* remove slelected rows all reduce over height check

* polish unittest

* refine jit.save/load to add support for other method, not only forward

* refine the code based on unit tests

* Add unit test for the code

* Add unit test for the code

* Modify the code according to the unit test

* Delete useless comments, save only one info file, etc.

* remove static_mode_white_list.pyc

* edit the code that generate 'extra_var_info'

* fix random failure

* use two input(x)

* fix model

*test=kunlun

* Add bfloat16 softmax and gelu

* Add pass attr bfloat16_enabled_op_types

* Changes from review

* Fix bug in convert_call

* refine unittest

* refine code

* refine code

* fix unittest failed

* add assert

* - Added sum bf16 oneDNN

test=develop

* - Fix to UT of sum bf16

test=develop

* add paddle.io.ComposeDataset & paddle.io.ChainDataset. test=develop

* rm cv2 dependence of normalize

* add conv_bias pass version test

* update according to reviews

modified sample code of CTCLoss & ctc_loss by remove disable_static() & print([.*].numpy()) & alias, test=document_fix (#28403)

* update DataLoader doc. test=develop

* skip reader op in mixed_precision decorator

* add ut

* add broadcast_shape api

* add ut

* follow comments

* add example code, test=dodument_fix

* update example code, test=document_fix

* enhance the op_version_registry, test=develop

* add unittests, test=develop

* enhance the op_version_registry, test=develop

* fix bugs, test=develop

* revert pybind_boost_headers.h, test=develop

* fix a attribute bug, test=develop

* fp16 result ok

* change -DWITH_NVINFER_PLUGIN toconfig.EnableTensorRtOSS

* auto detect special slice op converter for ernie with trt oss

* ernie oss only support fp16

* fix special_slice_plugin serialize bug

* matmul in tensorrt ok

* ernie unittest ok

* add matmul tensorrt unittest

* remove demo code

* pool2d cuda kernel supports fp16

* fix compile issue of template

* add ut

* Add rnn_op.
test=develop

* Fix rnn_op grad maker's drop_empty_grad.
test=develop

This PR is follow up of #28213. On that PR we tried to decrease GPU usage, however the CI still randomly failed. So I added retry logic for the initialization of nccl and cusolver. If the initialization failed, we can retry to avoid the random failure.

* add generate_proposals_v2 op

* Fix lr setting of AdamW when lr is an instance of LRScheduler.
test=develop

* Fix static graph test mode in test_adamw_op.py.
test=develop

* [Dy2stat] Support to modify value of buffer tensor

* remove "defaultTest"

* fix name confliction

* update version pybind to v2.4.3, test=develop

* update unittests, test=develop

* check and fix tensor and scalar type promotion

* fix else branch error

* fix scalar method error

* fix test_math_op_path unittest

* add future division for unittest

* rm useless bin file

* return pil image as default according backend

* Add median api.

* Add median api.

* Add median api.

* Add median api.

* Add median api.

* enable test_parallel_executor_fetch_isolated_var, test=develop

* add enable_static, test=develop

* set test_parallel_executor_fetch_isolated_var RUN_TYPE=DIST, develop=test

* Enhance multiclass_nms to support LoD for dygraph mode

* fix some error in multiclass_nms

* update GetLodFromRoisNum to GetNmsLodFromRoisNum