sserdoubleh
sserdoubleh
Add some usefule scripts. - [x] evaluation - [x] self-chat
1. Upgrade padding strategy for AMP. 2. Update AMP white/black list. 3. Support pure fp16 training. 4. Add more amp settings. 5. Add convert checkpoint tools. 6. Support fp16 inference.
### PR types Bug fixes ### PR changes ### Description Fix the allocated size for beam search in FT.
### 软件环境 ```Markdown - paddlepaddle: - paddlepaddle-gpu: 2.6.0 - paddlenlp: 2.7.2 ``` ### 重复问题 - [X] I have searched the existing issues ### 错误描述 ```Markdown 需要对 logits 除以 temperature,再进行 top...
### 请提出你的问题 https://github.com/PaddlePaddle/PaddleNLP/blob/37b4fe0285dbfd098fc29efcf3eef25984970cc0/paddlenlp/generation/configuration_utils.py#L149-L150 这里的 GenerationConfig 中,max_new_tokens 带 s,而 min_new_token 不带 s,参数名格式不统一,最好还是统一成带 s 的?