* Rename dropout is test (#43098)

* replace dropout_is_test with is_test.
* improve atol on a100.

* fused_attention fused_feedforward api support Model Tensor Parallel (#42985)

* fix is_test bug in fused_feedforward. (#43508)
Co-authored-by: NLi Min <11663212+limin2021@users.noreply.github.com>

[cherry-pick 2.3] Add fused_multi_transformer op to optimize transformer generation performance (#42311)

* Add fused_multi_transformer op to optimize transformer generation performance (#41814)

* fix fused_multi_transformer compile failed in cuda arch < sm53 (#42315)

* fix ci timeout

* implement autotune python API

* fix doc

* fix windows error

* fix doc and enable auto-tuning when config is None

* fix windows error

* fix doc

ref clipgradbynorm

update

Update __init__.py

cherry-pick from develop PR https://github.com/PaddlePaddle/Paddle/pull/42254

Cherry Pick #42220
Delete resnet_unit api from __all__

* fix recompute

* modify return

* fix moe apis (#41650)

* Moe ref (#41836)

* moe ref

* ref commit

* update; document_fix

* update;document_fix

* Moe ref (#41864)

* moe ref

* ref commit; document_fix

* update; document_fix

* update document_fix

* update; document_fix

[Cherry-Pick]Add hard_swish/kron/linspace/logit/graph_send_recv/multi_dot/maxout/multiplex op yaml file  (#41566)

* [Phi]Add graph_send_recv yaml file (#41206)

* add graph_send_recv yaml

* deal with confict

* fix compile bugs

* cherry-pick pr 41298

* cherry-pick pr41550

* fix compile bugs

* [Eager]Fix segment_pool/allclose/isclose/scale API bug (#41506)

* [Eager]Fix segment_pool/allclose/isclose/scale API bug

* fix kernel register problem

* add norm, segment_pool (#41465)
Co-authored-by: Nhong <43953930+phlrain@users.noreply.github.com>

* cut off relation between xk and initial_position's graph

* fix_bug

* add detach to cut off with original graph

* Add graph_reindex API

* add graph_sample_neighbors api

* Add buffer

* delete VLOG

* delete thrust::copy for output

* add ShareDataWith

* delete graph_reindex hashtable output

* add graph_reindex dispensable

* add reindex unittest, move memset to cuda kernel, change api

* fix conflict

* add reindex buffer for gpu version note

* fix conflicts for op_func_generator

* Add fisher_yates sampling, add dispensable, change infermeta

* add dtype for edge_id

* fix rocm ci and static check ci

* add unittest

* fix unittest

* fix unittest

* fix bug

Enhance vjp/jvp/Jacobian/Hessian API for supporting dynamic, static graph and batched, unbatched mode (#40692)

* modify vjp/jvp for both dynamic and static graph

* enforce jacobian class for supporting first/last batch

* add unittest for jvp, jacobian withlast batch, jacobian with first batch

* fix the incorrect shape when multi-index Jacobian

* enforce Hessian class for supporting dynamic graph

* add Hessian class unittest

* bugfix, jvp double_backward_trick zeros_like return stop_gradient=True in static graph

* add API beta warnnings

* add white_list for cuda11.x ci windows.

* optimize some code snippets and documments

* set unittest timeout to 100 seconds

* move vjp,jvp,Jacobian,Hessian to incubate

* fix vjp,vjp import path of sample code

* fix code style error of augtograd/__init__ file

* change vjp to paddle.grad

* use grad and gradients api

* fix preprocess for x

* fix a bug, val_and_grad should return a Tensor

* detach value and grad to avoid assign error
Co-authored-by: Nlevi131 <limaolin01@baidu.com>

* [New API]: miminize_bfgs and miminize_lbfgs

* modify for python module call correctly

* add functional package, add error raise in static_graph, change assign to set_value

* unify static_graph and dygraph, fix bug when x or H0 is float64

* now only accept input is tensor, put check args in utils.py, put exception test together

* temp

* add more detailed algorithm illustration and comment, reduce test case to limit test time in 15s

* change in_dygraph_mode to in_dynamic_mode

* fix bug of sample code; reduce test case to reduce test time

* change dir to incubate

* add random routing op

add _random_routing api in utils

add random routing ut

* # This is a combination of 10 commits.
# The first commit's message is:
add expert count op

add ut for expert_count

# This is the 2nd commit message:

update UT only for cuda

# This is the 3rd commit message:

fix for rocm

# This is the 4th commit message:

update ut

# This is the 5th commit message:

add moe module

# This is the 6th commit message:

add expert count op

add ut for expert_count

# This is the 7th commit message:

update UT only for cuda

# This is the 8th commit message:

update ut

# This is the 9th commit message:

add moe module

# This is the 10th commit message:

make expert count private

* add assign pos op

* fix upper num name

* add api _assign pos

* add ut for assign pos op

* update date

* add op about moe gate

update utils

add limit by capacity op

add ut for limit_by_capacity

add ut for prune_gate_by_capacity

add ut for limit_by_capacity

add ut for prune_gate_by_capacity

* fix for win

* fix bugs in test_limit_by_capacity_op

* update ut

* update for test (timeout)

* fix ut

* update

* update(fix) ut for win

* moe apis in incubate

* # This is a combination of 10 commits.
# The first commit's message is:
add expert count op

add ut for expert_count

# This is the 2nd commit message:

update UT only for cuda

# This is the 3rd commit message:

fix for rocm

# This is the 4th commit message:

update ut

# This is the 5th commit message:

add moe module

# This is the 6th commit message:

add expert count op

add ut for expert_count

# This is the 7th commit message:

update UT only for cuda

# This is the 8th commit message:

update ut

# This is the 9th commit message:

add moe module

# This is the 10th commit message:

make expert count private

* add assign pos op

* fix upper num name

* add api _assign pos

* add ut for assign pos op

* update date

* fix for win

* update for test (timeout)

* fix ut

* update

* fix ut for number count

* add apis and utils

* add gate apis

* add moe and grad clip apis

* update moe apis

* add ops for moe gate

* fix

* update for base moe layer api

* add random routing op

add _random_routing api in utils

add random routing ut

* fix for dygraph

* update with ranodm routing

* update

* fix ut for limit by capacity

* update

* update limit by capacity for easily to switch to single thread mode

* update api docs
Co-authored-by: Nhlygit66666 <2570058140@qq.com>

* add random routing op

add _random_routing api in utils

add random routing ut

* # This is a combination of 10 commits.
# The first commit's message is:
add expert count op

add ut for expert_count

# This is the 2nd commit message:

update UT only for cuda

# This is the 3rd commit message:

fix for rocm

# This is the 4th commit message:

update ut

# This is the 5th commit message:

add moe module

# This is the 6th commit message:

add expert count op

add ut for expert_count

# This is the 7th commit message:

update UT only for cuda

# This is the 8th commit message:

update ut

# This is the 9th commit message:

add moe module

# This is the 10th commit message:

make expert count private

* add assign pos op

* fix upper num name

* add api _assign pos

* add ut for assign pos op

* update date

* add op about moe gate

update utils

add limit by capacity op

add ut for limit_by_capacity

add ut for prune_gate_by_capacity

add ut for limit_by_capacity

add ut for prune_gate_by_capacity

* fix for win

* fix bugs in test_limit_by_capacity_op

* update ut

* update for test (timeout)

* fix ut

* update

* update(fix) ut for win

* moe apis in incubate

* # This is a combination of 10 commits.
# The first commit's message is:
add expert count op

add ut for expert_count

# This is the 2nd commit message:

update UT only for cuda

# This is the 3rd commit message:

fix for rocm

# This is the 4th commit message:

update ut

# This is the 5th commit message:

add moe module

# This is the 6th commit message:

add expert count op

add ut for expert_count

# This is the 7th commit message:

update UT only for cuda

# This is the 8th commit message:

update ut

# This is the 9th commit message:

add moe module

# This is the 10th commit message:

make expert count private

* add assign pos op

* fix upper num name

* add api _assign pos

* add ut for assign pos op

* update date

* fix for win

* update for test (timeout)

* fix ut

* update

* fix ut for number count

* add apis and utils

* add gate apis

* add moe and grad clip apis

* update moe apis

* add ops for moe gate

* fix

* update for base moe layer api

* add random routing op

add _random_routing api in utils

add random routing ut

* fix for dygraph

* update with ranodm routing

* update

* fix ut for limit by capacity

* update
Co-authored-by: Nhlygit66666 <2570058140@qq.com>

* refactor eager flags

* fix flags error when we switch from eager to dygraph

* fix ci problem

* fix ci

* fix ci

* merge develop and fix code style

* merge develop and fix code style

* fix op test error

* fix op test error

* fix op test error

* fix op test error

* fix op test error

* merge develop

* add out_size shape for graph_send_recv

* fix bug in register kernel: no const int& support

* add out_size in infermeta

* change unittest

* fix unittest

* fix out_size default value

* fix doc

* delete arg mapping

* add sig

* move -1 to 0

* move -1 to 0

* segment pool support for int int64 kernel.

* add support in python api

[multiprocessing] Add paddle.incubate.multiprocessing for sharing tensors  between python processes. (#37302)

* Add support for paddle.multiprocessing
* move multiprocessing to incubate.

* vectorize lamb kernel

* remove flags, add ut

* remove useless codes

* refine code, add param order

* add multi tensor apply l2 norm

* add multi_tensor_apply code

* make sizeof(TensorMeta) smalller

* move code to distributed_fused_lamb_op.cu

* remove useless FLAGS

* fix 'invalid escape sequence'

* fix assert error

* add DistributedFusedLamb op

* polish code

* fix compile error

* compatible with pten changement

* fix rocm compile error

* improve converage

* update upstream/develop

* fix cast_with_ptr.h

* add FLAGS_distributed_lamb_divide_nranks_when_allreduce=1

* fix clip before allreduce

* add use_master_param_norm

* code polish

* fix bug

* fix ROCM ci

* add the test case for the UVA

* add the context load for the uva

* Add graph_sample kernel

* Add graph_sample commit

* add new commit for graph_sample

* add unsigned long long int

* delete some remarks

* add cpu version

* add cuda eids

* add cpu eids

* delete _uva

* optimize speed: emplace_back, last_layer

* add to_uva_tensor

* add cpu return_eids choice

* add gpu return_eids choice

* add cpu reindex_nodes

* add gpu reindex_nodes

* rename op and add OMP for cpu

* add incubate api

* fix the compile problem for the PADDLE_ENFORE and different device

* fix the rcom and windows compile problem

* add unittest for graph_sample_neighbors

* fix cpu unittest and unique problem

* fix uva unittest, fix cuda unique problem

* fix the windows compile problem

* fix the windows rand_r compile problem

* add correct unittest, add src_eids dispensable

* delete black

* combine uva unittest

* mv Sample_index to Sample_Index; check input shape; fix random sample func

* delete memset & cudaMemset

* fix according to PR comments

* fix rocm ci

* modify function names according to the specification

* fix windows_openblas ci

* refine annotations, fix windows unittest, add default value for uva device_id, fix bug for input nodes with empty neighbors

* fix rocm ci

* rename graph_sample_neighbors as graph_khop_sampler, add incubate api doc

* add data type

* fix conflict
Co-authored-by: Nwawltor <fangzeyang0904@hotmail.com>

* fix bug:
1. atten: set the default value of attn_dropout_rate to None
2. ffn: add activation parameter

* for pure fp16

* Add a SparseCsrTensor

* remove unused functional

* remove const

* remove SetMemoberTensor

* remove non_zero_nums_, the number of non zero elements of each batch can be obtained from the crows

* SparseCooTensor

* add SetMember

* merge upstream; add SetMember

* merge upstream

* merge upstream; add newline at end of file

* add newline at end of file

* remove newline at end of file

* remove newline at end of file

* stash

* user pten::framework::make_ddim

* user pten::framework::make_ddim

* merge upstream; use the latest mutable_data

* merge upstream; use the latest mutable_data

* return mutable dense tensor

* Fix bugs when bias is none for static graph for fused_attention op.

Add support for bias is none for fused_attention op.

* GeneratePass support attr condition and mapping, test=develop

* fix coverage, test=develop

* Add fuse_resnet_unit pass, test=develop

* fix CI errors, test=develop

* fix CI errors, test=develop

* fix unittest error when compiling without CUDA, test=develop

* fix static ci error, test=develop

* limit kernel size must equal 1, test=develop

* add cpu version, using set: sum, min, max

* add cpu version: mean

* improve cpu code and fix dynamic memory allcation problem

* fix arg error, add index judge, delete fp16

* fix bug in CudaAtomicMax and CudaAtomicMin

* add CUDA version

* fix grad_op bug for index

* add op test, add correct cpu grad op

* Add correct CUDA Mean grad

* [Add] Successful MEAN and SUM

* [Add] Successful MIN and MAX in CPU

* [Add] Successful MIN and MAX in CUDA

* fix windows dtype ci

* fix ROCM ci by adding HIP flag

* rename fused_gather_scatter to send_recv

* unify name as send and recv

* change zero index return time

* add send_recv incubate api

* fix index data type, add unittest case for API

* delete redundant input tensor

* fix en example and docs, add default value in pool_type

* add shape judge and max grid judge

* fix comment

* fix index type bug

* add const &

* fix en docs

* delete numpy in examples

* add unittest for int input

* fix send_recv comment

* change send_recv to graph_send_recv

fused_attention_op的实现中，使用了bias_add，且其实现是通过使用kernel primitive来实现的，之后kernel primitive的WriteData api接口及函数内部实现发生了更改，将判断越界的逻辑移到了template的参数中，使得调用的分支有错误，产生了越界赋值操作，污染了别的显存空间的内容。具体表现为：test_fused_attention_op_api.py 单次执行基本上不会报错，多次循环执行不同shape的输入，结果计算不对，具有偶发性，bug不易察觉。

* fix bug:
1. atten: set the default value of attn_dropout_rate to None
2. ffn: add activation parameter

* Fix fused_attention english doc test=document_fix

本PR是fused_transformer的layer层代码，包含FusedFeedForward的layer层代码和FusedTransformerEncoderLayer的代码。

功能：本PR的目标是提高attention模块的计算性能。
为了减少框架层对op的调度开销，本PR通过在C++层手动实现attention模块，对外提供attention 大op；
为了减少防存开销，本PR采取了两种优化方法：
（1）在q,k,v计算时通过共享输入X，将该处的gemm，transpose和bias add从三次调用减少为一次；
（2）使用kernel融合优化技术，在不同cuda kernel之间通过寄存器传输数据；