* 添加rocm平台支持代码

* 修改一些问题

* 修改一些歧义并添加备注

* 修改代码格式

* 解决冲突后的代码修改

* 修改operators.cmake

* 修改格式

* 修正错误

* 统一接口

* 修改日期

* improve grad perf

* add op test case for complex

* polish code details

* add xpu set constant support

* fix argument rror

* remove useless pyc file

* add transpose double grad test=develop

* fix bug of matmul_v2 for broadcast

* add service, remove ut on mac

* fix heter_profiler & add heter stop method

* fix code style

* fix compile problem when cuda_arch < 6000

* refine code

* refine code

* support roi_align & affine_channel for kunlun

* minor

* Sharding add hybrid-dp feature

* update sharding in distributed_strategy

* update sharding unitest

* revise code format for sharding

* improve drop out

* add VectorizedRandomGeneratorWithGenerator

* fix bug

* modify according to comments

* Add the strategy of skipping cc/cu test compilation and execution in CI, test=develop

* fix if error with CI_SKIP_TEST, test=develop

* fix add properties to test error on Linux/MAC, test=develop

* fix set test properties of test_code_generator error, test=develop

* remove test codes and advance judgment of file modification on Linux, test=develop

* rename CI_SKIP_TEST to CI_SKIP_CPP_TEST, test=document_fix

* Add branch judgement on Linux, test=develop

* Add more value to calculate hash key

* fix size_t

* polish code

* 1.fix matmul bug 2. add one hot

* add xpu error msg

* - Added infrastructre for new test

 - Added UT for Multiple models prediction

- cosmetic fixes

- lint

- lint fixes

* - Removed timeout for MMP test

fix p_norm with empty shape (#29500)

* layernorm fw opt

* layernorm bw opt

* fix typo, test=develop

* remove const dim3 for windows CI compatibility

* merge develop
Co-authored-by: Nzlsh80826 <zlsh80826@gmail.com>

* add ps table

Change-Id: I468a04bd071d21ff52654926fcf4d5f3da19e178

* remove tensor copy in the update_loss_scaling op

* not use thrust.

* fix some cuda memory access error.

* support deepcopy for Layer/Tensor/Paramerbase

* fix some code

As the title

* vec softmax fw

* vec softmax bw

* add a message argument for compiler compatibility

* add tensor_indices in AssignGroupBySize

* add rebuild group in reducer

Add Sleep Time for CUDA Retry, which is similar to our GPU retry logic. This is a try to avoid init GPU allocation random failure in unit test.