[Bug]: PyTorch doesn't reserve all of available VRAM

[Spaceginner]

Is there an existing issue for this?

• [X] I have searched the existing issues and checked the recent builds/commits

What happened?

After switching to EndeavourOS (Arch Linux distro) from Windows, I noticed that PyTorch doesn't allocate all of available memory (even looking at nvdia-smi there is still like ~500 MB of free VRAM), and if I compare reported usage of python3.10 process and reported reserved memory by PyTorch they don't match.

Steps to reproduce the problem

idk how to reproduce this

What should have happened?

Allocate all of available memory

Commit where the problem happens

0cc0ee1bcb4c24a8c9715f66cede06601bfc00c8

What platforms do you use to access the UI ?

Linux

What browsers do you use to access the UI ?

Google Chrome

Command Line Arguments

--medvram --no-half --precision full --xformers --api --theme dark

List of extensions

a1111-stable-diffusion-webui-vram-estimator

Console logs

https://pastebin.com/b3Gwn2XD

Additional information

On Windows 11 it was reserving up to 5.5 GBs or so. Here, however, only ~3.5 GBs

[Xyem]

Why would expect it to reserve all available VRAM? It should only use what it needs.

[Spaceginner]

Because it was doing it on Windows. Also I resolved this issue by downgrading to 1.12.1 with CUDA 11.6. Upgrading to PyTorch 2.1 with CUDA 11.8 also resolves this issue.