Blacklist torch version breaking execution on RPi

[NeonBohdan]

@NeonDaniel please provide a description of RPi problem

It may be related to new functionality, then it may happen on x86 too, or get working again with the next torch hotfix

Or it's related to known numpy, torch versioning problem on RPi As torch uses numpy internally and numpy and torch should have the same version of numpy c++ precompiled stuff

[NeonDaniel]

@NeonDaniel please provide a description of RPi problem

It may be related to new functionality, then it may happen on x86 too, or get working again with the next torch hotfix

Or it's related to known numpy, torch versioning problem on RPi As torch uses numpy internally and numpy and torch should have the same version of numpy c++ precompiled stuff

I could only trace it back to a vague error somewhere in synthesizer.tts:

 > Text splitted to sentences.
['hello']
Illegal instruction

I didn't catch the exit code, but it did kill the Python process

[NeonBohdan]

Thanks, then it looks like low level error with bad logging

I will look into it on my RPi

[NeonBohdan]

@NeonDaniel for me RPi torch==1.13.0, cp39 breakes too It breakes on this line

This is a transformer layer, and torch==1.13.0 updates transformers behaviour a lot It's only RPi problem, so it's something related to compiled ARM operations

---

For now torch~=1.9,!=1.13.0 is fine If the problem will repeated with future torch versions we can rebuild pickle model files or investigate native coqui model behaviour on RPi