.disable_all_reduce_sequence(False)
#train_config.cudnn_conv_enable_pseudo_half(True)
#train_config.all_reduce_group_min_mbyte(8)
#train_config.all_reduce_group_num(128)
train_config.all_reduce_group_min_mbyte(8)
train_config.all_reduce_group_num(128)
# train_config.all_reduce_lazy_ratio(0)
# train_config.enable_nccl_hierarchical_all_reduce(True)
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册