add sync_batch_norm_bn and deliver indices_dict 

* cherry-pick #46359 and resolve conflict

修改一处Slice的bug

* fix inference perfermence problem caused by selecting cudnn kernel for softmax

* recover use_cudnn in opmaker of softmax

* return proper state

* fix for dims

* fix
Co-authored-by: Njakpiase <jakpia21@gmail.com>

[Cherry-pick][Release/2.4]Refine the memory usage of fused_attention and fused_feedforward ops (#47235)

* fix fused_attention fused_feedforward

* fix ci

* fix ci

* fix ci PADDLE_GET_CONST

* fix ci ut

* Fix bug of reduce_sum op. When input.numel() > INT32_MAX, its result
is wrong.

* support pure bfloat16

* support bf16 linear

* update PR to pass CI

* tiny fix where_grad_kernel.cu

* Support bfloat16 type for reducer and sharding.

* Fix some bug.

* Polish code.

* Polise code.

* Add bfloat16 datatype in fill_grad kernels.
Co-authored-by: Nsneaxiy <sneaxiy@126.com>
Co-authored-by: Nsneaxiy <sneaxiy@126.com>

* forbid Conv2D into Paddle-TensoRT when weight is not persistable.

* Add infer prune function

* add fusion op

* add audio feature dataset

* fix coding style

* fix coding style2

* rm librosa

* rm voxceleb

* rm librosa in test

* add scipy fftpack

* add functional

* fix setup

* fix setup2

* rm colorlog

* refactor dataset __init__.py

* fix converage

* fix librosa import error

* fix windows test

* fix windows ci

* rm datasets

* fix setup

* remove testdata

* add librosa in requirement

* add librosa in requirement2

* change librosa to 0.8.1

* update ci docker

* fix ci error

* fix ci error2

* fix ci coverage

* fix converage

* fix coverage

* rm audio_base in test, notest,test=coverage

* fix copyright

* rm backend

* add datast in __init__

* rm compliance&&add function test

* fix setup

* fix windows

* fix windows2

* fix test timeout

* add backend & datasets

* fix bugs

* fix ci time issue

* add dataset test

* rm test_audio_feature

* avoid windows isssue, tmp

* note windows isssue

* skip windows issue

* refactor dataset test

* add dataset.py

* fix dtype in layers.mfcc

* fix ci-static-check

* fix dtype in layers.mfcc && fix ci-static-check

* add relative accuracy

* modity API.spec

* skip cuda11.2 test

* skip cuda11.2 test2

* skip cuda11.2

* change dataset name

* fix format

* update api.spec

* update api.spec2

* fix coverage

* add dataset test

* rm download load dict

* rm download load dict in init

* update api.spec3

* fix dataset coverage

* fix coverage

* fix coverage2

* restore api.spec

* restore api.spec2

* fix api-spec 3

* fix api-spec 4

* fix api.spec

* fix api.spec6

* refactor init_backend

* fix typo

* change paddleaudio backend set

* fix get_current_audio_backend()

* fix format

* fix format2

* remove format in parameters

* fix format2

* add warning massage in wave_backend && remove redundant audio util

* rm audio util in print_signatures

* fix format3

* add tess dataset license

* format warning

* add more info in warning msg

* add paddleaudio version check

* replace dataset esc50 with tess

* add tess dataset && rm numpy transform in dataset.py

* fix set audio backend bug

* fix equal error

* fix format && coverage error

* add api example

* fix format

* fix error

* fix typo

* add noqa in __init__

* fix backend doc example error

* rm seed in dataset

* update bakcend example

* fix typo

* fix typo

* fix example err

* fix typo

* fix ci dataset test

* fix example fil

* try to fix ci

* clean dataset doc

* change get_current_audio_backend to get_current_backend

* creplace paddle.audio.backends.info with paddle.audio.info, same with load, save

* fix ci error

* repalce api in test_audio_backend

* fix save&&set_backend exmaple

* Simplify the codes of conv. (#45966)

* Enable to record whether the conv algo is got by exhaustive search to fix autotune cache bug. (#47065)

* add fused_attention_pass. test=develop

* support fp16. test=develop

* fix format. test=develop

* Fix quantize model deploy bugs when using MKLDNN (#45920)

* fix immutable op quantize bugs

* fix

* fix build bug

* fix test

* notest,test=inference

* fix ppyoloe acc drop bugs

* fix test

* fix test

* add test

* fix

* fix

* fix test

* fix refined name bug

* fix test

* bias fix

* fix matmul weight dequant bug

* re-ci

* fix tester

* fix test

* fix tester

* update weight dequantize func

* update code

* update test for converage

* update test

* update cmake

* update cmakelist

* update code

* rerun ci

* remove useless code

* re-ci

* update code

* update code

* fix header

* update code for log

* stride_to_24

* fix CI failing

* Enhance the layernorm shift partation fuse op when shift size > 0 (roll shifting)
* fix cherry-pick test

Fix some operators when the tensor.numel() > INT32_MAX

support pure bfloat16 for more ops

* Add unsigned int8 propagation

* Add or modify unit tests

* Correct concat scale checking

* Apply review suggestions

* Corrections
Co-authored-by: Njoanna.wozna.intel <joanna.wozna@intel.com>

* Fix bug of reduce_sum op. When input.numel() > INT32_MAX, its result
is wrong.

* Support allow_partial switch, which can be configure in
pipeline_configs. If sent tensor are not the same from
different hosts, they shouldn't been sent partially and
then concated as a whole tensor.

* Change name allow_partial to enable_partial_send_recv.

* Add global variable _enable_partial_send_recv

[CherryPick][Dy2St]Fix recurrent op eager deletion pass error in dy2st

Construct exec and ctx only once in cond op to speed up

Add symbolic shape deduction function for unfold, scatter_nd_add, p_norm, grid_sampler, pad3d, etc (#46291) (#47003)

新增sparse.is_same_shape、sparse.reshape、sparse.transpose 三个API

* [Dygraph] Fix performance of pp+mp by using send/recv_calc_stream instead of send/recv (#46116)

* [Dygraph] Fix Perf of FusedFeedForward and FusedAttention with AllReduce (#46780)

* update

* draft with debug print
* remove debug print
* bug fix for ci

* Support both use_calc_stream and sync_op in send recv APIs (#46023)

* Support both use_calc_stream and sync_op in allgather API (#46295)

* Support both use_calc_stream and sync_op in collective communication API (#46761)

* Move group and all reduce from collective to communication (#45848)

* Completes bfloat16 dtype for collective api in eager mode (#45844)

* Fix collective APIs cannot be recognized when building docs (#46962)
Co-authored-by: NLiYuRio <63526175+LiYuRio@users.noreply.github.com>

cherry-pick : #46322, #46245
Sparse API 支持静态图

* paddle-inference support custom-ops
Co-authored-by: NZhixin Yao <zhixiny@graphcore.ai>

* fix tolower
Co-authored-by: NZhixin Yao <zhixiny@graphcore.ai>
Co-authored-by: NZhixin Yao <zhixiny@graphcore.ai>

* [BUG]Fix expand_as_v2 bug while X and Y with different dtype

* fix commit

* fix reshape2 opteller;
add elementwise min/max register for tensorrt

Fix set_value failure when source tensor is fp16 Dtype and destiny value is a number
(dev PR link:#46801)

* Revert pool+grad oneDNN kernel conversion (#45989)

* [PHI] transpose2_grad op migration (#46139)

* op migrated, Copy(OneDNNContext, ...) added

* mutable_data & op registration in fluid removed

* refactoring

* OneDNNGetDataType to uppercase

* missing cpu check added, handler moved to .h file

* name changed to transpose_grad

* Copy changed back to TensorCopy

* Resizing corrected, Copy(OneDNNContext) removed
Co-authored-by: NPiotr Paturej <48731682+piotrekobi@users.noreply.github.com>
Co-authored-by: NPaulina Gacek <paulina.gacek@intel.com>