Open gravityrail opened 2 weeks ago
The crash occurs on any OS with transformers
>=4.41. This fix was originally shared by @pseudotensor in https://github.com/idiap/coqui-ai-TTS/issues/31. But I just opened https://github.com/idiap/coqui-ai-TTS/pull/46 in our fork (since this repo is not updated anymore) which more generally aligns the streaming code to the upstream transformers
implementation again, so it should be more resilient to future updates. It will be available in the next release.
Quick fix for streaming responses on macOS. It was crashing somewhere in pytorch. Unfortunately I can't find the original post where I found this patch to give credit - too many tabs open! But it works for me.
For this to work, you must run with
PYTORCH_ENABLE_MPS_FALLBACK=1
and callmodel.to(mps_device)
on the XttsModel