Keep getting OOM

[zono50]

I have a NVIDIA Geforce RTX 4070 with 12GB of Vram, using a conda environment with a python venv environment. I am using nvtop to monitor my vram using, and it's using between 35-40% vram til about epoch 2 or 3, then it jumps up to 100 and goes oom.

Is there a memory leak? or am i doing something wrong that is causing this? I have lowered all the settings to the bare minimum.

Epochs - 6
Batch size - 2
Grad accumulation steps - 2
Seconds 7

I'm using cudatoolkit with 11.8 and CUDNN 8.9.2.26. I'm doing the large v3 model, and it always happens between the 2/3 epoch

[Mycohl]

There's no information on this repo about how much VRAM this thing requires.
I can't even run a single epoch with 8GB. Immediate OutOfMemoryError.
Google search shows people variously claiming that "finetune" works on 8GB, or that it requires 16GB.

I can run the regular xtts repo and RVC without issue.

This repo, unlike the regular xtts repo, doesn't appear to support the --lowvram argument.

My only next option is to get it set up on Windows, because the driver apparently allows CUDA to share system RAM.