New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
17.0 Server execution error: expected query, key and value to have the same dtype (was #345) with face (worked in 16.1) and style transfer #635
Comments
Reference, Face, Style, Composition all use IP-Adapter (https://github.com/cubiq/ComfyUI_IPAdapter_plus) I think this issue has been reported multiple times, but can easily crop up again. It's probably not easy to test or fix because it's hardware-specific, seems to happen on the 10xx GPUs. Latest discussion here: cubiq/ComfyUI_IPAdapter_plus#108 (comment) |
Edit: Got it to work as well for SDXL Models with "server_arguments": "--dont-upcast-attention --force-fp16", Edit: As it seems, some systems have even more issus on 30xx than on 10xx, Link: https://www.youtube.com/watch?v=Ngt5Oqa7aqE&lc=Ugx0xb5uhdnqqbVi2z94AaABAg.A2gCz2uY35NA2gJLWifDEH Edit: Here i s al link to same (?) issue with automatic1111 and rtx20xx series, though i cant see if ip adapter plus is involved. Thanks for the link. Everthing is working with 1.5 models, i found out by your hint. |
So what is planned? According to other comments, esp the youtube comments on streamtabulous' post it not only a 10xx issue. In the opposite, esp. 30xx has more issues and of more kinds. |
I ran all tests on a 3060, and did not encounter issues. Reference/Style/Composition/Face are working. As for the 10xx cards, someone with the hardware needs to reproduce this directly in ComfyUI with a simple example workflow from the IP-Adapter nodes and report the full error there. I'm not familiar enough with that code to blindly guess where the issue might be. If it worked before as you say it would also help to test different version of IP-Adapter nodes to find out when the issue was introduced. |
Is the ip adaper involved in "fill" and did it change in 1.17 or 1,17. 1? I have the dtype error message first ime ever with fill this evening - für 1.5 and xl models. Getting annoying. Looks like i have to roll back to 1.16. |
It's used for fill/expand only if you don't provide a prompt (as a replacement of sorts) |
To says it clearly: this is a new issue. All my dtypes error so far were concerning ip face adapter V2 (SDXL).
plus: when changing sever arguments to "server_arguments": "--dont-upcast-attention --force-fp16" the json file opens automatically with a line "server_arguments": "null" beore it. Is this correct? |
Need to see the full error trace from the server.log to say for sure, but since it's related to reference/face/style it is likely that it's an issue with IP-Adapter. And yes that has been updated. Style transfer is one of the new features.
Don't really understand that part, |
"null" is a copy and pace error, sry for that. It inserts a line with null, not with "null". Here my server log from today, first use of fill feature with "server_arguments": "--dont-upcast-attention --force-fp16", second without. 2024-04-30 07:33:54,132 INFO Starting server Hardware / OS: GTX 1070, X470 Chipset, Ryzen 7 2700x, Win11 Just for general info, not due to actual intentions: Is it possible to have the new and the oold adapter in one directory and chose manually which one to work with? Or are there some config settings that oblige to use the new one with 1.17.1? |
They are different models (those were not changed) but are both handled by the same extension nodes and share a lot of code. So it's not possible to update them independently.
It's possible to some extent. But a while ago IP-adapter nodes significantly changed their interface, and version 1.16+ of the plugin is not compatible with the old one. |
Hi Acly! |
It just confirmed that it's a disagreement between IP-adapter and the base ComfyUI sampler. |
sounds like a bug, question is on which level |
edit: same with reference, randomly working or not. Always worked in previous versions.
on GTX 1070: I had the Error (see refrence line) in 16.1 with face adapter. First i solveed it with forced FP16, then i discovered, it works without forced fp when GPU settings by switching from low to medium cuda. (Was on low following some youtube tips.)
In 17. 0 none of both ways work anymore, neither for face adapter nor for the new style transfer. Having worked face before seems to indicate a solvable bug - and lets me hope that it can repair style transfer too. (Was very hot to try and very dissapointed now.)
Nevertheless great work, acly, and incredible speed of development!
(btw. where can i see,wwhich samplers are defactav ailable?)
thx.
The text was updated successfully, but these errors were encountered: