* support pure bfloat16

* support bf16 linear

* update PR to pass CI

* tiny fix where_grad_kernel.cu

* add bfloat16 to selu_grad to pass CI

* fix selu grad compilation error

* 之前的pr合并了大量错误代码，重新提交一份

* 之前的pr合并了大量错误代码，重新提交一份

* 修正格式问题

* 改回原来的格式

* 按照要求修改

* 按照要求修改格式

* 修复注释的问题

* 更新格式

* 测试自动格式化

* 修正英文注释

* fix docs build error

* pre-commit

* for docs build

* for docs build

* 修复mantissa计算错误的bug

* 修复误判exponent可能存在负数，导致计算量增加的情况
Co-authored-by: NLigoml <39876205+Ligoml@users.noreply.github.com>

Add index_select, index_select_grad, reduce_min kernel and their unittests for kunlun. Add registers of index_select, index_select_grad, reduce_min, sqrt, sqrt_grad to xpu2_op_list.test=kunlun. (#46557)

* [CodeStyle][F401] remove unused import in unittests/collective

* empty commit, test=document_fix

* empty commit

* remove calibration file path

* remove useless code

* [CustomDevice] add to_static, amp ut

* update

* fix failed ut

* update

[Eager, Performance optimization] support mod / matmul ( % and @ operator) to sink to Cpp layer (#46565)

* [Eager, Performance optimization] support mod ( % operator) to sink to Cpp layer

* fix mod logic

* support matmul math operator

* rm LOG(warning), use VLOG(6)

* fix conflicts mistake

* [AutoParallel] fix dist_split

* add unittest

* update cmakelist

* remove needless using tensor

* remove needless using tensor

* resolve conflict

* replace tensor using

* fix format error

* revert needless changing

* fix rocm and npu compile error

* fix cinn compile error

* fix format error

* fix mkldnn format error

* fix mkldnn format error

* fix cinn compile error

* fix cinn compile error

* fix cinn compile error

* resolve conflict

* fix p_norm n=1 issue

* fix p norm test error

[Eager, Performance optimization] support less_than & less_equal( < & <= operator) to sink to Cpp layer (#46542)

* added fc int8 tests

* CI fix

* added skipping UTs for GPUs

* fixes for CI

* added support for residual connections inside fc

* fix for quant int8 bias

* - lint
Co-authored-by: Njakpiase <jakpia21@gmail.com>

* add gpu kernel for transfer layout

* comment error throw

* fix: flag setting in testcase; add condition check for raising error

* fix typo

* fix: add error type for PADDLE_THROW

* remove kernel fallback in data_transfer.cc

* remove useless variable definition

* [CodeStyle] reformat test_error

* update lineno

* remove test_error from yapf whitelist

* try format test_origin_info

* try to fix origin info test case

* remove future import

* remove test_origin_info from yapf excludes

* empty commit, test=document_fix

* empty commit

* fix clip_extra code in remove_training_info

* revert rnn opmaker clear

* [Eager] math op sink to Cpp level

* fix ci errors

* draft version

* support + and - operator under cpp directly

* add static test

* polish code

* promote types or unify right type to left

* recover static test case

* polish code and fix some ci errors

* support complex and polish code

* fix conflicts

* fix windows ci errors

* fix windows-inference-ci errors

* polish and fix tests

* fix test case

* polish code

* [Eager, Performance optimization] support multiply( * operator) to sink to Cpp layer

* rm useless glog

* [Eager, Performance optimization] support divide( / and // operator) to sink to Cpp layer

* polish code

* polish code and fix code-format

* polish code

* fix CI

* polish code

* update test

* support div operator under cpp

* fix scalar as input

* Polish div logic, fix ci test

* fix errors

* init dropout

* small format fix

* fix pr comments

* add value test

* [Auto Parallel] Imporve the user-defined fetches and logging

* [Auto Parallel] Make Engine class callable

* [Auto Parallel] Update the data loading of tuner

* [CodeStyle] remove all future import

* revert test_error.py

* restore future import in example code

* fix libpaddle soname mismatch error

* fix windows failed

* polish linux and windows make impl

* unify winddows lib name

* fix windows error

* revert copy dst change

* revert naming change

* revert windows change

* fix gpups compile failed

* [Eager] math op sink to Cpp level

* fix ci errors

* draft version


* support + and - operator under cpp directly

* add static test

* polish code

* promote types or unify right type to left

* recover static test case

* polish code and fix some ci errors

* support complex and polish code

* fix conflicts

* fix windows ci errors

* fix windows-inference-ci errors

* polish and fix tests

* fix test case

* polish code

* [Eager, Performance optimization] support multiply( * operator) to sink to Cpp layer

* rm useless glog

* polish code

* polish code and fix code-format

* polish code

* fix CI

* polish code

* Support both use_calc_stream and sync_op in send recv APIs (#46023)

* add batch_norm prim2orig rule
Co-authored-by: NWen Sun <35923278+HermitSun@users.noreply.github.com>

[Eager, Performance optimization] math op sink to Cpp level ( + and - operator as an example ) (#45811)

* [Eager] math op sink to Cpp level

* fix ci errors

* draft version

* draft version

* draft version

* support + and - operator under cpp directly

* add static test

* polish code

* promote types or unify right type to left

* recover static test case

* polish code and fix some ci errors

* support complex and polish code

* fix conflicts

* fix windows ci errors

* fix windows-inference-ci errors

* polish and fix tests

* fix test case

* polish code

* polish code

* polish code and fix code-format

* polish code

* polish code

* polish code

* polish code

* enable eager mode on xpu, test=kunlun

* add numpy support to xpu

* fix tensor using error

* fix  error, test=kunlun

* fix failed tests, test=kunlun

* support rsqrt_p

* refine code and ut

* add_prim_rsqrt

* fix ut

* open the clip_extra flag in paddle.static.save_inference_model, test=allcase

* set the defalut value of clip_extra in TracedLayer from False to True, test=allcase

*  Disable TF32 to solve accuracy for test_trt_conv_pass and test_trt_deformable_conv in A10 GPU.