* support load state_dict from save_params/persistables

* remove failed unittest

* add load eof check & unittest

* remove eof check

* fix error message

* add input_spec & output_spec for translated_layer

* update error message

* fix ports conflict when launching multi-nodes in paddlecloud;test=develop

* add DISTRIBUTED_TRAINER_ENDPOINTS env for cloud;test=develop

* optimize slice TRT plugin

This patch removes unnecessary barrier for data transfer of needed offset,
so data transfer can be overlap with GPU kernel execution.

This patch also fixes incorrect name of slice plugin. That is, replaces
"layernorm" with "slice"

test=develop

* add serialize/deserialize to slice plugin

* add static shape slice trt plugin

* fix slice trt op convertor dynamic shape bug

* fix format by clang-format

* fix pylint format error

* fix problems commented by peiyang
Co-authored-by: NRyan Jeng <rjeng@nvidia.com>

* optimize errror report

* add test case for pad op converter

* fix some spelling mistake commented by peiyang

* replace sequence length attr to input

* Remove the cache in post_traning_quantization, test=develop

* add check for sparse parameters with weight_decay

* move sparse check to adam.py

* move worker loop to top level

* move reader process loop to top level

* fix failed unittests

Update amp_check_finite_and_scale_op and add an updating_loss_scaling op for static graph amp training. (#26240)

* update amp_check_finite_and_scale_op for static_amp.

* use amp_check_finite_and_scale in static graph amp.

* update grads to zero when grads own infinite values(as for amp_checkout_finite_and_scale op).

* add update_loss_scaling op in cpp.

* add update_loss_scaling_op unit test.

* update the doc of the check_finite_and_unscale op

* Update the process of gradients updating skipping if the gradients have infinite values.

* update the way to zero grads.

* update test_update_loss_scaling_op.py

* add log info when find infinite grads.

* add the unit test for UpdateLossScaling Layer.

* rm auto from localsgd

* Add int8 GRU kernel with UTs

* Lint fixes

* More lint fixes

* temporally disable zero_copy

* add test

* follow comments

* fix calcu_gradients

* fix code place

* fix embedding interface usage

* support to_static(model)

* add warning and unittest

* Use the single GPU card to execute the test_fuse_bn_act_pass UT.

* add double grad for tile, test=develop

* add double grad for expand_v2 op, test=develop

* add double grad for expand, test=develop

* add some unittest cases ot verify jit.save, no_test

* add more unittests

* add test with example inputs

* polish implement details

* remove useless blank

* fix fetch random error

* parameter_server_optimizer support auto_strategy

 fix the CudaPinMemory bug for the equal op and add the test case for the equal op

Decrease the number of running iterations to reduce CI time.

CI system shows it decreased the unittest time from about 90 seconds to about 30 seconds