* Add XPU plugin to support the customized ops or improve the performance of the fusion ops based on hand-written xpu micro kernels.

* refine README.md

[XPU] Fix the topk, set_value ops that using temporary tensors avoiding the memory overlaps during multi-stream inference (#54851)

* [CustomDevice] release device manager in py::atexit

* fix hip_version macro

* update

* update

* add lerp bf16 support

* fix bug

* Update test_lerp_op.py

modify the input dtype

* modify the test_lerp_op.py

* Update test_lerp_op.py

* fix bug of import

* add user_defined_grads

* Update test_lerp_op.py

* fix bug of grad

* fix bug of grad

* fix bug of grad

* add the check for bfloat16 dtype

* add linear_compress API

* fix_fetch_op_and_null_type_bug

* fix compile bug

* add test case

* Fix compiling on XPU related to MPTypeTrait.

* Unify the use of MPTypeTrait.

* Fix compiling error.

* style

* more

* update ctest

* Update legacy_backward.yaml

* Update legacy_ops.yaml

* Update legacy_ops.yaml

* update

* update

* update for move

* add int_with_ll to conv

* fix bugs when output_size is specified for conv2d_transpose

* add pos_weight in kernel

* fix unittest

* fix xpu

* fix bce unittest, change infermeta order

* [ROCM] fix cupti, hipcub

* update

* update

* delete swish_raw

* fix

* Update activation_kernel.cc

* fix

* add all_to_all phi operator, kernel, api

* add all_to_all ut

* tinyfix

* remove ops from OpsWithFluidKernelNeedMoveToPhi set

* open static build flag

* OpsWithFluidKernelNeedMoveToPhi

* open new_executor_static_build

* add infermate for cudnn_lstm

* fix

* update

* fix

* update

* update

* update

* fix pow2 decay

* fix pow2 decay

* recover analysis_predictor.cc

* fix pow2 decay

* fix cudnn lstm

* add output register info for svd

* fix pow2_decay_with_linear_warmup_kernel

* recover test lstm cudnn

* recover svg register codes

* fix register info

* fix reduce sum register info

* add output info for adadelta

* add output info for adadelta

* add output info for adamax

* fix complex abs register info

* add register info for cudnn_lstm_grad

* recover

* fix lstm cudnn

* fix

* fix xpu output registe info

* remove std::cout

* add backend

* remove output info in pow2_decay_with_linear_warmup_kernel

* add judgment in TensorShouldBeFakeInitialized

* recover power_

* close new_executor_static_build

* fix set_value_xpu

* add fetch kernel

* support fetch var in new ir

* fix bug

* polish code

* change array equal to np.testing

* static graph autogen code support for matmul op

* fix bug

* fix bug

* fix bug

* fix bug

* fix bug

* fix bug

* fix bug

* add unique kernel for xpu

* add unique kernel for xpu

* update uniittest

* add xpu support for unique with axis

* fix lamb always_adapt

* fix optest

* fix all optests

* fix for log xxx

* add int32/int64 for cpu/gpu; add float16/bfloat16 for cpu forward

* fix docstring

* fix bug

* fix bugs

* fix bugs

* fix bugs

* fix bugs

* fix bug

* using cast

* fix test

* fix api

* fix other bugs

* fix ci bug for not using dygraph guard

* add bfloat16 test

* fix ut

* bf16

* exp/expm1 support int32/int64

* fix ut

* fix ut

* fix ut

* [Zero-Dim] paddle.nanmedian support 0D Tensor

* fix CI

* fix A100 CUDA12 ut

* fix ci uts

* fix test_sync_batch_norm_op

* fix sync bn op ut again by separating 2 files

* fix codestyle ci

* combine other PRs

* fix codestyle

* fix codestyle ci

* fix for log xxx

* add int32/int64 for cpu/gpu; add float16/bfloat16 for cpu forward

* fix docstring

* fix bug

* fix bugs

* fix bugs

* fix bugs

* fix bugs

* fix bug

* using cast

* fix test

* fix api

* fix other bugs

* fix ci bug for not using dygraph guard

* add bfloat16 test

* fix ut

* bf16

* support bias's rank equal to input's rank