Issue: Sudden Increase in Loss During Training GPT-SoVITS GPT S1 Model for a New Language

[kikozi2000]

I'm training the GPT-SoVITS GPT S1 model on a new language. I've made the following modifications to support the new language:

Added new phonemes in symbol2.py
Extended the text embedding dimensions in the model

Problem: During training, the loss initially decreases as expected but then suddenly increases significantly. I've tried experimenting with different learning rates and scheduling strategies, but the same problem persists across all variants.

What could be causing this sudden increase in loss?
Are there additional steps or configurations I should consider when extending the model for a new language?
Any guidance or suggestions would be greatly appreciated!

[XXXXRT666]

You may load the model directly and ignore size mismatch