Why is this slower than before?

[sergiohzph]

I haven't used this for 1-2 months and now this is slower than before. With Euler a at 512x512 and nothing more I get around 11it/s with a 4090. I remember that I used to get around 18it/s.

A batch of 100 images used to take one minute, now it takes 3 minutes. The slowdown is most noticeable in resolutions such as 1024x512 or in landscape. Where a batch of 100 images takes half an hour, it takes quite a long half hour.

Why is it slower?

Yeah yeah I've a 4090 and it's faster enough... I didn't come for this kind of replies so save it. I just want to know why it's slower now.

PS: I've made sure that I've enabled xformers, and yeah, it's enabled.

[missionfloyd]

18 seems low for a 4090. I can get that with a 4070Ti. Have you tried --opt-sdp-attention instead of xformers?

[sergiohzph]

Seems to be faster than xformers 18-20it/s. Thanks, this is strange, xformers always was the fastest one.

[ewebgh33]

I also can't get more than around 11-13its on a 4090 despite trying many things.

From the sys info tab:
arch: AMD64
cpu: Intel64 Family 6 Model 79 Stepping 1, GenuineIntel
system: Windows
release: Windows-10-10.0.22621-SP0
python: 3.10.8
device: NVIDIA GeForce RTX 4090 (1) (compute_37) (8, 9)
cuda: 11.8
cudnn: 8700
driver: 528.02
528.02
xformers: 0.0.20
diffusers:
transformers: 4.30.2

My args are:
set COMMANDLINE_ARGS=--opt-sdp-no-mem-attention --opt-channelslast
as this seems to get me 1-2its more than --xformers.

But I've never been able to hit 18, let alone something in the 20s. And the people who are saying they can get 30+ ??? How?

I was getting 4-5its, but then did the following:

• Nvidia power options to max (Nvidia Control Panel)
• Win11 hardware accelerated GPU scheduling disabled.
  Once those options were done I got around 12 consistently. Can't get any more.

Any clues as to what else to do, appreciated!!

[sergiohzph]

@AUTOMATIC1111 Could you explain us this? Why is it being slower than before?

[ewebgh33]

Ah I see what you are saying. In my case it's "still slow". Apologies I seem to be in the wrong issue thread.

[UrdaxAlDiomis]

Same here. On linux with AMD 6900xt, a 512 x 768 image used to take a maximum of 20 seconds per image, now, no matter what I've tried, the best I can get with 512 x 512 is over 3 minutes per image. Doing the exact same image, seed, dimensions on comfyUI, I get an upscaled x4 of that image within 35 seconds.

What changed in the couple of months to degrade performance by so much?

[JoseQui77]

I haven't used this for 1-2 months and now this is slower than before. With Euler a at 512x512 and nothing more I get around 11it/s with a 4090. I remember that I used to get around 18it/s.

A batch of 100 images used to take one minute, now it takes 3 minutes. The slowdown is most noticeable in resolutions such as 1024x512 or in landscape. Where a batch of 100 images takes half an hour, it takes quite a long half hour.

Why is it slower?

Yeah yeah I've a 4090 and it's faster enough... I didn't come for this kind of replies so save it. I just want to know why it's slower now.

PS: I've made sure that I've enabled xformers, and yeah, it's enabled.

I have tried to upgrade to the latest version several times and up to the last time I tried few weeks ago is always slower, so for now I'm stuck and go back to the version where it performed faster for me which is the following.

version: v1.5.1  •  python: 3.10.11  •  torch: 2.0.1+cu118  •  xformers: N/A  •  gradio: 3.32.0  •

git checkout 68f336b

commit 68f336b (HEAD, tag: v1.5.1, origin/multiple_loaded_models)
Merge: a3ddf46 50973ec
Author: AUTOMATIC1111 [email protected]
Date: Thu Jul 27 09:02:22 2023 +0300

Merge branch 'release_candidate'

Also, I am using the following driver, newer drivers seem to be slower too.

+---------------------------------------------------------------------------------------+
| NVIDIA-SMI 531.68 Driver Version: 531.68 CUDA Version: 12.1 |
|-----------------------------------------+----------------------+----------------------+
| GPU Name TCC/WDDM | Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|=========================================+======================+======================|
| 0 NVIDIA GeForce GTX 1660 Ti WDDM | 00000000:01:00.0 On | N/A |
| N/A 90C P0 78W / N/A| 5532MiB / 6144MiB | 100% Default |
| | | N/A |
+-----------------------------------------+----------------------+----------------------+

I think something was changed between the following versions that caused the performance drop:

v1.5.2
on Aug 23 c9c8485 zip tar.gz Notes
v1.5.2-RC
on Aug 23 31f2be3 zip tar.gz

This seems to be the last fast one (at least for me):

v1.5.1
on Jul 27 68f336b zip tar.gz Notes

Check if that version works as it used to do for you.