Dec 2023 - can someone please advise me on the latest optimal setup for RTX3090

[TheWallop]

I would be so grateful for help regarding my points of confusion.

Two queries:

1. All currently works fairly well but is it optimal?

My specifics: RTX3090 card, Intel 12 Gen i5, 64Gb RAM, A1111, W10 22H2

My problem: there is so much legacy advice out there on what cudnn to install, what pytorch vsn to use, whether to still use xformers on older setups vs opt-sdp-attention being potentially better (although VRAM hungry), browser h/w accel on vs off etc

If I was to start from scratch again this December in 2023, with a general requirement for say 1024 x 1024 type output, what should I do comprehensively please? Let us assume I have started with a vanilla w10/11 PC - so not even nVidia is installed let alone A1111.

I would be so grateful for a (((detailed step-by-step))) ;) on the speediest setup I could use with TensorRT Extension running the gens.

2. Entirely separate (or possibly part of the problem) is the 4 x procedure call errors I always get on any install of A1111:
   

Is there a reason? And is there a properly working fix please? It doesn't seem to stop A1111 from working for my needs.

---

Thanks in advance for any help. The area I most want information is the optimally correct way to set up A1111 with an RTX3090 at this current time.

[TheWallop]

Have found a solution to part 2 at least: https://www.youtube.com/watch?v=ssPhVOgd1Qc

If you also follow this, I chose to install tensorrt==9.0.1.post12.dev4 which worked correctly versus the version instructed.

[missionfloyd]

Just follow the instructions. You don't need to mess with cudnn anymore.

The TensorRT extension can be found on the Extensions > Available tab.

[w-e-w]

irrc this NVIDIA/Stable-Diffusion-WebUI-TensorRT#12 (comment)

[TheWallop]

All good now thanks. Since following the youtube method after already having followed
https://github.com/AUTOMATIC1111/stable-diffusion-webui?tab=readme-ov-file#automatic-installation-on-windows , my performance is excellent again and with no errors at any point.

It will no doubt go flaky again sometime, with some upgrade of this or that, but the main thing is I have gained a better understanding of what affects what.

Performance changed from 10-15 It/s to around 28 - 35 It/s so cannot grumble for an RTX3090 :)