* cherry pick from #43397

* fix code

* Rename dropout is test (#43098)

* replace dropout_is_test with is_test.
* improve atol on a100.

* fused_attention fused_feedforward api support Model Tensor Parallel (#42985)

* fix is_test bug in fused_feedforward. (#43508)
Co-authored-by: NLi Min <11663212+limin2021@users.noreply.github.com>

* fix bug of strided_slice (#43388)

* fix stride_slice bug

* fix bug

* fix bug of infer shape for slice (#43443)

* [EinsumOp] Polish forward logic and backward logic for optimize (#42603)

* change logic for optimize

* modifty

* merge

* change einsum_v2 as default and add new flags: FLAG_einsum_opt=1|0 (#43010)

* [EinsumOp] Make EinsumOp support bfloat16. (#43085)

* change einsum_v2 as default and add new flags: FLAG_einsum_opt=1|0

* make EInsumOP support bf16

* add unittest for BF16

* add condition for test_BF16

* fix bugs

* fix

* change the backward api to fit einsum op

Reduce amax/amin frobenius_norm_kerne原始实现为Eigen实现，文件编译时间较长，因此本PR将其替换为KP实现
删除DefaultElementwiseOperator中重复功能支持，减少elementwise_double_grad OP编译时间

删除在2.3 对比whl包大小。

解决onnxruntime后端依赖的protobuf跟框架或外部protobuf版本冲突问题

Delete ElementwiseKernel in BroadcastKernel
减少所有Broadcast中重复功能调用，同时减少编译时间和问题体积

删除Broadcast function中rank例化以及Elementwise调用，降低编译时间。
从develop分支中的#42645 PR修改而来，由于develop分支与release分支相差较大，无法实现cherry-pick，因此针对release2.3重新提交PR.
Broadcast中关于rank的例化会导致底层模板展开较多，造成reduce_sum_grad_kernel.cu.o文件体积过大，修改后可以降低.o体积及编译时间

删除判断build目录大小和预测库大小检查功能。该功能是和develop比较，会存在差异，在release任务中取消判断

* Fix cond_block_grad error when handle no need grad vras

* Add comment and UT

* fix_multi_int8 (#42977)

* cherry-pick fix_multihead_int8

cherry-pick PR #42700

* pdnode_compare

* panode compare

* pdnode_compare

* [MLU] add mlu new profiler (#41138)

* [MLU] add mlu new profiler

* fix format

* [MLU] support add callback to stream (#41831)

* [MLU] add gather mlu kernel (#41969)

* [MLU] add mlu activation kernels (#41751)

set attr ignoreIndex type to string for custom_nllloss_op

部分 cheery-pick of #42534

    add class NameScopeHelper for adding namescope info
    添加更多 种类优化器状态的映射
    为 IpuStrategy 添加 compilation_progress_logger option 用于输出 编译进度
    部分代码清理和杂项优化

Reduce the number of threads per block of deformable_psroi_pooling to solve the bug where too many resources requested for launch (PaddlePaddle#42531) (#42533)

* Reduce time variation for cuda_managed_memory_test (#42458)

* Disable standalone executor for test_tensordot (#42476)

* Fix the race condition in cumsum operator

* Optimize cumsum operator
Co-authored-by: NLeo Chen <39020268+leo0519@users.noreply.github.com>

* enable graph-engine to return all id (#42319)

* enable graph-engine to return all id

* change vector's dimension

* change vector's dimension

* enlarge returned ids dimensions

* change sample result's structure to fit training (#42426)

* enable graph-engine to return all id

* change vector's dimension

* change vector's dimension

* enlarge returned ids dimensions

* add actual_val

* change vlog

* fix bug

* bug fix

* bug fix

* fix display test

* singleton of gpu_graph_wrapper

* change sample result's structure to fit training

* recover sample code

* fix

* secondary sample

* add graph partition

* fix pybind
Co-authored-by: NDesmonDay <908660116@qq.com>
Co-authored-by: NDesmonDay <908660116@qq.com>

* fix bug of batch_norm_grad kernel with fp16

* format code

* fix paddle-ort python bug

* fix paddle-ort python bug

* Fix test_cudnn_norm_conv and test_cudnn_bn_add_relu in CUDA11.2

* no throw in V100 for some cases