set net.forward to original forward function in flops when net is a dy2stat model.

* fix cusparse compile bug in CUDA11.2, test=develop

* modify sparse_attention docs, test=document_fix (#36554)

* modify sparse_attention docs, test=develop

* add warning

* add warning ,test=document_fix

* fix_qkv_plugin: half_scale

* [Paddle-Inference] fix_qkv_plugin: fix half scale

修复了fused_transformer_encoder_layer fine-tune过程发现的一些问题：

    fused_attention_op添加attn_mask=None的支持：PR
    pre_layer_norm处理问题：PR
    参数处理，计算错误的问题：PR
    add_bias计算错误问题：PR
    添加pure fp16的支持：PR

修复了一维Tensor在使用省略号(...)索引时维度检测异常的问题。

Co-authored-by: NPei Yang <peiyang@baidu.com>

* add mlperf optimization PRs

* update

* fix rnn grad bug when num_layers is set 2 and dropout_prob is set 0

* add more test for rnn

    Renamed the variable and function
    Removed the original template function
    Removed the tests_properties in CMakeLists.txt

att,Fix issue:36902

* fix cusparse compile bug in CUDA11.2, test=develop

* fix bug

Cherry-pick #36525

2. add complex data type support for paddle.shape at graph assembly.

Cherry-pick PR #36511

* fix device docs;test=document_fix

* update __init__.py

* add paddle.version.cuda and paddle.version.cudnn API

* fix little bug

* fix bug

* add doc string

* fix mkdir error

* fix windows path

* fix new paddle/version path

* fix unittest

* fix format

* [cherry-pick 2.2]support quantization of bert

support quantization for maumul_v2

* Update quantization_pass.py

* Fix fused_attention english doc test=document_fix

* change api for support trt8

* Revert "Align CTC grad scale same with ESPNet (#34729)"

This reverts commit 10f9644c.

* ctc grad compute on gpu

* Fix bug when pre_layer_norm is false.

* update fft api path (#36219)

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

* fix fft axis (#36321)

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

* use unified external error message for cufft api (#36114)

* fft: modify sample code result (#36325)

* dynamic load mkl as a fft backend when it is avaialble and requested (#36414)

* add rocm support for fft api (#36415)

* move signal apis

* move fft and signal API path (#2)

* move signal apis

* move fft.py and signal.py to paddle/, fix typos

* fix relative imports from fft.py and signal.py

* fix typos in signal.py (#3)

* move signal apis

* move fft.py and signal.py to paddle/, fix typos

* fix relative imports from fft.py and signal.py

* fix typos

* disable Cache when CUFFT_VERSION >= 10200 (#4)

* move signal apis

* move fft.py and signal.py to paddle/, fix typos

* fix relative imports from fft.py and signal.py

* fix typos

* Add LRUCache for fft plans

* add LRUCache for cuff and hipfft (#5)

* move signal apis

* move fft.py and signal.py to paddle/, fix typos

* fix relative imports from fft.py and signal.py

* fix typos

* WIP: add cache

* delete move constructor and operator= for CuFFTHandle and FFTConfig

* remove log from CuFFTHandle and FFTConfig

* add lrucache for fft rocm backend

* disable LRUCache when CUFFT_VERSION >= 10200

* disbale copy and move for hipFFTHandle; format code
Co-authored-by: NXiaoxu Chen <chenxx_id@163.com>

* remove debug message of cufftHandler

* roll_op: support Tensor as input for shifts (#36727)

* fix fftshift/ifftshift on static mode

* update roll_op version

* add more test cases for fftshift/ifftshift
Co-authored-by: Nzhiboniu <31800336+zhiboniu@users.noreply.github.com>
Co-authored-by: Nchenfeiyu <chenfeiyu@baidu.com>
Co-authored-by: LJQ❤️ <33169170+lijiaqi0612@users.noreply.github.com>

show paddle traceback after last user code traceback

Co-authored-by: Nzlsh80826 <zlsh80826@gmail.com>

本PR是fused_transformer的layer层代码，包含FusedFeedForward的layer层代码和FusedTransformerEncoderLayer的代码。

* bugfix: only check backend when mode == Collecive

Update `cond` English document

* fix BatchNorm for fp16

功能：本PR的目标是提高attention模块的计算性能。
为了减少框架层对op的调度开销，本PR通过在C++层手动实现attention模块，对外提供attention 大op；
为了减少防存开销，本PR采取了两种优化方法：
（1）在q,k,v计算时通过共享输入X，将该处的gemm，transpose和bias add从三次调用减少为一次；
（2）使用kernel融合优化技术，在不同cuda kernel之间通过寄存器传输数据；