* implement of matmul using cublasLt instead of cublas

* Update matmul_kernel_impl_via_blasLt.h

---------
Co-authored-by: Nzhangbopd <1299246947@qq.com>
Co-authored-by: NBo Zhang <105368690+zhangbopd@users.noreply.github.com>
Co-authored-by: NLiu Yiqun <liuyiqun01@baidu.com>

* Enable matmul + bias fusion in fused_gat_attention.

* Add a variable to control whether using fused matmul + bias.

* change outputs and grads from fp16-fp16-comparision and fp16-fp32
comparision

* support grad comparision fp16-fp32

* the change of reference dtype only occured from np.float16 to np.float32

* fix the list type can not infer the dtype by attribute dtype by transfer
the list to array

* adjust the default atol and rtol of float16 to 1e-3

* Polish code

* fix error

* fix

* Polish code

* fix the _is_cal_ref and np.float16

* fix the combination of is_calc_ref and np.float16

* remove unuseful codes in op_test.py

* fix ci

* fix the rtol set in the dygraph checker and eager checker

---------
Co-authored-by: NZzSean <18818272991@163.com>

* rename elementwise_heaviside to heaviside

* delete __init__.py

* fix bug

* Revert "fixoptminizer _set_auxiliary_var bug (#50335)"

This reverts commit c44005f0.

* Revert "refine optimizer create accumulators (#50188)"

This reverts commit 244e7546.

* Revert "fix found_inf bug for custom optimizer (#50158)"

This reverts commit 64573f9f.

* Revert "refine amp scaler found_inf (#49864)"

This reverts commit 382e9a06.

* fix code format

* fix conflict

* fix

* and others

* more ops

* reset distribute_fpn and precision_recall

* reset fc

* modify arange test

* modify reshape&reduce

* add fill_any and sigmoid_cross_entropy

* reset linear_interp_v2

* reset reduce

* modify

* modify arange

* modify cast

* allow fall back to fp16 when int8

* refine code

* refine code

* refine code

* ConvertToFusedOp

* change static to inline
Co-authored-by: NTomasz Socha <tomasz.socha@intel.com>

---------
Co-authored-by: NTomasz Socha <tomasz.socha@intel.com>

* change amp with to_prim

* fix prim amp

* fix rules

* fix liear

* add amp test

* add test

* disable this test on cpu

* disable this test on cpu

---------
Co-authored-by: Ncyber-pioneer <chenzhuo@tju.edu.cn>

* support 'backend' in static ops

* change bitwise_xx comment in python

* change bitwise_xxx comment in python

* change 'backend' and 'data_type' in GetExpectedKernelType

* Add bert prim and cinn test

* support prim test in OpTest

* fix cmake

* fix op test

* fix test_input_spec

* disable cinn in reduce_sum unit test

* add bfloat16 dtype for sum

* add approve rules

* polish code

* add clear jit program function

* convert grad out from tensor to numpy

* remove unnecessary code

* add only_prim flag

* fix flag

* fix op test

* add attr

* fix optest comp inplace error

* fix op test

* fix op test with guard

* add initialization of check_comp flag

* fix comp inplace error in op test

* rename check_comp with check_prim and add bfloat16 dtype convert

* rename comp_op_type to prim_op_type

* rename comp to prim

* remove useless code

* skip ci check for only prim

* add no_grad_vars and grad_outputs in prim test

* fix var_dict

* fix op test for only_prim

* fix dy2static bugs

* polish some code

* temp

* modify op test

* except cinn test

* modify bfp16

* modify pad grad

* add pad_grad dtype

* start cinn part

---------
Co-authored-by: NCharles-hit <wanghao107@baidu.com>

* fix attrs loss in creating op

* add comment

* add case

* add case

* remove unused case setting

* perfect translated layer

* perfect code according comment

* change protobuf version in pslib mode and link libjvm.so fot libps.so

* keep protobuf version same with pslib and enable compile with pslib

* [XPU] add expand_grad, isnan, meshgrid kernels

* update

* fix multihead

* fix multihead

* [CINN]Enhance CacheKey hash logic by considering input dtypes

* add unittest

* fix typo

* fix typo

* fix map.at

* fix find

* fix test

* fix cinn cache key structure realize

* using ordered map for attributes

* add test by review advice

---------
Co-authored-by: Njiangcheng <thisjiang@qq.com>

* got rid of save_quant_model

* review changes

* change phi tensor_gen->tensor_operants_gen

* [Tensor API] Support multiple Tensor C++ api

* [Tensor API] Unsupport prob Tensor API

* accept reviewers comment of #50731

* delete tensor_api.yaml

* move fluid generator to phi

* move fluid generator to phi

* update .gitignore

* fix bugs

* fix cannot find "glog/logging.h" in "generator.h"

* fix bugs

* first commit

* main codes has been developed

* fix all bugs

* add vectorize input&output

* a test for optimization_of_layer_norm_fwd

* add some changes

* fix memory coalesced access for more optimization.

* fix addition ctest error

* fix according to ci-approval

* remove change on slice

* fix accurary diff issue when XPU op batch_norm is added to XPU blacklist

* remap op output tensor to input tensor when the op has fallen back to CPU

* rename function name and fix bug causing by InplaceCounter