* [Auto Parallel]  Add the unified cluster representation

* [Auto Parallel] Add the graph class for physical mapping

* [Auto Parallel] Add the simple physical mapper

* Set the timeout of the mapper

* Merge the upstream develop unittests cmake files

* Fix a bug of the process group

* Remove mapper unittest from platforms which is not GPU

* Move the instantiation of process group after resharding

* Add the local id for devices

* Update the rank mapping format

* Add some comments

* Remove the related files about mapping

* Remove unused rank_mapping unittest

* Improve the unittest coverage

* fix data parallel when VOCAB var in program

* test

* test

* rm test

* update

* update

* update

* add unittest

* update

* update save

* Fix bugs when bias is none for static graph for fused_attention op.

reset_inplace_version removes all inplace related records to VarBase/VariableWrapper, the essential purpose of which is to let you use inplace operations as if using its non-inplaced version, which of course will cause unexpected consequences if not used with care.

This is essentially a hack interface to satisfy one specific request

* add tdm sample

* add tdm sample in c++

* update tdm sample

* modify sample count

* fix conflict

* add set_date

* fix cmake error

* fix bug of proto

* update index_dataset proto

* update cmake

* fix error cmake

* fix cmake mkldnn

* fix cmake proto

* update cmake proto

* update cmake

* update rec

* update dataset

* update dataset

* update dataset

* updata dataset

* updata dataset

* updata coverage

* updata ci

* goback4

* fix npu ci

* add xxhash dep

* fix dropout static when axis != None

* update dropout test

* add dropout test

* fix test

* Update test_dropout_op.py

* Update test_dropout_op.py

* fix testcase

* fix testcase

* Update test_dropout_op.py

* fix testcase

* fix testcase

* optimize perf

* add new test

* fix testcase

* [NPU] add int64 support for argsort op

* [NPU] delete debug codes

* [NPU] add NPU kernel for prior_box op

* [NPU] delete debug codes

* fix. test=develop

* fix. test=develop

* fix. test=develop

* fix. test=develop

* fix. test=develop

* [heterps]bug fix for _run_from_dataset

* fix heter_server.cc

* fix launch_utils.py

* fix heter_section_worker.cc

* fix. test=develop

* fix. test=develop

* add new API paddle.nn.initializer.Dirac

* fix doc

* fix program cache key

* bug fix

* fix cache problem

* remove unused code

* hot fix for dataloader thread error

* polish comment

* fix type in comment, test=document_fix

* add zeropad2d v0.1

* add zeropad2d v0.2

* add zeropad2d v0.3

* add zeropad2d v0.3

* add zeropad2d v0.3

* add zeropad2d v0.4

* add zeropad2d v0.5

* add zeropad2d v0.5 codestyle

* add zeropad2d v0.5 codestyle

* add zeropad2d v0.6 functional

* add zeropad2d v0.6 functional

* add zeropad2d v0.6 functional

* skip compiled program

* fix ut

* pybind core

* set use psgpu

* matmul_convert_int8

* matmul_convert_int8

* matmulconvert_int8

* Matmul_int8_convert: tensor*tensor

* Matmul_int8_convert: tensor*tensor

* Matmul_int8_convert: tensor*tensor

* adapt auto search

* adapt auto search

* fix matmulv2 compatible

* del debug

* [Auto Parallel]  Add the unified cluster representation

* Add the local id for devices

* Add some comments

* run dy2stat pure fp16 in Linear model

* no use self._pure_fp16_inputs

* add test and fix Adam error in dy2stat pure fp16 training

* use paddle.optimizer.Adam

* run test in gpu

* change test time for CI

* enlarge atol for test_resnet_pure_fp16

* refine code and enlarge atol

* make custom_white_list and custom_black_list take effect for AMP and pure fp16

* check tracer is not None

* use default atol

* change filter_size

* change atol and add some NOTE

* fix inplace bug

* fix custom grad input error

* add unittest

* fix inplace bug

Add support for bias is none for fused_attention op.

`paddle.utils.download` ：change to call `extractall` on tar/zip compressd file  to speed up the uncompress process when they includes many files

--- result of decompression speed comparison ---
1. dataset：https://paddlenlp.bj.bcebos.com/datasets/cnn_dailymail/cnn_stories.tgz, decompression time
：5m50s vs 20s
2. dataset：https://paddlenlp.bj.bcebos.com/datasets/cnn_dailymail/dailymail_stories.tgz, decompression time：33m20s vs 47s

* skip compiled program with places > 1

* fix corner case and add ut

* fix_nearest

* fix_nearest

* fix_nearest

* fix_nearest

* Added HCCL backend support in dynamic graph mode

* fix segmentation fault

* add ut

* Bug fix for snapshoting VariableWrapper with initialized tensor but empty allocation

* Added unittest for inplace&clear_gradient

* Add transfer_layout/dtype op

* clean useless codes

* fix unused var

* add optest in white.txt

* split into data_transfer.cc

* fix cmake

* modify according reviewer comment

* replace cast_op with transfer_dtype_op

* fix autoconvert

* fix merge parameter

* add isclose op, test=develop

* add isclose op, test=develop

* add isclose api, test=develop

* rm useless code

* rm useless code

* update python api of isclose

* add some unittest of isclose op, test=develop

[Dy2stat]Allow users to switch eval/train mode when using @to_static to decorate a function (#37383)

* Allow users to switch eval/train mode when using @to_static to decorate a function

* refine code for train() and eval()

* support zero dim for slice op

* support zero dim Tensor in set_value op

* polish some debug log