* add multi tensor for adam

* add merged_adam op

* refine code

* refine adam compute logic

* support multi precision update for LAMB

* hide some api

* fix ci uts

* fix lamb output of dygraph

* remove some changes to some PR

* try to fix Py3 CI compile error

* fix test_imperative_optimizer, add lars ut, add layer_norm ut

* fix ut, fix format

* fix ut

* fix windows ci

* refine structure for cuda and rocm

* update

* update

* update

* update

* add regularation and Nesterov for mergerd_momentum

* refine unittest for use_nesterov attr

* refine op check

* refine code

* fix bug

* refine code of regularization_flag

* delete useless code

* [NPU] reorganization for device API abstraction

* [NPU] delete old files

* [NPU] fix npu_collective_helper

* [NPU] fix collective_helper

* [NPU] fix ut

* [NPU] mod memory allocation and hccl_helper

* [NPU] fix place_type

* [NPU] split enfoce.h

* move acl* call into npu_info

* merge conflict

* fix merge

* merge conflict

* merge conflict

* copy beta pow to same place when skip_update=1

* fix xpu

* add pow2_warmup op

* remove contrib __all__

* add AttrT

* rename

* follow comments

* fix duplicate PADDLE_RESTRICT

* remove wrong restrict

* remove master_param_out __restrict__

* update

* merge momentum ops

* update

* add ut to improve coverage

* remove optimizer change

* fix error msg

* update ut

* add __restrict__ for CUDA

* update ut

* move merged_momentum_op to optimizer dir

* fix coverage

* A leap of try for cudaLaunchCooperativeKernel

* fix bugs

* Totally replace the lar cuda kernel

* Fix bugs

* a test for lars merge

* Adding las_op_momentum infer_shape

* Fix codes

* use avg_numel instead of max_numel to acquire grid num

* modify unittest files about lars op

* Finally converge when merged-lars works

* fix ctest files

* add merged_operation kernel when cuda version is older than 11

* Fix code style

* fix ctest failure

* fix error

* fix all ctest error and change lars compute code of cpu

* fix bugs on v100.

* revert python modififation about lars

* revert python modification codes

* A leap of try for cudaLaunchCooperativeKernel

* fix bugs

* Totally replace the lar cuda kernel

* Fix bugs

* fix code according to comments

* fix codes according to  review comments

* adding some function overload

* relocate the power operation.

* Create stateful OneDNNAXPYHandler object.

This makes it possible to call it multiple times without recreating the
oneDNN primitives every time.

* Prepare SGDOpKernel to reuse its implementation from OneDNN kernel.

* OneDNN SGD kernel.

* Update call to use new OneDNNAXPYHandler object api.

* Setup seed in proper place.

* Enable OneDNN kernel only for single case.

* For dense param and sparse grad.

* Small refactor.

* Enable oneDNN by op attr or by cmd line flag.

* Use int64_t type for number of elements.

* Support dense param and grad from OneDNN kernel.

* Enable SGD OneDNN kernel when use MP BF16 optimizer.

* Force non-copyable/movable OneDNNAXPYHandler.

* Reuse OneDNNAXPYHandler for spare tensors in SUM op.

* Fix SFINAE rules.

* Remove recording event inside AXPY.

* Get rid of internal primitive caching.

* Stop use PP cache mechanims to store mem and primitive obj.
* Handler obj store and reuse needed desc & prim

* Do not derive from MKLDNNHandlerT

* add layerwise learning rate for adamw

* fix format

* add unitest

* add NotImplementedError

* add gpu unitest

* update gpuplace

* sparse_momentum_op is used to save w@GRAD memory for gather_op when gather from a large parameter

* adamw support cuda

* adamw support cuda

* [npu]add rmsprop op

* add momentum_op_npu and test

* update

* fix hang

* copy found_inf to cpu in advance to improve performance

* add npu test

* add npu test

* refine code

* refine memcpy op

* fix adam

* adam add input SkipUpdate

* add unittest

* add npu unittest

* fix xpu compile

* remove param stream

* Add oneDNN AXPY handler.

* Add fallback for small tensors.

* Fix ifdefs

* Remove unnecessary namespace prefixes and add missing headers.

* Guard handler_axpy with proper ifdefs.

* Compilation of this function is possible only when Paddle is not build
with CUDA nor HIP.

* Move AXPY handler code to separate files.

* Use oneDNN AXPY handler in SGD op.

* Use axpy handler only when Paddle is built with oneDNN.

* Add test for SUM BF16 with big rows.

* Fix SFINAE rules for elementwise_add_to.

* Add test case for SGD with big rows.

* update

* update
Co-authored-by: NAdam Osewski <adam.osewski@intel.com>

* refine ~npuOpRunner

* implement destructor and forbid copy

* use reference to avoid copy

* use const reference

* relax adam precision

* fix top_k

* add use_global_beta_pow

* add use_global_beta_pow

* update npu kernel

* update python api

* refine code

* add ut for use_global_beta_pow

* fix npu kernel

* add ut for api

* add ut for exception

* add ut for save/load

* add input EpsilonTensor for adam

* update python api

* add unit test

* add npu test

* add more ut

* [NPU] support GarbageCollector for npu (#31874)

* support GarbageCollector for npu

* fix typo

* fix gather_grad

* disable NPUDefaultStreamGarbageCollector on NPU

* [NPU] support npu for memcpy op (#31808)

* support npu for memcpy op

* add ut

* fix ut

* fix typo

* 【NPU】fix bug of using temp vector (#31963)

* fix bug when beta1_pow on cpu (#31995)

* [NPU] support npu profiler (#31684)

* support npu profiler

* add python api

* fix bugs

* add wrapper for incomplete type

* update profile proto

* record npu wait

* add xpu placeholder

* fix adam (#32016)

* [NPU] enable async copy and  add wait before sync operation (#31956)

* enable async copy and  add wait before sync operation

* remove unneccessary wait

* add FillNpuTensorWithConstant

* refine

* fix fill_constant

* make TensorFromVector/TensorToVector sync

* [NPU] Support dataloader on npu place. (#31867)

* [NPU] Wait on NPUPlace (#32086)

* [NPU] fix cast op (#32121)

* fix npu kernel of cast op to handle casting to same dtype

* add comments

* [NPU] support cann 20.3 (#32044)

* fix compile problem on cann 20.3

* fix ut

* fix test_mul

* fix check_finite_and_scale

* fix lookup_table_v2_grad

* fix cmake

* support print op

* [NPU] Support npu save load (#31893)

* support save load for NPU

* add save load npu unittest

* support np.array transform in NPU

* fix errors

* delete dygraph in unittest

* add Wait

* fix unittest

* fix review comment

* fix unittest problem

* fix little problem

* change aclrtSynchronizeDevice to aclrtSynchronizeStream for better performance (#32196)

* change aclrtSynchronizeDevice to aclrtSynchronizeStream for better performace

* refine code

* fix NPUDeviceContext in all c++ unittest (#32198)

* fix NPUDeviceContext in all c++ unittest

* refine log
Co-authored-by: Npangyoki <pangyoki@126.com>

* [NPU] Remove TensorFromVector and avoid sync copy in npu op kernel for better performance (#31994)

* enable async copy and  add wait before sync operation

* remove unneccessary wait

* add FillNpuTensorWithConstant

* refine

* fix fill_constant

* change TensorFromVector to FillNpuTensorWithConstant

* fix ignored api

* delete extra unittest

* fix little error

* fix update_loss_scaling_op_npu and check_finite_and_unscale_op_npu

* change TensorCopySync to TensorCopy

* delete useless Wait and add StreamWait

* fix npu_stream error

* fix check_finite_and_unscale_op_npu TensorCopy

* only save stream wait

* fix NPUDeviceContext in all c++ unittest

* delete wait
Co-authored-by: Nzhiqiu <chenqiuliang@baidu.com>

* delete useless unittest file (#32206)

* Fix op test (#32231)

* fix conditional block (#32243)

* fix adam bug again (#32246)

* fix compile

* fix ut

* fix ut
Co-authored-by: Nliym27 <33742067+liym27@users.noreply.github.com>
Co-authored-by: Npangyoki <pangyoki@126.com>