Unload model does not close cuda process

[spaghiajoeojo]

Using --unload-timer 300 and --preload-model xtts_v2.0.2 I can see the model correctly loaded in memory through nvidia-smi.

After 5 minutes (300 seconds) GPU memory drops almost by 95% (from 1938MiB to 104MiB) but the python process remains active not allowing the GPU to enter idle state drawing much more power from PSU.

[matatonic]

I'm aware of this issue, the fix requires an architecture change to use a worker process to load/use the model, I'm working on the architecture change for this already.