Memory spike 100% RAM on 2060 with 512×768, GPU CUDA available

[rwynx]

Hey there. When generating images at 512×768 using Euler A, 20 sampling steps, no highres fix, with --xformers --medvram enabled, my system RAM spikes to 100% while GPU VRAM usage is low. Only things using my GPU at the moment are DWM and CSRSS. It causes Windows to freeze and sometimes crashes Opera GX (other browsers too, mostly) and some other apps like Discord, Steam etc. at the same time. PyTorch correctly detects my GPU, so the generation is supposed to run on VRAM.

With my settings I think even my 2060 can handle 512x768 and it does most of the time, but this crash thing and RAM spike also happens frequently. Using Python: 3.10.6, OS: Windows 10 (19045.6216), Torch: 2.1.2+cu121, xformers: 0.0.23.post1.

If there is something wrong with my setup please guide me to fix it. Thank you in advance.

About torch and cuda:

>>> import torch
>>> print(torch.cuda.is_available())
True
>>> print(torch.cuda.get_device_name(0))
NVIDIA GeForce RTX 2060
>>> print(torch.version.cuda)
12.1

Webui-user.bat:

@echo off

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS= --xformers --medvram --force-enable-xformers
set XFORMERS_FORCE_DISABLE_TRITON=1

call webui.bat