[Bug] PermissionError: [WinError 32] The process cannot access the file trainer_0_log.txt

[P3p1d]

Describe the bug

trainer_0_log.txt
Hi!
I am running the ljspeech recipe for xtts2 on a custom dataset (fine-tuning with about 45 minutes of audio and metadata).
After running for one epoch, multiple errors seem to rise up. I have tried different batch sizes and settings but it still occurs. The NotImplemented Error should be internally handled well, but I do not know what to do with the other ones. The log file is not opened during training so this should be an error on the coqui side.
Thank you in advance for your time

To Reproduce

I was running the train_gpt_xtts.py from Windows command line with admin rights. This is code provided in the source under recipes, ljspeech and xtts_v2.

train_gpt_xtts.py

The dataset is custom.

Expected behavior

I think a checkpoint should be created after completing all the steps.

Logs

Environment

coqui-tts 0.27.0
coquit-tts-trainer 0.3.1
torch 2.8.0+cu129
torchaudio 2.8.0
GPU is GTX 1650
python 3.11

Additional context

No response

[eginhard]

Could you share the exact error messages you get? There aren't any in the log file you shared.

[P3p1d]

I am sorry, I attached the wrong file. Here I have copied the console in the text file. The last line is in Czech, it is the classic Windows error that the process does not have the access to the file because another process is using it.
Thank you for your patience

error_report.txt

[eginhard]

Thanks! This should have been fixed by #462, which wrapped the training recipes with main() functions. But your script already has that fix, so I'm not quite sure why it still occurs...

[para133]

Hello, I have check the previous issue, and wrapped the training recipes with main() but still have the same problem. I also try to set num_loader_workers=0 and freeze multiprocessing，but not work. I will be thankful if you can help! @eginhard

[eginhard]

I don't have a Windows machine, so I'm not able to reproduce this unfortunately. But if someone figures out the exact issue, I'm happy to merge a fix.