Describe

Add feed_forward for fused attention op.
(1) Encapsulate matmul impl (forward and backward) used in attention op.
(2) Implement bias_add (forward and backward) used in attention op.

* add api

* temp save

* revert

* copytocpu async ok

* fix style

* copy sync ok

* fix compile error

* fix compile error

* api done

* update python async api

* fix compile

* remove async python api; add c++ async unittest

* remove python async api

* update unittest

* update unittest

* add C++ unittest for copytensor

* add unittest

* update namespace utils to class TensorUtils

* add unittest

* update unittest

* update unittest

* update code style

* update code style

* update unittest

* add_roi_align_npu

* update

* update

* update

* use spinlock in auto growth, test=develop

* refine,test=develop

* fix cmaklist for new executor, test=develop

* refine, test=develop

* refine, test=develop

* Modify ci time count & fix resnet50_quant multi_thread tests

* fix wrong time variable

* temporary change

* fix for expand_v2

* changes after review, activated ppyolov inference test

* fix index tensor leak in __setitem__

* fix another usage of PyTuple_Pack

* refine code

* refine code

* handle None index

* add Py_DecRef

* revert ut

* refine code

* merge develop

* use RAII

* follow comments

* add fetch, test=develop

* fix fetch2op, test=develop

* fix fetch2op, test=develop

* refine, test=develop

* fix fetch ctx, test=develop

* add wait, test=develop

* rename fetch2 to fetch_v2, test=develop

* merge, test=develop

* Add no_sync in data parallel for dynamic graph

* modify UT of no_sync

* delete test_parallel_dygraph_dataparallel_no_sync.py

* add test_parallel_dygraph_no_sync.py

* modify run_trainer_with_spawn in UTs

* Add UT of complex control flow in no_sync

* add specific descriptions and notes for no_sync

* check code style

* modify UT's TIMEOUT in CMakeLists.txt

* fix bmm bug

* bmm style

* fix bmm

* - concat refactoring draft

* - cmpilation fixes

* - yet another compilation fix

* - fix

* - compilation fix

* - fixes to compilation

* - another compilation fix

* - fix

* - Added overloaded AcquirePrimitiveDesc for concat

* - fix

* - reserve introduced

* - UT fixes

* - test concat int8 improved

* - fixes

* - fix to crash

* - lint fixes

* - fixes after review

* - some other fixes from review

* add conv_op_npu and test

* add more tests

* clean headers & support fp16

* update

* add pool2d_op_npu and test

* update

* update pool2d_backward_navie

* clean headers

* add auto_parallel dir

* mv to paddle.distributed

* add shard_xx api

* add distributed attrs for var

* add ut, test=develop

* add dist

* update

* update

* update

* update

* update

* update, test=develop

* update, test=develop

* update, test=develop

* update, test=develop

* update, test=develop

* update, test=develop

* update, test=develop

* update

* update

* update

* update

* update

* update, test=develop

* update, test=develop

* update

* update

* delete unused proto

* resotre op_desc

* restore type_defs

* update var_desc

* remove dimss_mapping for proto_pybind

* update interface.py

* update framework.py

* update

* update

* add auto_parallel dir

* mv to paddle.distributed

* add shard_xx api

* add distributed attrs for var

* add ut, test=develop

* [WIP] Add the auto completion feature and related codes

* [WIP] Improve the auto completion and related codes

* [WIP] Make the auto completion to support data-parallel

* [WIP] Make the completion support mp and dp+mp

* [WIP] Refactor auto completion unit test for MLP

* [WIP] Refactor the implementation of DistributedOperatorImpl

* [WIP] Improve dims_mapping update rule and fix a bug

* [WIP] Support auto completion for one transformer decoder layer

* [WIP] Add a minor change

* [WIP] Fix a bug within the uint test

* Shard XShape tensor, add embedding completion and refactor code

* Add the distributed_operators dir to setup.py.in

* Improve the completion process and add the unittest for gpt

* fix process_mesh ut

* fix process_mesh ut

* update

* update, test=develop

* Add support for automatically completing distributed attrs of special ops

* update

* update

* update

* fix doc sample codes, test=develop

* improve coverage, test=develop

* add static_mode check, test=develop

* Model the cluster for cost model and physical mapping

* update, test=develop

* add set_placement, test=develop

* Add the check to make sure the candidate tensors' size is great than zero

* update doc, test=develop

* update doc, test=develop

* update doc, test=develop

* update doc, test=develop

* update, test=develop

* Auto mark dist attrs annotated by user

* update ndarray to nested list, test=develop

* update, test=develop

* Add auto-completion module for auto-parallel (based on PR#33804)

* Remove unnecessary files

* Remove unrelated files for the auto completion pr

* Update the unit test to improve the coverage

* Modify codes based on reviews

* Minor changes for CI

* Improve some codes based on new comments

* Fix bugs caused by shallow copy in attributes.py
* Imporve amend_distributed_attr_for_program in context.py
* Other changes for weihang's comments
Co-authored-by: Nsandyhouse <lilong12@baidu.com>

* - disabled interpolate onednn

* - compilation fix

* - draft of batch_norm cache disabling

* - fixes to UT

* enable infer_ut on windows

* remove lib calculation & time

* unset http_proxy when download bos file on windows

Refactor the organization of layer_norm cuda impl so that it can be reused in fused attention op.

    Extract the layer_norm cuda impl form layer_norm_op.cu to layer_norm_kernel.cu.h.
    Define fused/attention_layer_norm.h, which can be used in fused attention op in next PR.

* Support getitem by Bool index

* delete some debug info of bool index

* support the case that the shape of bool index is different from indexed tensor