Yam(长琴)

@Abhinay1997 I modified the code. `strict=True` is always a good habit. I just found another issue (it's my own problem, i modifed the arch of xtts). now waiting for the...

modify this in `hifigan_decoder.py`: ```python - resblock_type, - resblock_dilation_sizes, - resblock_kernel_sizes, - upsample_kernel_sizes, - upsample_initial_channel, - upsample_factors, - inference_padding=5, - cond_channels=0, - conv_pre_weight_norm=True, - conv_post_weight_norm=True, - conv_post_bias=True, - cond_in_each_up_layer=False, +...

maybe ```python + conv_pre_weight_norm=True, + conv_post_weight_norm=True, + conv_post_bias=True, ``` when training. (means not remove_parametrizations)

非vLLM 版本流式输出效果不降，LLM推理大概610ms，flow开启trt190ms，听到第一段Speech大概810ms。

@tiaotiaosong A100哦，按官方教程稍微改改就行