Is the flow model trained with half-precision works well?

[JohnHerry]

In the pretained opensource stack, the llm support half inference but not flow model. If I train the flow with deepspeed and get a half-precision float model, will it works well? will it get good speech quality?

[aluminumbox]

not amp trained, we will update cosyvoice-300m and coysvoice-300m-25hz with amp training later

[JohnHerry]

ok， thank you for the help

[JohnHerry]

ok， thank you for the help
by the way, what I mean is to train the base model from scratch. not to fine tune on the released pretraines.