Zhong Hui

### PR types Function optimization ### PR changes Others ### Description Enhance ERNIE pretrain with additional chinese char

### PR types New features ### PR changes Models ### Description Support GPTForGeneration for dygraph. - [ ] support attn_mask as input - [ ] support stop with eos_id

[文本显示] VDL 文本数据只能显示一行，而且无法下载

1

如图所示，目前文本只显示一行，长一点的文本看不全。而且不能下载，用户很难查看数据

[Model] [PinSage] Add pinsage model.

1

Add PinSage model.

[Model][LGNN] Add line graph neural network.

1

[Model][LGNN] Add line graph neural network.

### PR types Others ### PR changes Others ### Description Fix amp usage for evaluation.

### PR types New features ### PR changes APIs ### Description support sharding for trainer. stage1: 可以支持 stage2：部分支持 - [x] offload 暂不支持，需要修复pure_fp16 stage3：暂不支持 - [ ] 模型保存存在问题

[Bug]: 混合精度类型提升不符合预期

1

### 软件环境 ```Markdown - paddlepaddle: - paddlepaddle-gpu: - paddlenlp: ``` ### 重复问题 - [X] I have searched the existing issues ### 错误描述 ```Markdown 类型提升不符合预期 ``` ### 稳定复现步骤 & 代码 ```python...

[Performance] Optimize unified checkpoint save/load speed.

3

### PR types ### PR changes ### Description For `facebook/llama-7b` UC/PD Save/Load 对比 | UC save | PD save -- | -- | -- testDP8 | 0 | 0 testPP4DP2...

Improve performance for `PyByteArray::new`

5

Hello, recently i found the loading speed of `safetensors` is more than 50% slow than loading pickle. https://github.com/huggingface/safetensors/issues/460 for a `1GB` numpy array, loading by pickle only need `1s`, but...