* Add pure fp16 training with master weights. (#27712)

* add the weight decay func for the momentum op

* Add the multi_precision function in Momentum Optimizer.

* Make sure that the initial value of master weights are same with the fp16 weights.

* add static loss scaling.

* add the rescale_grad function in the pure fp16 training.

* use the original momentum updating method.

* Polish some codes, such as variable names.

* add docstring for apis.

* update the var creation details of _create_master_weight.

* not modify codes about imperative momentum updating.

* Fix the error of test_dist_sparse_tensor_load_momentum UT.

* add unit test for multi precision fp16 training.

* add more unit tests for CI.

* Use lower threshold values for allclose comparing in test_multi_precision_fp16_train UT.

* add lamb optimizer and unittest

* fix momentum resume training

* fix momentum acc

* The leaf tensor concept is exposed and the gradient accumulation of leaf tensor

* The leaf tensor concept is exposed and the gradient accumulation of leaf tensor

* fix coverage

* fix api doc

* fix CI unittest

* fix CI unittest

* fix unitest

* empty tensor does’t need inner_var_

* fix some error message

* refactor momentum op to combine weight_decay (scale op and sum op)

* add doc; notest

* fix doc; notest

* update doc; notest

* refine optimizer && adam

* refine optimizer; notest

* add adam

* fix doc

* fix doc && add adamw; notest

* add error message

* bug fix

* refine rmsprop && adamax

* fix ci

* buf fix

* update comment

* unify arguments place; notest

* fix ut, test=develop

* bug fix

* fix conflicts, test=develop

* add examples code

* bug fix

* fix comments

* fix sample code

* add sample code for Optimizer

* add adamax ut, test=develop

* fix rmsprop ut, test=develop

* add ut for optimizer.py and adamw.py

* first commit of adadelta optimizer

* fix learning rate

* fix adadelta doc and add sgd momentum

* remove unused fluid

* fix codestyle

* Update test_adam_op.py

* Update test_adam_op.py

* fix SGD in 2 unittests

* fix SGD in 2 unittests

* fix ci

* fix ut
Co-authored-by: NMRXLT <xlt2024@gmail.com>
Co-authored-by: Nmapingshuo <mps2012@yeah.net>

* wip

* wip

* wip

* wip for test

* add fp16 tests test=develop

* fix cpu build test=develop

* fix test=develop

* fix py3 tests test=develop

* fix lr_scheduler dtype test=develop

* fix test=dvelop

* test fix ci compile test=develop

* fix build and merge test=develop

* fallback momentumop change to general test=develop

* make fp16 lr schedule simple test=develop

* fix ut test=develop

* fix tests test=develop

* remove fp16 learning rate cast test=develop

This reverts commit 3d750f9c.

* wip

* wip

* wip

* wip for test

* add fp16 tests test=develop

* fix cpu build test=develop

* fix test=develop

* fix py3 tests test=develop

* fix lr_scheduler dtype test=develop

* fix test=dvelop

* test fix ci compile test=develop

* fix build and merge test=develop

* fallback momentumop change to general test=develop

* fix nccl2 lars dist support

* put lars in momentum op

* add tests lars

* fix ci

* fix cpu kernel

* soft warning

* remove lars in test_recognize_digits.py

* move to another op

* add file

* update api.spec test=develop

* update test=develop

* fix api.spec test=develop

* wip

* wip, finish grad merge ops

* wip, finish graph build

* wip test running

* work on 1 gpu

* workable version

* update

* fix tests

* fuse broadcast op

* fix compile failed

* refine

* add batch merge test mnist

* fix CI test=develop

* fix build

* use independent bn params for batch merge test=develop

* update api.spec

* follow comments and for test

* wip

* refine tests test=develop

* follow comments test=develop

* remove startup bn modify test=develop

* follow comments test=develop

* fix merge test=develop

* Fix serious bug in nesterov momentum optimizer.

Currently, our tests run with 2 GPUs, the init time is absurdly long:
about 4s for each process.  Currently, we run each OP test on
different processes. This PR:

1. create cmake function py_test_modules which will generate the
Makefile that runs a list of Python unittest module in a single Python
process.

2. move all "python unittest compatible" (e.g., used the unittest
package, not just a regular python file). from fluid/tests to
fluid/tests/unittests.

3. cmake now will run all OP tests in fluid/tests/unittests in a
single process, except the time-consuming tests, they are separated
into different processes to utilize parallelism. Please make sure to
use the unittest package if you put the python test file in
fluid/tests/unittests

4. remove all exit(0) from fluid/tests/unittests/*.py, exit(0) is used
to disable unittest, we can not do it when running all tests in a
single process since it will terminate the process without running the
other tests. Instead, the test is disabled in
fluid/tests/unittests/CMakeLists.txt. FIXME is added for each disabled
item. Please disable the unittest from
fluid/tests/unittests/CMakeLists.txt, instead of adding exit(0) to the
Python file, for all Python file in fluid/tests/unittests/.

5. add an option WITH_FAST_BUNDLE_TEST. When OFF, will run the unit
tests in separate process so that they can be tested individually.

* "fix decode bug"

* "follow commnet"

* "fix error"

* "fix hook bug"

* fix based comment

* fix copyright

* fix based on comment

* add copyright hook

* add copyright hook

* refine copyright hook

* "test copyright hook"

* fix check style

* fix ci

* init commit

* change some dir name

* Fix attribute naming for momentum_op

* Fix minor typo in comment

* Fix attribute name

* Fix names in test_optimizer

* Fix python wrapper