[AMD/LINUX] Slow generation after boot, but improves drastically by launching a game. (idk what to title this oddity) #732
Replies: 8 comments 15 replies
-
that is really curious, but lets step a bit? Since you're talking about GPU utilization and AMD GPU, I assume this uses ROCm and since you have A1111 and this repo both installed on the same system, its the same ROCm? Can you confirm version? second, since each repo is running in its own VENV, they probably use private installs of Torch itself, so lets check torch versions for each? what are the command line flags you're using for each repo? |
Beta Was this translation helpful? Give feedback.
-
only differences are that you're starting a1111 with |
Beta Was this translation helpful? Give feedback.
-
I can confirm my 6700xt has dropped in speed dramatically, I went from around 7.6 it/s to anywhere between 6-10s/it, rolling back to commit 5148c5b returns to full speed, though that was just the commit I was on before I updated this morning. Here are my cuda settings, in addition I'm using the --listen cmd arg. I can grab more details when I go on lunch here in an hour or so. "precision": "Autocast", |
Beta Was this translation helpful? Give feedback.
-
good to know last confirmed good commit hash, i'll check the differences. |
Beta Was this translation helpful? Give feedback.
-
I had experienced a similar oddity when I reinstall the repo from the ground up. Half of the time, my generations are slow at 0.5-1 it/s. The other half of the time, I’m at full speed of 6.5it/s. I have no idea what causes it, no differences at all between the installs, same settings and arguments. I will note, I haven’t tried updating in a week so I’m not sure if this has been fixed or is a related ongoing issue. |
Beta Was this translation helpful? Give feedback.
-
Just chiming in to report the same issue (6700XT). I did a clean install and rolled back to 5148c5b, but unfortunately that didn't work for me; 512x512 images still took 8 to 10 seconds per step. I also tried playing two games (one via Proton, one natively), but had no success. What did work for me was loading up Automatic1111 and generating a few images there; then I was able to generate images at normal speed with this UI. I then went back to the latest commit, and still have been able to generate images quickly. Edit: For the avoidance of doubt, I should clarify that I was able to generate images at normal speed with this UI after generating with A1111 for a bit. I was able to replicate this today after a reboot; I was generating at 4.5 s/it on this UI, then generated 24 images with A1111 at normal speed, 6.3 it/s, over a span of about 6 minutes. Then after that, I immediately got normal speeds again on this UI, 7.2 it/s. |
Beta Was this translation helpful? Give feedback.
-
I had this same problem but somehow it magically resolved itself. One thingI remeber on linux though is if you use CoreCTRL and set the GPU to power saving mode, the performance will drastically reduce in a similar way. Maybe something is triggering that? (I don't have access to my main pc at the moment for a while) |
Beta Was this translation helpful? Give feedback.
-
So, a lot of these comments seem like answers to a different question. Regardless, I might have been wrong about the game part, in the sense that its not the game, but the time. |
Beta Was this translation helpful? Give feedback.
-
I have a puzzling problem, that while surmountable, is quite annoying and Im not sure where to find answers to it, so I started here. (GPU:6800XT, Fedora37)
After booting up the PC and starting the webui, generation is really slow, it uses GPU but even at default(512x unipc:10 or eulera) the console writes ~0.55it/s or ~1.5s/it. Hardware monitor shows really low and intermittent usage/activity of the GPU(bright green line should be straight maxed out at the top like the pinky line):
Other UIs work directly after boot, even if slower than what I can get with this version, but Im also using sub quadratic or opt-split there. ComfyUI and a1111 both get around 4-5it/s(512x eulera:20). (a1111 and ComfyUI have their own venv).
There are a few ways I can gain speed with this UI, I think just loading a model with Blender worked, but that didnt reliably work, what always works is to start a game in via Steam. I havent tested around a lot, but launching and running a game with proton, while the UI is still active(though not generating), works all the time.
Once the game ran for a few minutes, I can turn it off and use this UI at 8-9it/s at default(512x eulera:20).
Once its running fast, nothing else influences it, as long as I dont restart the PC. I thought/think it might be a driver issue, but as neither a1111 or ComfyUI have any problems directly after boot, I think it might be more of a problem with this UI.
Im willing to thoroughly test this more, but I dont know where to start, so if anyone has an idea or the same Issue Ill try it out.
Beta Was this translation helpful? Give feedback.
All reactions