* fix some matmul that X and Y both other ops outputs, do not dequantize the Y.

* fix CI format

* fix according to review

* modify xpu.cmake,*test=kunlun (#41832)

* modify xpu.cmake,*test=kunlun

* modify xpu.cmake,*test=kunlun

* modify xpu.cmake,*test=kunlun

* modify xpu.cmake,*test=kunlun

* support bilstm,*test=kunlun

* [cherry-pick]support multi_layer of bilstm,*test=kunlun

* [cherry-pick]refactor sum unit test,*test=kunlun (#43561)

* cherry pick from #43397

* fix code

* place all save/load path into temporary directory

* rm no need unittest

* fix test

* fix test.

* fix pg bugs

* update

* Rename dropout is test (#43098)

* replace dropout_is_test with is_test.
* improve atol on a100.

* fused_attention fused_feedforward api support Model Tensor Parallel (#42985)

* fix is_test bug in fused_feedforward. (#43508)
Co-authored-by: NLi Min <11663212+limin2021@users.noreply.github.com>

    Use tempfile for unit test & custom op test to replace temporary files to ensure that all temporary files will be deleted normally after a single measurement, avoiding the usage of disk files.
    The PR only involves single-test and op test modifications and does not affect existing functionality.
    Release/2.3 branch modified in PR43521;

* fix unit test temp file, test=develop (#43155)

* add cleanup code, test=develop (#43305)

* Fix numpy 1.20+ deprecation warnings (#42929)

* Replace np.bool/np.bool8 with np.bool_

* Replace np.object with np.object_

* Replace np.complex with np.complex128

* Replace np.float with np.float64

* Replace np.int with np.int_

* Rerun pre-commit for newer pre-commit configuration

* Use builtin bool instead of np.bool_ based on the context

* fix mode dtype
Co-authored-by: Nzlsh80826 <rewang@nvidia.com>

* Add progress bar and speed up Quantization Pass

* fix typo

* fix bug of strided_slice (#43388)

* fix stride_slice bug

* fix bug

* fix bug of infer shape for slice (#43443)

* [EinsumOp] Polish forward logic and backward logic for optimize (#42603)

* change logic for optimize

* modifty

* merge

* change einsum_v2 as default and add new flags: FLAG_einsum_opt=1|0 (#43010)

* [EinsumOp] Make EinsumOp support bfloat16. (#43085)

* change einsum_v2 as default and add new flags: FLAG_einsum_opt=1|0

* make EInsumOP support bf16

* add unittest for BF16

* add condition for test_BF16

* fix bugs

* fix

* change the backward api to fit einsum op

    使用 tempfile 替换临时文件，保证在单测结束后，所有临时文件都会被正常的删除，避免占用磁盘文件。
    此 PR 仅涉及单测修改，不影响现有功能。
    develop 分支修改在 PR 43376

删除无用信息

* support fuse conv and bn in QAT (#42255)

* support skip_op_list in PostTrainingQuantization (#42378)

* fix unittest

Reduce amax/amin frobenius_norm_kerne原始实现为Eigen实现，文件编译时间较长，因此本PR将其替换为KP实现
删除DefaultElementwiseOperator中重复功能支持，减少elementwise_double_grad OP编译时间

删除在2.3 对比whl包大小。

解决onnxruntime后端依赖的protobuf跟框架或外部protobuf版本冲突问题

Delete ElementwiseKernel in BroadcastKernel
减少所有Broadcast中重复功能调用，同时减少编译时间和问题体积

删除Broadcast function中rank例化以及Elementwise调用，降低编译时间。
从develop分支中的#42645 PR修改而来，由于develop分支与release分支相差较大，无法实现cherry-pick，因此针对release2.3重新提交PR.
Broadcast中关于rank的例化会导致底层模板展开较多，造成reduce_sum_grad_kernel.cu.o文件体积过大，修改后可以降低.o体积及编译时间

删除判断build目录大小和预测库大小检查功能。该功能是和develop比较，会存在差异，在release任务中取消判断

* Fix cond_block_grad error when handle no need grad vras

* Add comment and UT

* fix_multi_int8 (#42977)

* cherry-pick fix_multihead_int8