* Fix fused_attention english doc test=document_fix

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

This reverts commit 10f9644c.

* ctc grad compute on gpu

* add trainer desc config to distributed strategy

* code style modified

* data_feed set lod

* fix bug

* code style

* fix bug

* save load

* save load

* save unittest

* add unittest of the_one_ps

* unittest

* add todo in communicator sendsparse

* support inference for quantized matmul_v2

* undate code style

* code style

* add align for WorkQueue

* add spinlock

* merge develop

* merge

* Add EventsWaiter

* update

* update

* update Error MSG

* update EventsWaiter

* Add Cancel For ThreadPool

* Add UT for Cancel

* fix Cancel

* Fix bug when pre_layer_norm is false.

* Refactor InterpreterCore code

* make tuple

* change api for support trt8

* fix:change api

* fix device docs;test=document_fix

* update __init__.py

* [ROCM] add custom op support, test=develop

* remove debug codes, test=develop

* GeneratePass support attr condition and mapping, test=develop

* fix coverage, test=develop

* 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

* add dcnv2 plugin

* revise completion for backward

* revise completion for update

* revise completion for update

* update unitest

* Add WIP version of elementwise_div_mkldnn without working dy grad

* Add dy gradient calculation implementation, disable broadcast tests

* Readd removed tests from static_mode_white_list

* Add bfloat16 gradient tests, remove int8 and uint8 support

* - Change the way dy grad is calculated to improve performance
- Refactor BinaryMKLDNNHandler to use a default parameter

* Change copyright year

* Refactor as suggested

* Attempt to bypass CI Approval
not accepting max_relative_error

* Fix formatting issue

* gloo hdfs set check & gloo connect retry

* add vlog

* print gloo connect addr & add vlog

* .

* modify vlof

* modify vlog

* modify vlog

* Update __init__.py

deleted extra clear_model

* add DenseNet

* 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

* clean debug code
Co-authored-by: NXiaoxu Chen <chenxx_id@163.com>

* Layer.to reutrn self

* add device required

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

* fix bug

* fix matmul dim error

* fix wrong dim check in matmul

* fix fftshift/ifftshift on static mode
* update roll_op version
* add more test cases for fftshift/ifftshift

* add eigvalsh with is_test

* add eigvalsh op

* fix backward bug

* forward and backward, float and complex, unittest

* remove eigvalsh_helper.h

* remove changes of cusolver.h

* fix unittest

* fix unittest bug

* update code following eigh

* fix test

* update lapack

* pull develop

* update funcor

* fix unittest bug

* fix details

* add tensor_method_func

* fix notes

* remove additional warnning in layer.to

* remove additional warnning in layer.to

* remove additional warnning in layer.to

* remove additional warnning in layer.to

* remove additional warnning in layer.to

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；