* add log for memory stats

* fix string_split in einsum

* Try to increase the repeat of autotune and fix the setting of allow_tf32_cublas.

* Change the repeat of cublaslt to 10.

* Use FLAGS_cublaslt_exhaustive_search_times as repeats.

* Fix compiling error on CI.

* Polish the key and simplify codes.

* move raw kernels to legacy

* Update elementwise_add_kernel.cu

* fix

Fixed the bug in the api.cc file where there was an inconsistency between the specified type (std::vector<DenseTensor*>&) in the function pointer kernel_signature and the type of the phi kernel parameter (std::vector<DenseTensor*>) when the phi kernel is set to output as std::vector<DenseTensor*>. test=kunlun (#54053)

Removed unnecessary header files introduced

[XPU][PHI Kernels] bind bitwise_add kernel & add int32/int64 support to scatter_nd_add kernel for xpu (#54066)

* bind new kernels to xpu

* refine code

* fix bugs in unittest

* [AMP OP&Test] Support float16 in selu

* fix

* merge code from forsish

* polish

* paddle/fluid/pybind/auto_parallel_py.cc

* polish

* polish

* polish

* polish

* polish

* polish

* polish

* polish

* polish

* polish

* polish

* [phi] autogen code tril_triu

* [phi][api]fix tril_triu_grad args

* [fluid] clean cmake; [phi] fix infer_meta

* Fix typos

* Fix

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update

* update HostAlloc

* update param name

* update cpu kernel

* remove kernel header

* update

* update

* update

* update

* update

* set out dtype

* static graph autogen code support for pad3d op

* bug fixed

* add ut for pad3d mkldnn op

* fix coverage

* fix bug

* fix bug

* Delete test_pad3d_mkldnn_op.py

* fix the static op generation for group_norm

* fix bug of mismatch

* fix bug of AssertionError

* fix setting of composite

* update_c++14_to_c++17_on_windows

* disable test_audio_logmel_feature and test_audio_mel_feature

* fix empty bugs for xpu

* fix empty bugs for xpu

* fix transfer_layout when input size if too big
* do not add TransferLayoutKernelGPU
* add int64 and add check

* Add GPU kernel for multiclass_nms3 op

* Make multiclass_nms3 gpu kernel output consistent with cpu kernel

* Fix API incompatibility

* Fix unittests on builds without CUDA

* Fix ROCM build

* Remove fluid headers; Use default atol for unittest

* Change function and variable naming

* Add comments; Reduce redundant code

* Use paddle test framework

* bind xpu op: 3D grid sample

* fix edge cases in xpu op: reshape & slice

* add minimum grad composite rules

* add public python api

* fix format

* fix format

* update testcase

* fix testcase

* fix format

* fix cmakelist.txt

* fix format

* fix param problem

* fix op and composite rule

* fix bf16 cpu support problem

* fix bf16 cpu issue

* fix axis error log

* add axis for maximum

* revert commit

* remove .orig

* fix generic problem

* revert max op

* fix axis error

* fix maximum axis

* fix test_check_output

* fix cinn

* fix minimum maximum axis check

* Reorganize the forward codes of flash-attention.

* Fix forward.

* Remove some noused codes.

* Simplify codes and fix backward.

* Change all LOG(INFO) to VLOG and fix the backward.

* add scale for AF2 flash_attn, much thanks to xreki and shaojie for debug these codes

* decrease the effect of debug print on performance

* Unify the initialize of flashattn arguments.

* Rewirte the reshape of temp_mask and temp_bias.

* API support use_flash_attn.

* Fix compiling error on CI.

* Try to crop the flash-attention lib.

* Correct the condition of whether can use flash-attn.

* Remove the softmax_out argument.

* Remove is_causal.

* Polish codes.

* Fix qkv_transpose_out's shape and scaling of Q * K.

* Update commit of flash-attention.

---------
Co-authored-by: NLiu Yiqun <liuyiqun01@baidu.com>