* add legacy_api.yaml

* set merged_momentum inplace only

* support inplace optional<vector<tensor>>

* add dygraph_mode api

* optimize TensorToConstDenseTensorPtr

* 修复生成动态图代码时，如果输出没有配置名字，会导致下标越界的问题。

* decide forward_return[0] is not none

* 修改反向yaml前向输出只有一个时，未配置名字，那么输出自动生成为out

* modify code style

* fix_stop_gradient

* [Eager draft] forward_only interface migrate to autograd_api

* strings api add dygraph forward function

* rm useless comments

* draft version for check CI

* fix ci

* forward-only no need compute_require_grad and pass stop_gradient, rm useless comments

* polish yaml and using CPUPlace = phi::CPUPlace

* rm useless comments

* polish yaml and update some test case

* rm useless funcs

* polish eager_gen code

* polish code

* support optional<vector<Tensor>> in yaml and eager

* delete useless comments in eager_gen.py

* fix api_base.py support optional<vector<TTensor>>

* python_c_gen.py support optional<vector<tensor>>

* transfer linear_interp_v2 yaml from fluid to phi

* fix op_test typo error

* change linear_interp_v2 testcase

* fix args in final_state_linear_interp_v2

* fix zeropad2d typo. test=document_fix

* Polish black_ops_list logic in eager_gen

* update black_ops_list

1. change the codegen code to avoid conversion from heterogeneous 'initializer list' to tuple, which fails on gcc 5.4; (#45036)

2. add a template CheckTensorHasNanOrInf to handle arbitary tuple of supported types.

* [Eager] use final_state_fill_constant_

* fill_constant use str_value

* add fill_constant_ to no_amp_list

* use float(value) as input

* support final state full_ same as fill_constant

* clean tensor.h

* fix gather_nd

* [Eager] just need to run needed nodes

* [Eager] just need to run needed nodes

* manuel_fused_gemm_epilogue

* remove cudaDeviceContext

* remove more template

* fix rocm compile

* remove alias name CUDADeviceContext

* fix compile

* fix tests

* revert changes

* [Phi] Add yaml for assign_value

* [Phi] Fix the bug of the assign api and modify the unittest

* [Phi] Fix the bug when the tensor does not have the backend info

* [Phi] Replace the functional-style cast init by the brace-init

* [Phi] Cast the data explicitly

* refactor general_grad and fix some bugs

* add TODO: support prune logic deeper

* move average_accumulates op to phi kernel

* fix new dygraph record event for op

* update unit test

* fix bug of amp code_gen

* fix bug

* [Dy2Sta]Fix Segment Fault while training multi-card if params have no grad

* fix unittest

* release gil when run backward

* Rename BOOST_GET macros

* Fix conflicts

* eager variable back sync

* sparse support amp

* EagerAmpAutoCasts support sparse

* Compilation optimization

* [CustomDevice] add custom ccl api

* add ut

* add amp logic in python_C

* fix inplace bug

* split coed gen for eager fluid_generated

* fused_gate_attention manual code in eager

* fix deriv with inplace

* fix double grad bugs

* remove additional file

* fix compat dygraph mode

* fix yaml remove additional yaml

* fix slice double grad error and auto code gen logic error for higher order differentiate

* fix fix_higher_order_deriv

* remove additional include

* fix fix_higher_order_deriv