RuohengMa

[PHI] add float16 support of accuracy; add NHWC support of pool2d and pool2d_grad

2

### PR Category Others ### PR Types Others ### Description Add float16 support of accuracy; add NHWC support of pool2d and pool2d_grad.

[FUSION_OP] add improved resnet_unit

2

### PR Category Others ### PR Types Others ### Description Add improved resnet_unit.

[fix] DP8 training

2

### PR types Bug fixes ### PR changes Others ### Description Llama2 13B 4k can not run in pure DP8 configuration on XPU. Fix this issue by setting no_sync() for...

- 系统环境/System Environment：Ubuntu 18.04 - 版本号/Version：Paddle：bb3fb69c56ddc8235b6f84325a1802ab9a295bfe PaddleOCR：b3912fcf7a567109a605f6be25a286e78da8952f 问题相关组件/Related components： - 运行指令/Command Code：python3.7 tools/train.py -c test_tipc/configs/sr_telescope/sr_telescope.yml -o Global.use_cpu=True Global.save_model_dir=./test_tipc/output/sr_telescope/lite_train_lite_infer/norm_train_gpus_0_autocast_null Global.epoch_num=1 Train.loader.batch_size_per_card=16 - 完整报错/Complete Error Message： - Traceback (most recent call last):...

[XPU]Restructure code for using paddle_xpu

4

### PR types Others ### PR changes Others ### Description Restructure code for using paddle_xpu.