A question about the performance

[Shijun-droid]

Thanks for your work and the pretrained models!

I'm playing around the pretrained models, however, the quality of the generated samples is not very good. Perhaps I did something wrong.

The only two changes I made are:

VoxInstruct/model/ar.py

Line 42 in e953295

_attn_implementation="flash_attention_2"

_attn_implementation="flash_attention_2"
->
_attn_implementation="eager"

VoxInstruct/model/nar.py

Line 42 in e953295

_attn_implementation="flash_attention_2"

_attn_implementation="flash_attention_2"
->
_attn_implementation="eager"

The reason for this change is that I only have V100 GPUs, which do not support flash-attn. The other hyperparameters are followed by infer.sh.

Thank you for your help!

[Value-Jack]

me too, i only change the _attn_implementation and get the bad performance, do you solve this problem？