* 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；

* fix wrong trt dim when input dim is 2

* update leaky_relu and instance_norm converter unit test

* add instance_norm input dim check

[Paddle-Inference]Add MatmulV2ToMatmul convert Pass, fix (matmul_v2, matmul, mul) convert pass, fix (matmul, mul) op_teller (#36652) (#36737)

* Add FasterTokenizer Operator (#34491)

Add Tokenizer related functionalities for Transformer model in order that the process of training and predicting is consistent.

* support the text string as an input Tensor
* support the "VOCAB"unordered_map<wstring, int> as an input Tensor to lookup tokens
* Tokenizer used for BERT. This tokenizer applies an end-to-end, text string to wordpiece tokenization.
* It first applies basic tokenization, followed by wordpiece tokenization.

* optimize fast tokenizer

* remove const_cast
Co-authored-by: Nzhoushunjie <zhoushunjie@baidu.com>
Co-authored-by: Nwawltor <fangzeyang0904@hotmail.com>

* add op: fused_feedforward(backward) (#35611)

这个PR是fused_feedforward反向的代码

相关kernel实现：fused_dropout_act_bias, fused_residual_dropout_bias, fused_layernorm_residual_dropout_bias

fused_feedforward是一个融合算子，该算子对transformer模型的feed forward层的算子进行融合和封装，使得前端只呈现一个接口，通过融合减少部分访存和kernel launch的时间，以此提升性能。

* Move fused_attention and fused_feedforward functional api path to incubate (#36704)

将 #35905 和 #35843 PR中新增的的python api接口移到incubate目录下。

* fix bugs in HybridParallelClipGrad of hybrid_parallel_optimizer (#36237)

* fix bugs in HybridParallelClipGrad of hybrid_parallel_optimizer

* update

* update

* fix bugs in mp_layers、pp_layers and HybridParallelClipGrad (#36144)

* 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

* [HybridParallel]Rebuild code for pipeline (#36396)

* add no_sync for parameters sync

* add pipeline for moe

* [HybridParallel]Support fp16 in dygraph hybrid parallel (#36420)

* [HybridParallel]Support fp16 in dygraph hybrid parallel

* update

* update

* update for recompute

* add unittest of pp+fp16

* add unittest of recompute+fp16

* update

* modify ut

* modify ut of cond (#36475)

* fix bugs of ClipGradByGlobalNorm in HybridParallel (#36555)

* fix bugs of ClipGradByGlobalNorm

* add unittests

* add unittests

* [HybridParallel]fix bug of check_inf in fleet_base.py (#36651)

* fix bug of check_inf

* fix allreduce

* support ClipGradByGlobalNorm in sharding (#36012)

* support ClipGradByGlobalNorm in sharding

* support ClipGradByGlobalNorm in sharding

* test=allcase

* Update test_linalg_cond.py

* Update hybrid_parallel_util.py

* Update hybrid_parallel_util.py
Co-authored-by: NShenLiang <1422485404@qq.com>
Co-authored-by: Nzhaoyingli <86812880+zhaoyinglia@users.noreply.github.com>

This is a fusion operator to compute feed forward layer in transformer model architecture.

* Add bincount op

* upload cpu version

* fix unitest

* fix unittest

* fix unittest

* fix en doc

* add more test

* fix en doc

* add more test case

* fix test

* fix input vailidation

* fix input check

* fix unittest

* fix test

* fix en doc

cherry-pick

Support various length support for SelectedRows in GLOO::AllGather (#36637)

    In cpu parallel using gloo, add various length support for SelectedRows

* refine amp level

* fix typo

* update tracer._amp_level

* add slotrecord datafeed (#36099)

* fix multi-node (#36329)

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；

[cherry-pick] Support CPU Parallel in DataParallel Interface by GLOO to speed up training (#35745) (#36605)

* User specified backend (#35745)

* remove tensordot

* Revert "Add fused_dropout wrapper to ease use. (#36185) (#36640)"

This reverts commit 05d7e2fd.

* [hybrid] seed and dropout op support force-cpu (#35820)

* [HIP] fix op not support AMD GPU bug, the flag PADDLE_WITH_ROCM is invalid

* [HIP] fix op not support AMD GPU bug, the flag PADDLE_WITH_ROCM is invalid

* [HIP] fix op not support AMD GPU bug

* [hybrid] seed and dropout op support force-cpu

* [hybrid] seed and dropout op support force-cpu

* [hybrid] seed and dropout op support force-cpu

* [hybrid] seed and dropout op support force-cpu

* [hybrid] seed and dropout op support force-cpu

* [hybrid] fix seed ci failed issue

* add AsExtra for force_cpu of seed op

* Add fused_dropout wrapper to ease use. (#36185)

* [hybrid] static model parallel dropout support deterministic RandomSeedGenerator (#36228)
Co-authored-by: Nxiayanming <41795079@qq.com>
Co-authored-by: NLi Min <11663212+limin2021@users.noreply.github.com>

* Fix grid sampler

* Fix code format

cherry-pick prs

#36568
fix fc fuse compat problem

#36610
support lite xpu choose device id

#36010
update lite branch

#36628
add file exists check

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。

Add fp16 kernel for clip_op.

Refine comments for GradScaler state_dict.

* Add new API tensordot
cherry-pick #36273

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；