RuntimeError: CUDA error: invalid configuration argument

[tin2tin]

RuntimeError: CUDA error: invalid configuration argument

[tin2tin]

Hmm... automatic1111 users are also affected: AUTOMATIC1111/stable-diffusion-webui#12582

[tin2tin]

Yes, I'm also see that error right now. A wild guess, maybe it is the update of Diffusers to 0.20 which caused it. It was updated yesterday. I'll try to downgrade the Diffusers version.

[mischaschaub]

Wow, what instant reaction! Many thanks and hoping for the best.

[tin2tin]

Downgrading diffusers didn't seem to work. So it might be the Torch version.

[tin2tin]

What is your gfx driver version?
I'm on 531 bc later versions automatically moved a lot of the AI processing over on the CPU and that caused serious slowdowns.
Maybe upgrading those NVIDIA drivers will solve it. You, the reporter on auto1111 and I all got RTX cards - maybe there is something there.
Now I'm trying to update pytouch to the nightly to see if that is making any difference.

[mischaschaub]

I am using version 536.40 Thanks for all your efforts! Am Sa., 19. Aug. 2023 um 16:53 Uhr schrieb tin2tin ***@***.***

…

: What is your gfx driver version? I'm on 531 bc later versions automatically moved a lot of the AI processing over on the CPU and that caused serious slowdowns. Maybe upgrading those NVIDIA drivers will solve it. You, the reporter on auto1111 and I all got RTX cards - maybe there is something there. Now I'm trying to update pytouch to the nightly to see if that is making any difference. — Reply to this email directly, view it on GitHub <#38 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AMF56BODDSU6VYDQXJYATUTXWDHNJANCNFSM6AAAAAA3WR2CNE> . You are receiving this because you authored the thread.Message ID: ***@***.***>

--

[tin2tin]

In the add-on preferences, I clicked Uninstall Dependencies. Let it do it's job. When finished, I closed Blender. Restarted it(as admin) and installed all dependencies again. And this worked. I still have the code for installing the previous version(0.19.3) of Diffusers active, tho. Not sure if that made a difference.

[mischaschaub]

Thanks for both ideas, will try them tomorrow first in the morning! Am Sa., 19. Aug. 2023 um 19:49 Uhr schrieb tin2tin ***@***.***

…

: In the add-on preferences, I clicked Uninstall Dependencies. Let it do it's job. When finished, I closed Blender. Restarted it(as admin) and installed all dependencies again. And this worked. I still have the code for installing the previous version(0.19.3) of Diffusers active, tho. Not sure if that made a difference. — Reply to this email directly, view it on GitHub <#38 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AMF56BO2FV7ZZ6IBF2RYJXLXWD4A3ANCNFSM6AAAAAA3WR2CNE> . You are receiving this because you authored the thread.Message ID: ***@***.***>

--

[tin2tin]

I updated it with some changes which might address those problems.
I made memory enhancements kick in at 8 GB including this line:
upscale.unet.enable_forward_chunking(chunk_size=1, dim=1)
If you got plenty of VRAM and feel everything has gone too slow, then comment out that line from everywhere is occurring.
Let me know how it goes.

[mischaschaub]

Not yet there. Tried now for an hour all combinations of restarts, even a new blender from scratch, then I installed the latest version from Github of your add-on and this finished OK, but when I try to run it I get this error: Read prefs: "C:\Users\mischa\AppData\Roaming\Blender Foundation\Blender\3.6\config\userpref.blend"
register_class(...):
Warning: 'NODE_MT_group_Geometry_/Math' doesn't have an alpha-numeric suffix
reloading addon: Generative_AI-main 1692512799.8565087 1692514277.5294695 'C:\Users\mischa\AppData\Roaming\Blender Foundation\Blender\3.6\scripts\addons\Generative_AI-main\init.py'
Modules Installed () from 'C:\Users\mischa\Downloads\Generative_AI-main (2).zip' into 'C:\Users\mischa\AppData\Roaming\Blender Foundation\Blender\3.6\scripts\addons'
vae\diffusion_pytorch_model.safetensors not found
Loading pipeline components...: 100%|████████████████████████████████████████████████████| 5/5 [00:02<00:00, 1.70it/s]
Generating Video
Traceback (most recent call last):
File "C:\Users\mischa\AppData\Roaming\Blender Foundation\Blender\3.6\scripts\addons\Generative_AI-main_init.py", line 1302, in execute
video_frames = pipe(
File "C:\Users\mischa\AppData\Roaming\Python\Python310\site-packages\torch\utils_contextlib.py", line 115, in decorate_context
return func(*args, **kwargs)
TypeError: VideoToVideoSDPipeline.call() got an unexpected keyword argument 'height'
Error: Python: Traceback (most recent call last):
File "C:\Users\mischa\AppData\Roaming\Blender Foundation\Blender\3.6\scripts\addons\Generative_AI-main_init_.py", line 1302, in execute
video_frames = pipe(
File "C:\Users\mischa\AppData\Roaming\Python\Python310\site-packages\torch\utils_contextlib.py", line 115, in decorate_context
return func(*args, **kwargs)
TypeError: VideoToVideoSDPipeline.call() got an unexpected keyword argument 'height'

[mischaschaub]

Getting stranger by the minute; get now always this error when I try to install your add-on. I even created a brand new windows-user on my second PC, installed blender and get this error. So what can I do - would love so much to use your app which would solve many of my problems.

[tin2tin]

In the bottom line of your screenshot is a path. Goto the side-packages folder. This is where all of the python modules are located. With Blender closed, delete all folders with dates corresponding with your installation of the add-on. Then start Blender(as admin), install the add-on(grab the latest from 10 hours ago). Hopefully that will install the add-on okay. Then click the install dependencies. I hope this should work.

If you see the height error again, send me a screenshot of the selected models(add-on preferences) and the sidebar options, so I can try to reproduce it.

[tin2tin]

Found the reason for the "height" error.

[mischaschaub]

Great. So I wait for the next update, I guess? Am So., 20. Aug. 2023 um 02:09 Uhr schrieb tin2tin ***@***.***

…

: Found the reason for the "height" error. — Reply to this email directly, view it on GitHub <#38 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AMF56BN5N4G2ZUCILGCRJVLXWHH6LANCNFSM6AAAAAA3WR2CNE> . You are receiving this because you authored the thread.Message ID: ***@***.***>

--

[mischaschaub]

In the meantime I finished the new installation, which worked flawlessly. But on the first run I get the following error:

......
Info: Installation of dependencies is finished.

vae\diffusion_pytorch_model.safetensors not found
Traceback (most recent call last):
File "C:\Users\mischa\AppData\Roaming\Blender Foundation\Blender\3.6\scripts\addons\Generative_AI-main_init_.py", line 1144, in execute
pipe = DiffusionPipeline.from_pretrained(
File "C:\Users\mischa\AppData\Roaming\Python\Python310\site-packages\diffusers\pipelines\pipeline_utils.py", line 908, in from_pretrained
cached_folder = cls.download(
File "C:\Users\mischa\AppData\Roaming\Python\Python310\site-packages\diffusers\pipelines\pipeline_utils.py", line 1446, in download
use_onnx = use_onnx if use_onnx is not None else pipeline_class.is_onnx
File "C:\Users\mischa\AppData\Roaming\Python\Python310\site-packages\diffusers\utils\import_utils.py", line 575, in getattr
return super().getattr(cls, key)
AttributeError: 'super' object has no attribute 'getattr'. Did you mean: 'setattr'?
Error: Python: Traceback (most recent call last):
File "C:\Users\mischa\AppData\Roaming\Blender Foundation\Blender\3.6\scripts\addons\Generative_AI-main_init.py", line 1144, in execute
pipe = DiffusionPipeline.from_pretrained(
File "C:\Users\mischa\AppData\Roaming\Python\Python310\site-packages\diffusers\pipelines\pipeline_utils.py", line 908, in from_pretrained
cached_folder = cls.download(
File "C:\Users\mischa\AppData\Roaming\Python\Python310\site-packages\diffusers\pipelines\pipeline_utils.py", line 1446, in download
use_onnx = use_onnx if use_onnx is not None else pipeline_class._is_onnx
File "C:\Users\mischa\AppData\Roaming\Python\Python310\site-packages\diffusers\utils\import_utils.py", line 575, in getattr
return super().getattr(cls, key)
AttributeError: 'super' object has no attribute 'getattr'. Did you mean: 'setattr'?

[tin2tin]

Just updated the code with (hopefully) a fix for the height thing. As all the models are passing through the same code, I hope this does not break something else, but I do not have time to check up all of the different variations.

No clue about the latest error. Looks like it is related to diffusers. Let me know if you still see this after the update.

[tin2tin]

Could be related to some problem downloading the models. If you need to clean out the models, they're located here:
C:\Users\YourName\.cache\huggingface\hub

Btw. if you get the cuda out of memory error, things might get wonky(Diffusers can crash Blender), so I tend to restart Blender after those errors(which you encounter when trying to see how far you can go in ex. number of frames)

[mischaschaub]

Great, this seems to work now. I installed blender and your add-on with a new admin-account, and then your latest version did install flawlessly. But at the moment I just get some kind of black and white dotty universe, rather nice, but I think this might be SDXL.ckpt missing in my fresh install. So far I just have used SD within VOC, so my knowledge level is rather basic. So do I have to install the SD-blender-add-on from github as well to get your tool running?

I have another question: Could I use a SDXL.cktp of my own, or a Lora? I got 700 historical photos of a town square in my hometown, and I would like to set my movies in this environment, if feasible. What do you think?

Just for your info of what I just did with zeroscope XL so far under VOC, and what I would like to continue doing with your add-on: https://vimeo.com/856147115?share=copy

with supermany thanx for your speedy support

[tin2tin]

If it is producing images, it is working. Great! if it produces some abstract patterns, often it means you should try to change the number of frames and/or the resolution. As a rule of thumb, try to start with the resolution stated in the model selector and +1 on the frames. There is a lot of trial and error in tweaking your way into settings working for you. cktps are not (yet?) supported. As mentioned is this add-on utilizing the Diffusers python module: https://github.com/huggingface/diffusers , and they got their own model format and the models needs to be uploaded to hugging face: https://huggingface.co/cerspense/zeroscope_v2_XL/tree/main

So, this is what you should you with your model, and then from line 629 in the script, you can see the way to add additional models pointing to the location on Hugging Face.

If you need to convert a model to this format, then this project should be used(as I don't have the hardware for training I never looked into this): https://github.com/ExponentialML/Text-To-Video-Finetuning

I haven't read up on loras yet, but I think they're supported by Diffusers.

Wow, very fascinating video. Totally satisfying to look at.

What is VOC?

[mischaschaub]

VOC=voices of chaos, my AI-paradise. Go to softoloy.pro and have a look - this guy is just incredible. I act as one of his patreons, but you get everything for free as well. There is a fantastic discord group. So I misunderstood completely, I guess, as I could choose a ckpt model in the setup menu of the add-on. Now that I use the standard resolution of Zeroscope XL of 576 x 320 everything renders out perfectly. One more question: I do not understand the seed function, as I cannot choose it, if I understand this correctly. But I would exactly love to choose seed myself - for instance, to repeat a rendering with a slightly different prompt, or do do it again in a higher resolution. Is this not feasible? Am So., 20. Aug. 2023 um 16:33 Uhr schrieb tin2tin ***@***.***

…

: If it is producing images, it is working. Great! if it produces some abstract patterns, often it means you should try to change the number of frames and/or the resolution. As a rule of thumb, try to start with the resolution stated in the model selector and +1 on the frames. There is a lot of trial and error in tweaking your way into settings working for you. cktps are not (yet?) supported. As mentioned is this add-on utilizing the Diffusers python module: https://github.com/huggingface/diffusers , and they got their own model format and the models needs to be uploaded to hugging face: https://huggingface.co/cerspense/zeroscope_v2_XL/tree/main So, this is what you should you with your model, and then from line 629 in the script, you can see the way to add additional models pointing to the location on Hugging Face. If you need to convert a model to this format, then this project should be used(as I don't have the hardware for training I never looked into this): https://github.com/ExponentialML/Text-To-Video-Finetuning I haven't read up on loras yet, but I think they're supported by Diffusers. Wow, very fascinating video. Totally satisfying to look at. What is VOC? — Reply to this email directly, view it on GitHub <#38 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AMF56BIK6WBWIEC4W3Y5M23XWIN23ANCNFSM6AAAAAA3WR2CNE> . You are receiving this because you authored the thread.Message ID: ***@***.***>

--

[tin2tin]

Thanks for the hint on VOC. The question mark is a toggle button to randomize the seed. When depressed you can edit the seed value.

However, if you use generated strips for img2vid or vid2vid, the seed value is saved in the filename and automatically inserted i the seed value while rendering. If you need at stronger variation, then lower the Strip Power value.

As it is now you can only select between the exposed models in the UI(but as mentioned, it is in most cases a trivial code thing thing to add more)

[tin2tin]

Closing, as this seems to be solved.