VRAM Estimator Extension

[space-nuko]

I made an extension that estimates VRAM usage for txt2img and img2img supporting Hires Fix as well

https://github.com/space-nuko/a1111-stable-diffusion-webui-vram-estimator

Related issue: #1730

I don't know if it will work flawlessly on other peoples' systems yet but I get pretty accurate results (within 1 GB) with my 3090. It takes into account that using non-latent upscalers with Hires Fix will use more VRAM than latent upscalers, and that img2img consumes more VRAM than txt2img.

How it works is you run a benchmark for a set of image/batch sizes on txt2img/img2img and the extension will extrapolate those results to the config options you set

Also an interesting finding to come out of this, it seems like batch size does not correlate with VRAM usage with txt2img/latent Hires Fix in an obvious way, although the growth seems consistent within the same batch size. However it does appear to correlate for img2img and non-latent Hires Fix

[kirmm3la]

I ran the benchmark on 8 Max Batch Count on 3090Ti and just kept doing it for 10 minutes, and was getting messages that cuda is out of memory. Should it go for that long? Tried on 4 Max Batch Count and was fine, took around 5 min.

[rdwulfe]

It does just seem to run forever for me as well, though it's begun showing an estimate in other windows, I can't seem to stop it.

[Mikerhinos]

This is normal it's running many tests of different resolution and batch size, keep it running it's pretty long.

[kirmm3la]