* fix calling bug of HybridParallelClipGrad

* fix bugs of HybridParallelClipGrad

* add unittest of pp with HybridParallelClipGrad

* fix bugs in mp_layers.py

* update

* fix bugs in pp_layers.py

* update

* Fix stop_gradient in RunProgramOp

* fix reference

* heterps:add fuse_allreduce op; test=develop
* add program_mode in minimize for pslib mode;test=develop

* add FLAGS_allreduce_record_one_event

* add more comments

* fix ut

* improve coverage

* fix ut, improve coverage

Add paddle.nn.functional.sparse_attention API

    本个PR主要将sparse_attention功能在python层进行了一层封装，OP的主体代码见：#PR35676

    此外，对于封装的python 接口，增加了相应的单测。

Sparse tensor core for convolution requires the input channel dimension is 2:4 structed sparse.
So we have to mask the input channel dimension for using sparse tensor core

* add reshard module

* fix conflict

* update reshard module

* update and add unitest

* update reshard module and unitest

* add more unitests

* enhance yolobox plugin

* [NPU] fix matmul_v2 and utils.run_check, test=develop

* remove debug files, test=develop

* fix install_check, test=develop

* fix doc, test=develop

* fix review comments, test=develop

* [NPU] fix set_value, test=develop

* fix typo, test=develop

* fix typo, test=develop

fix: `-1` is used when fft's axis is `0`

* add mish trt plugin, compile & install success, run error. test=develop
* modify code according to review
* add TRT_NOEXCEPT for mish trt plugin
* add unittest for mish trt plugin
* remove unnecessary check of mish in op_teller.cc
* fix some problem of trt8
* add check and modify unittest while converting mish to trt plugin
Co-authored-by: Ndengkaipeng <dengkaipeng@baidu.com>

* add skip case in trt converter ut

* disable group_norm trt plugin

Add use_cinn flag and use it to control whether we run PaddlePaddle using CINN.

Also add:

Replace PaddlePaddle graph with a CINN graph in a pass
PE Method to feed data and run the graph by CINN

* Add new API tensordot

* Set timeout value 400 for UT; Fix format for EN docs

* Set timeout value 1000 for UT; Fix format for EN docs

* Remove some input check

* Coding style improve: don't compare boolean values to True or False
using ==

* update fft api path
* add sample code for ihfft2
Co-authored-by: Nchenfeiyu <chenfeiyu@baidu.com>

* support ClipGradByGlobalNorm in sharding

* support ClipGradByGlobalNorm in sharding

* test=allcase

对于__getattr__重载后不满足条件的参数，全部抛出AttributeError异常，达到与未重载版本一致。

* support CUDA Graph on PE

* add ut, fix CI compile

* reduce memory consumption

* fix CUDA 10 CI

* improve coverage

* improve python coverage

* [NPU] support NCL and NCL for BatchNorm, test=develop

* [NPU] remove debug files, test=develop

* update, test=develop

Add python interface of subgraph: 1. all_sub_graphs() 2. get_sub_graph(idx)

* Added oneDNN BF16 relu

* fixed typo

* refactored test, review fixes

* fix bugs in HybridParallelClipGrad of hybrid_parallel_optimizer

* update

* update

* Remove unused header.

* Use ConvMKLDNNHandlerT for conv2d INT8.

* Use absolute module path to import.

* tmp

* added concat BF16/FP32 BWD oneDNN kernel

* minor change

* minor change

* fix for CI

* added formatting

* Reverted deleting static keyword

* added reviewers suggestions

* reverted deleting concat bf16 test file

* fixed concat tests

* [NPU] modify transpose2 and index_select_grad kernels for model xlnet

* add transpose2 int64_t unit test

* add more transpose2 unit tests

* update test_transpose_op_npu.py

* fix raw optim

* pre-commit test file
Co-authored-by: Nsneaxiy <sneaxiy@126.com>

* add basic support for CUDA Graph

* fix ci compile error

* fix LOG print, fix windows CI

* follow comments and update

* small fix for default ctor

* fix rocm compile error

* fix CPU compile error

* update func name

* skip cpu

* update unittest

* update unittest