hifigan的训练无法直接使用cpu，且修改代码后无法接着训练

**Summary[问题简述（一句话）]**
A clear and concise description of what the issue is.
hifigan的训练无法直接使用cpu，且修改代码后无法接着训练
**Env & To Reproduce[复现与环境]**
描述你用的环境、代码版本、模型
最新环境、代码版本，模型：hifigan
**Screenshots[截图（如有）]**
If applicable, add screenshots to help
将MockingBird-main\vocoder\hifigan下trian.py中41行torch.cuda.manual_seed(h.seed)改为torch..manual_seed(h.seed)；
42行 device = torch.device('cuda:{:d}'.format(rank))改为device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')。
之后可以以cpu训练，但每次运行相同代码无法接着上一次训练。

Start with vocoder/hifigan/train.py, especially the reported lines 41-42, and inspect how device selection and checkpoint loading are handled. Reproduce the CPU training attempt and the second run with the same code; done means HiFi-GAN trains on CPU and resumes from the previous run without restarting.

Written by the indexing model from the issue text.