* [CustomPlace] fix amp

* [CustomPlace] fix amp

* fix ut because of too long time matmul fp16

* fused qkvBiasAdd and transpose with split qkv

* fix typo

* fix format

* fix name

* add annotation

* fix comment

* cleanup unused code

* unify is_int8 is_bfloat16

* Simplify matmul_v2 FWD kernel

* remove RunKernel methods

* remove import namespace

* remove headers

* clean fluid/phi cross imports

* remove fluid axpy_handler

* delete fluid methods

* activations

* OneDNNMemDesc

* MKLDNNFormatForSize

* MatchShapeToLayout

* MKLDNNMemoryFormat

* MKLDNNFormat

* ReorderMKLDNNHandler

* to_void_cast

* review suggestions

* interpolate

* remove fluid depedency

* init

* ExecuteMatMulV2

* rm fluid kernel

* matmul_grad

* remove mutable_data

* Migrate conv_transpose to phi

* Move handler to kernel

* kernel m

* Fix formatting

* handler

* remove fluid

* revert tcp_store

* tcp_store

* remove unused

* Fix declaration

* add dnn input

* Fix typo
Co-authored-by: NSławomir Siwek <slawomir.siwek@intel.com>

* fix bug of zero_allocator in host

* fix test compile bug

* add unittest

* update test

* Add quick gelu and fused bias add kernel

* fix annotation

* remove useless code

* add fast gelu option and set it in multi transformer op

* add flag to restrict if use fast gelu approximate

* fix flags conflict

* fix use tanh function instead

* add cudart version limit

* use phi fast tanh func

* fix comment

* move "paddle/phi/backends/gpu/gpu_device_function.h" to phi

* update copyright years

* rm "fluid/platform/device/gpu/gpu_device_function.h" in phi

* fix rocm-complie bugs

* correct sync behavior for XPU distributed training

XPU support event mechanism similar to cuda event, so it is advisable to
use an event to sync compute/comm streams for performance. However this
mechanism is never fully tested, and inconsistent loss/ending_epochs are
reported. Therefore, this PR replaces event sync with stream waiting as
a temporary solution.

* remove compile warning

* fix device id issue for xpu eager

xpu device id is not correctly set in eager mode, thus vars are on dev0 unless
XPUDeviceGurad is called, leading to this error message for all node rank != 0:
"NotImplementedError: (Unimplemented) Place Place(xpu:0) is not supported."

* fix typo

* fix pybind error

* Refactor conv_kernel and conv_grad_kernel to provide interface for CUDNNv8 implementation

* Fix macro

* Add implementation for conv_kernel and conv_grad_kernel

* Modification after rebase onto latest develop

* Modify plan cache to comply with the API of phi::autotune

* Refactor to reduce duplicate code

* Review fix:
- move functions in  conv_kernel_impl_v8.h and conv_grad_kernel_impl_v8.h to conv_kernel.cu and conv_grad_kernelk.cu
- add const specifier for input tensor
- add logging when plans fail to execute
- move CudnnConvBwdFilterV8 and CudnnConvBwdDataV8 to conv_cudnn_frontend.h

* - move plan building outside of cache

* Fix ROCM build

* remove "gpu_primitives.h" in fluid namespace

* fix PR-CI-GpuPS fail

* fix PR-CI-GpuPS fail

* clip extra and intermediate output of op

* fix bug

* fix bug

* polich code

* polich log

* [NPU] add _npu_identity op and api, test=develop

* fix doc

* address comments

* int scale

* round

* revert commit

* standard api

* fix xpu bugs

* add_cast_bool

* cast

* Implement a common dims simplifier.

* Fix the include position error.

* Reduce the cpu overhead of broadcast computing.

* rm "paddle/fluid/operators/math.h" in phi

* rm "paddle/fluid/operators/math.h" in fluit

Support bfloat16 for adamw and adam optimizer. Fit the lr for pure bf16 training with tensor fusion. (#48041)

* add bfloat16 for adamw

* set lr not to bfloat16 for pure bf16 training

* update the logic

* update the adamw optimizer

* support bfloat for adam

* add vectorized bfloat16 atomicAdd

* fix compile error

* fix compile error again

* fix V100 compile error

* fix V100 compile again

* [NPU] update npu prop, test=develop

* remove ddim.h

* remove diff

* update storage prop, test=develop