* clean unused code: piece.cc/h

* clean usage

* Rewrite strided_slice converter  using  shape tensor 
* clean code

* add embedding range check

* change head file

* change head file

* fix

* first version, accuracy corrected

* disable debug print

* use blockReduceSum in phi

* add UT

* add opCompat

* code style

* code refine

* bug fix

* code refine

* test fix

* bugfix

* codesytle fix

* code style

* code-style

* code-style

* code-style

* git

* style

* leave default relu in kernel

* style

* cleanup FCMKLDNN pattern

* merge conflicts

* update develop

* update develop

* add const

* rename to oneDNN and adjust attributes

* whitespace

* cond infer apply exec seprate

* fix bugs

* [XPU] update xpu cmake to 1016. test=kunlun

* fix special case of transpose op. test=kunlun

* support generating code of opmaker for backward op invoke forward op

* gsupport code-gen of opmaker for sparse op

* refind logic of choose phi kernrel

* fix complie budg

* fix code_gen bug

* fix bug

* fix kernel signature code-gen

* fix complie bug of VarType

* fix complie bug of VarType

* fix test_sparse_conv_op

* fix test_sparse_norm_op

* Fix bug of reduce_sum op. When input.numel() > INT32_MAX, its result
is wrong.

* Support allow_partial switch, which can be configure in
pipeline_configs. If sent tensor are not the same from
different hosts, they shouldn't been sent partially and
then concated as a whole tensor.

* Change name allow_partial to enable_partial_send_recv.

* Add global variable _enable_partial_send_recv

* Fix bug of reduce_sum op. When input.numel() > INT32_MAX, its result
is wrong.

* support pure bfloat16

* support bf16 linear

* update PR to pass CI

* tiny fix where_grad_kernel.cu

* Support bfloat16 type for reducer and sharding.

* Fix some bug.

* Polish code.

* Polise code.

* Add bfloat16 datatype in fill_grad kernels.
Co-authored-by: Nsneaxiy <sneaxiy@126.com>

This reverts commit 166ff39a.

* namespace modify

* update by comment

* add sparse reshape

* change the dtype in all test cases to int64

* just one test case

* modify comments

* Update test_sparse_reshape_op.py

* chang the type of "shape"  from  vector<int64_t>  to  IntArray

* check whether sp_out.to_dense() is the cause  of error

* print sp_out

* Update reshape_kernel.cc

* use numpy to generate the equal paddle tensor

* just check dense_tensor.numpy()

* check cpu and cuda versions

* Update test_sparse_reshape_op.py

* supply all test cases for cpu forward coo kernel

* test forward coo cuda kernel

* change configuration of cuda kernel

* keep only one test case

* test coo cpu kernel (forward and backward)

* row major or column major ???

* test cuda coo forward kernel

* complete declaration and registration

* Update __init__.py

* rebuild

* retrigger CI

* add cudaMalloc and cudaMemcpy  in  ReshapeCooKernel  and change back to row major order in a cuda dense tensor

* midify minor error

* test only cpu coo forward kernel

* add all test cases for coo forward kernel  (both cpu and gpu)

* test all forward kernels (coo, csr; cpu, gpu)

* add all test cases for all kinds of kernels

* just retrigger CI

* Update sparse_ops.yaml

* Update sparse_ops.yaml

* Update sparse_ops.yaml

* resolve conflicts

* Update sparse_ops.yaml

* don't specify tensor place

* new shape has -1 or 0 in it

* Update unary_grad_kernel.h

* correct lvalue error

* code style

* Update sparse_backward.yaml

* Update sparse_ops.yaml

* Update unary_kernel.h

* Update unary.py

* Update sparse_backward.yaml

* Update unary.py

* code style

* code style

* code style

* Update unary.py

* specify tensor place explicitly

* do not use numpy array

* use numpy array in unit test again

* modify example code in docstring

* first version of ln_s_p with s>0

* refine and UT

* pass opt draft

* pass opt

* code refine

* code-style

* bug fix

* fix ci test

* code style

* skip ReplaceAllReduceOp in GraphtoBlock when nccl_ctxs_ is nullptr

* update ut

* test_dist_allreduce_op failed

* fix test_dist_allreduce_op

* add ut

* fix nccl cpu compile

* fix

test=kunlun

* add singleton to custom device

* Update custom_device.cc

Init device_init_flag_ in default

* simplify conv_mkldnn op registration

* remove custom type value in conv grad op

* draft with debug print