* fix distribute_fpn_proposals

* fix bug

* Add fused_feed_forward pass for semi-automatic static graph training.

* Add fused_feedforward property in parallel_executor.cc

* Polish code.

* Polish fused feed_forward pass code. Support use_dropout1 and
use_dropout2 option.

* Support model parallel in fused_feedforward pass.

* extract common methods to reuse

* add header for transpose ops

* fused_transpose

* Split big function

* transpose2 tests

* fused_transpose

* Apply extra attributes

* add pbtxt file

* update pbtxt

* Merge develop

* add more strict op compats

* code  style

* remove mkldnn_data_type

* unify SetOutMemDescWithReshape2FuseSupport

* adjust quantize-dequantize for transpose

* remove appendact

* transpose2 quantization

* fix int8 tests

* adjust transpose_op to current develop

* delete fusion code from transpose_kernel

* add fused transpose to NHWC unittest

* change order

* add nms3 register output defs

* remove nms from set

* remove nms from set

* test_logit_op

* add cudaKernel to replace eigen impl

* bf16 unit test CI

* [AMP OP&Test] Fix fp16 check_grad when user_defined_grads are not None

* fix cond

* [CustomOP Optional] CustomOP supports optional Tensor

* fix test_custom_concat, add pytest to CMakeLists

* remove net_drawer.py

* remove memory_analysis.py

* remove test_memory_analysis.py

* add autogen code for index_add op

* bug fixed

This reverts commit 3b2cd23a.

* fix raw_program_optimizer not apply when using amp

* fix CI

* supoort auto generate p_norm

* fix bug in backward

* support auto generate for dirichlet

* use uppercase in args

* use op_compat for name mapping

* max comp

* fix

* add test

* fix

* fix

* fix

* fix

* fix test

* fix api

* fix ninja error

* fix_ninja_error

* fix ninja error

* fix r-200 ci ninja error

* add fused_linear_param_grad_add_kernel

* fix compile error

* remove flag

* fix ci compile error

* fix ci compile error

* revert pylayer revision

* fix ci ut

* improve performance

* add Attribute system to new ir

* set StorageType to Storage in Type and Attribute

* refine strAttr

* refine name of StrAttribute

* add DictionaryAttribute

* refine code

* refine dic_attr

* refine code

* Set DictionaryAttribute ParamKey is map

* refine code

* refine code by comment

* refine code

* refine code

* refine code

* refine code

* fix complie bug

* refine code

* add const for Attribute storage

* allow return none when stop_gradient=True

* remove useless code

* refine code

* refine code

* fix test cast

* change more test

* add more tests

* test_accuracy_op

* add create_test_fp/bf16_class

* cast after calculation

* change convert_uint16_to_float_ifneed

* delete TestAccuracyOpFp32 according to PR comment

* fix the rtol setting rules in bfloat16 forward

* [AutoParallel] support bloom

* fix import

* align amp and bf16

* update func name

* clipbyglobalnorm and add_n support bf16

* upgrade amp strategy api

* update bf16 unittest

* fix static clip

---------
Co-authored-by: Nliangjianzhong <liangjianzhong@baidu.com>
Co-authored-by: NAurelius84 <zhangliujie@baidu.com>

* remove DataFeeder.feed_parallel and decorate_reader

* fix CI

* fix CI

* replace assert false with AssertionError

* 修改配置文件多余的部分

fix embd plugin: S = mask_id.d[1]

* fix batch_norm custom_vjp dtype

* add nhwc test example

* fix typo