Merge pull request !3687 from yankai10/merge123

Merge pull request !3705 from vlne-v1/amp_doc

Merge pull request !3657 from yangzhenzhang/add-stridedslice-op

Merge pull request !3650 from lijiaqi/add_calculation_formula

Merge pull request !3512 from Guomenghao319/add_python_distribute_pretrain_script

Merge pull request !3711 from fangzehua/topkgrad

Merge pull request !3699 from fangzehua/fix_scatterop

Merge pull request !3700 from zhouyaqiang0/master

Merge pull request !3655 from chenweifeng/broadcast_to

Merge pull request !3707 from chenweifeng/LayerNorm

Merge pull request !3679 from huangbingjian/report_float16

Merge pull request !3667 from cjh9368/main_master

Merge pull request !3701 from ling/pad

Merge pull request !3405 from gziyan/fix_optimizer_parallel

Merge pull request !3656 from limingqi107/master

Merge pull request !3691 from XunDeng/pp_poc_v3

Merge pull request !3631 from kisnwang/fix-subgraph-maketuple-error

Merge pull request !3653 from panfengfeng/add_epoch_num_description_master

Merge pull request !3672 from hexia/fix_input_check

Merge pull request !3661 from shenwei41/sw_master

Merge pull request !3664 from hanjun996/master

Merge pull request !3668 from ougongchang/fix_collect_dataset

Merge pull request !3522 from wanghua/c75_tinybert

Merge pull request !3660 from meixiaowei/master

Merge pull request !3665 from zhouyaqiang0/master

Merge pull request !3341 from tom_chen/maxpool_with_argmax

We collect graph and dataset graph in begin stage before,
If there compile graph fail in GPU, we also collect graph
and dataset graph to summary dir, it will confuse user.

So we collect graph and dataset graph in step end stage now,
If there compile graph fail, we will not collect graph and dataset
graph.