Is it possible in modifying RVC to allow 6GB of VRAM GPU to train RVC models?

[wraith saddle]

Currently I heard RVC Disconnected recently has no longer will have any updates or fixes but I believe it has a risk of being broken whenever Google Colab does any changes. And I want to maintain the quality as I have for my RVC models I made instead of it being trained in a lower quality in Weights.gg.

I've also been wanting to train them locally for awhile without dealing time limits from Google Colab. I saw someone tried to modify RVC to lower usage of VRAM slowing the training time down to prevent getting errors or computer freezes. and I'm trying to look for the file that can change the self.gpu_mem or something that slows down training. One way was to adjust the amount of batches but otherwise no idea what I can do to reduce enough to train slowly.

I recently got a GPU upgrade a few months ago and I was able to do the voice to voice conversion and RVC Realtime but otherwise I haven't train any models yet. The GPU I have currently is a NVIDIA GeForce GTX 1660. And the current RVC software I use is "RVC1006Nvidia".

Here's this link I found that relates to it.
https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI/issues/649

[primal ether]

Just use applio and enable checkpointing

[wraith saddle]

Nvm I just used 3 batches and have caches off. Currently my rvc model is 34 epochs in.

I can probably use Applio for to use the pretrains.

But gonna start with the regular for now.

Actually I'll give Applio a go just for training. But I'll see if it works with just 3 batches. But I'll install it in another time.

[spare thicket]

batch size 4-8 with checkpointing recommended

the batch recommendation as half of vram seems misleading tho

[wraith saddle]

I use a NVIDIA GeForce GTX 1660 just fyi and my model is 153 epochs in