* fix compare op in for in the cuda device

* fix the paddle compare op for the broadcast

* fix precision of mp

* fix bug of seed

* fix dp

* print group

* BugFix with ParseInputDataType from LodTensorArray

* BugFix with ParseInputDataType from LodTensorArray

cherry-pick of #32972

* cherrypick for #32640 :add profile and fix dataset hang in heterps;test=develop

* cherrypick for #32640 :add profile and fix dataset hang in heterps;test=develop

* cherrypick for #32640 :add profile and fix dataset hang in heterps;test=develop

fix error log for reducer

fix doc

fix bug of utest

fix spawn

fix converage

* implement MHA order same as training

* fix fp16 compile issue on old architecture
Co-authored-by: Nzlsh80826 <rewang@nvidia.com>

* fix compile error on jetson platform

* remove unused head file

* rm decode_jpeg op on jetson platform

 [CHERRY-PICK2.1]Remove paddle_custom_op dynamic libraries, and link to FLUID_CORE on windows (#32583) (#32769)

* Remove paddle_custom_op dynamic libraries, change link to FLUID_CORE on windows, and check copy_to

* fix CI

修复了py_layer_op由于没有析构PyLayerContext造成内存(显存)泄露的问题。

原始pr：#32707

* clear 'BasicEngine' when an exception occurs in the backward. (#32546)

* clear 'BasicEngine' when an exception occurs in the backward.

* deal with conflict.

* deal with conflict.

* forward return any type. (#32661)

* base changes for fix

* minor change

* fix for bwd kernel

* removed unnecessary import

* implemented reviewers suggestions

* CI fix

cherry-pick:change softmax_with_cross_entropy_op's parameter name from softmax_switch to use_softmax (#32750)

* change parameter name from softmax_switch to use_softmax, test=develop

* cherry-pick:change parameter name from softmax_switch to use_softmax, test=develop

cherry-pick #32561

* add relu6_ hardsigmoid_ leaky_relu_ Inplace APIs

* add softmax_with_cross_entropy_ Inplace API

* add clip_ scale_ add_ subtract_ Inplace APIs

* add wlist

* fix parameter of scale api

* add add_n_ Inplace API and remove log_ Inplace API

* fix elementwise_add_ and elementwise_sub_ broadcast problem

* elementwise inplace api give error message before run the op

* use broadcast_shape in elementwise inplace op

* add 8 inplace apis that is auto generated

* add unittest for all inplace apis

* add decorator for inplace apis in static mode

* fix windows blas fail of exp inplace api, change array_equal to allclose

* add flatten inplace api

* add flatten unittest

* fix flatten unittest

* add decorator

* fix grad.numpy in test_pylayer_op

* unsupport softmax_with_cross_entropy_

* add test_inplace_softmax_with_cross_entropy to static_mode_white_list

* delete __all__ in inplace_utils

* delete activation inplace function and add Tensor.inplace_func

* change paddle.inplace_ to Tensor.inplace_

* fix little problem

* add paddle in inplace_utils

* add op read_file and decode_jpeg

* Add dlnne engine runtime

* Remove <const_cast> and remove unrelated modify with dlnne, +clang-format

* Add copyright message

* Add some paddlepaddle_pass to support more networks

* Add delete dropout_op pass
Co-authored-by: Ndenglin-github <82362191+denglin-github@users.noreply.github.com>

* Add bf16 uniform random initializer

* Remove duplicated section

* Change UT to CPU place only

* Put detail functions into anonymous namespace

This is cherry-pick of #32281

cherry-pick of #32666

后续修复计划是啥

- Executor is nt always having FLAGS_use_mkldnn set to true

* optimize update_loss_scaling_op by fused for loop to one kernel, test=develop

* remove useless while loop and optimize variable name, test=develop

* optimize variable name from out_addrs_tensor to out_addrs_mem, test=develop

* optimize variable name for readable by change prefix identifier from t_ to local_