Kaipeng Deng

**add fleetx/inference and InferenceEngine** - add `fleetx/inference` directory and `fleetx/inference/engine.py` - add `TensorRTConfig` and `InferenceEngine` API **test code** ```python def main(): from paddlenlp.transformers import GPTTokenizer fleet.init(is_collective=True) infer_engine = InferenceEngine("./inference_model_pp1mp2") tokenizer...

### PR types New features ### PR changes Others ### Describe **Add fused multi transformer pass fleetx**

fix hybrid export doc

1

**fix hybrid export doc**

add cn doc for GPU DataLoader

2

**add cn doc for GPU DataLoader**

**refine doc for metrics**

Is these speed testing on Kirin 990 CPU or NPU，and could you please give the benchmark testing code demo？

PointRCNN rpn use train_aug data

1

**PointRCNN rpn use train_aug data** use train_aug data also in RPN training, this will improve the AP [Uploading rpn_train.log…]() [rpn_eval_train_aug.log](https://github.com/PaddlePaddle/models/files/4024480/rpn_eval_train_aug.log) [rcnn_offline_train.log](https://github.com/PaddlePaddle/models/files/4024481/rcnn_offline_train.log) [rpn_eval_199_proposal.log](https://github.com/PaddlePaddle/models/files/4024482/rpn_eval_199_proposal.log) [rcnn_offline_eval29.log](https://github.com/PaddlePaddle/models/files/4024483/rcnn_offline_eval29.log) [kitti_eval.log](https://github.com/PaddlePaddle/models/files/4024484/kitti_eval.log)