* Add multihead fuse pass for ernie opt

* Refine softmax

test=develop

* Refine cuda kernel

* Refine cuda version

* Refine cmake

test=develop

* refine header file

* refine test case and pass
* refine comments

* Add ConvTranspose + BatchNorm fuse pass
test=develop

* Add tests for conv+bn and conv_transpose+bn passes
test=develop

Fix parse content in CreatePreLoadReaders. Before this fix, if you use dataset.set_parse_content and dataset.preload, parse content didn't work.

* add new save load interface; test=develop

* add new save interface; test=develop

* add save load interface ;

* fix save load error;

* fix dygraph set dict bug;

* add save load unit test; test=develop

* fix test_imperative_optimizer bug; test=develop

* fix unitest optimizer bug; test=develop

* fix code coverage; test=develop

* fix converage; test=develop

* add document for apis; test=develop

* fix unitest error; test=develop

* fix save load unit test error; test=develop

* fix error message; test=develop

* change set_parameter set_optimizer to save_dygraph; test=develop

* add load_graph check; test=develop

* fix api spec; test=develop

add executor.FetchHandler for train/infer from the dataset

test=develop

* How to write custom op needs to follow framework OP spec.
* Package fluid_framework.so and headers into whl.
* Add paddle.sysconfig.get_include() and paddle.sysconfig.get_lib() to get include dir and lib dir.
* Export some C-APIs to merge OpInfo between core.so and custom_op.so.
* Add unit testing.
* Update API.spec.

* Follow Wangzhen's comment in PR 18970, test=develop

* Review comments, test=develop

* Leave fake quantization around mul

test=develop

* Replace Fake with Real Quantized Mul

test=develop

* Fix bug in quantize placement pass

Nodes in the graph now have checked type instead of node name when they are to be marked for quantization test=develop

* update operator compatible info, test=develop

* revert cmake/version.cmake, test=develop

* add unit_tests and fix bugs, test=develop

* update ../paddle/fluid/framework/framework.proto, test=develop

* fix bug of paddle/fluid/inference/api/analysis_predictor.cc, test=develop

* update paddle/fluid/framework/version_test.cc, test=develop

* add comments and rename interfaces, test=develop

* Fix conv2d+dequantize squash for residual fusion

test=develop

* Disable conv-requant squash

test=develop

* codegen code for reconstruction test=develop

* fix the cmake test=develop

* fix review advice test=develop

* add a base class for the Communicator
* add AsyncCommunicator Impl for async distributed training

* shape and optimize paddle error message stack, test=develop

* limit exception type & add unittest, test=develop

* fix multi-platform problem, test=develop

* fix related unnitest failed, test=develop

* add doc & fix unittest errors, test=develop

* fix function name error, test=develop

* update tensor test exception msg compare, test=develop

* remove unittest on win32, the dir format is different, test=develop

* remove useless package, test=develop

* add paddle enforce handler unittest, test=develop

* add exception checkout, test=develop

* fix coverage failed, test=develop

* fix op registry test failed, test=develop

* refactor whole pr, test=develop

* remove test in CMakelist, test=develop

* fix coverage, test=develop

test=develop

Add dtype for coalesce_tensor_op

The new "fluid.data" changes old "fluid.layers.data":

1. Add shape and dtype check.
2. Remove "append_batch_size" parameter. We won't offer this in the new data layer because other deep learning platforms don't have this kind of data layer pre-processing. It may confuse users.
3. Remove "stop gradient" parameter because the data layer doesn't do back-propagation

TODO：
Now data layer feeded by executor is checked, will we want to check the feed data of readers in the future?

fix memory leak in HogwildWorker,  whose ops are  explicitly deleted in destructor

* support change shuffle thread num
* support change train thread num
* fix receive shuffle data of each channel
* data norm stop gradient
* add check thread_tensor type and root_tensor type when merge metric
* remove sleep in shuffle, add config
* add config of pslib client to client communication
* fix xbox str
* add data norm op testcase
* add flush in trainer finalize

test=develop

* Add RecordHistoryLocalExecScopes
test=develop

* remove the useless warning for user to avoid confuse test=develop

* add op compatible infomation; test=develop

* add enum type

* add enum type; test=develop

* Set states of recurrent op as dependent vars in prune of save inference model

This PR will fix the save/load inference model problem of RNN models.

The reason of the bug is that save_inferenc_model will prune OPs that doesn't contribute to Output. But in recurrent_op, States are not Output, OPs refers States will be pruned. 

This fix adds States of recurrent_op as dependent var so that OPs referring States won't be pruned. 

* Fix conv2d+dequantize squash for residual fusion

test=develop

* Change condition

test=develop

Add boost as dependency of prune

fix #19862

* Add fc_elementwise_layernorm_fuse pass and unittest.

* Add fused_fc_elementwise_layernorm op and its GPU kernel.
test=develop

* Apply fc_elementwise_layernorm_fuse_pass to GPU inference.

* Add the setting of attrs in the definition of binary_op.
test=develop

* Add comment.

* Implement the unittest.
test=develop

* Change the unittest name of layer_norm.
test=develop

* fix memory reuse bug on feeding variables, test=develop

* add comments to reference count members, test=develop

* rm return in vfork

* rm return in vfork
test=develop