[Issue]: Shared Memory Usage (Intel Arc)

[tclakk]

Putting this here as do not wish to spam issues...

A few months ago (I would guess August last year) when I was running a SDXL model + Loras things were running pretty smoothly. On Windows 10 able to gen a 1280 x 960 image every ~30s.
Came back to using SDnext start of Jan 2025 -> initially unable to start, resolved by re-cloning repo into new directory and following documented install steps.
However, now when I generate a prompt I have an issue where the 1st image generated in a batch uses enough memory that the system dips into shared memory (slowing it down significantly ofc). After 1st generation memory usage drops back down and I get similar performance to a few months ago.

Is this a known issue? If so how do I tackle it?
Thanks in advance!

Windows 10 shared memory issue

WSL attempt

I then also attempted following instructions for Ubuntu 24.04 WSL install, again cloning fresh repo. After being forced to reduce image size to generate, found similar shared memory issues. The following I believe was done with a 768 x 576 image:

Environment

Intel Core i7-6700K 4 GHz Quad-Core Processor
MSI H110M Gaming Mobo (PCIe Gen 3.0)
2 x 16GB RAM
Intel Arc A750 8GB GPU

Brave browser
Windows 10
Python: version=3.10.6
Version: app=sd.next updated=2025-01-29 hash=46464c4c branch=master url=https://github.com/vladmandic/automatic.git/tree/master ui=main
Platform: arch=AMD64 cpu=Intel64 Family 6 Model 94 Stepping 3, GenuineIntel system=Windows release=Windows-10-10.0.19045-SP0 python=3.10.6 docker=False
Args: ['--use-ipex', '--medvram', '--debug']
IPEX: Intel OneAPI toolkit detected
Startup: standard
Extensions enabled: ['Lora', 'sd-extension-chainner', 'sd-extension-system-info', 'sd-webui-agent-scheduler', 'sdnext-modernui', 'stable-diffusion-webui-rembg']
Env flags: []
Engine: backend=Backend.DIFFUSERS compute=ipex device=xpu attention="Scaled-Dot-Product" mode=no_grad
Torch parameters: backend=ipex device=xpu config=Auto dtype=torch.bfloat16 context=no_grad nohalf=False nohalfvae=False upcast=False deterministic=False fp16=pass bf16=pass optimization="Scaled-Dot-Product"
Device: device=Intel(R) Arc(TM) A750 Graphics n=1 ipex= driver=1.6.31896
Torch: torch==2.7.0 torchvision==0.22.0.dev20250110+xpu
Packages: diffusers==0.33.0.dev0 transformers==4.47.1 accelerate==1.2.1 gradio==3.43.2

[vladmandic]

to start with default model offloading method and memory offloading settings have changed (change was documented in release notes).
try setting balanced offload low watermark to 0 in settings.

[tclakk]

This appears to have worked thanks! :)))
Just want to say I really appreciate your work and how active you are in the repo 🙏

Setting Balanced Offload GPU high watermark to 0.25 I have gotten the following result. Looks like I can probably set it a bit higher.

Model Offloading: Balanced
Balanced Offload GPU low watermark: 0
Balanced Offload GPU high watermark: 0.25

[vladmandic]

fyi. offloading is always a "balancing act" - more aggressive offloading equals more memory savings, but slows down processing a bit.
i needed to pick default values that work for most.