add gpt2 model for the paddlenlp

* refine transformer cpu

* delete run_pretrain.sh

* update docs

* fix readme

* update doc

* add / for dir

* refine eval batch size

* support AMP training

* change the dir name

* add static step final save

* refine doc

* alter method name

* update train

* update eval

* fix doc

* fix version bug

add peoplesner, fix msraner (#5183)

fix lac bug (#5190)

* add pure fp16 training.

* Add lookup_table in custom_black_list during the pure fp16 training.

* add transformer for rc0

* predict no_grad

* add benchmark readme

* add predict for static

* update doc for fleet api

* typo

* add dygraph amp support for transformer

* according params use_amp to decide use dygraph amp

* add fleet, test=develop

* update codes for senta benchmark

* support pad seq and bsz multi

* alter vocab size

* alter md5sum

* root for reader

nlp bert benchmark add gradient merge

* add static predict

* rm position_encoding_init

* fix static predict

* add the xpu support the the paddlenlp

* add the support xpu for the ernie and bert
Co-authored-by: Nwawltor <fanzeyang0904@hotmail.com>

change the format for the bert benchmark (#5160)

* fix incorrect usage of data loader

* refine code

* remove unused code

* refine code

add the suport the log format for the benchmark

* fix dygraph

* update decription

* upgrade MapDatasetWrapper

* minor fix

* refine code

* new chunk_eval

* np first

* add embedding example in embeddings.md

* add embedding.md link in toke_embedding readme

* fix windows url bug

* add pad factor to pad vocab

* shuffle batch

* update reader

* update benchmark