This is driving me nuts | Invoke | Page 1

mossy fossil Feb 4, 2023, 9:48 AM

#

WARNING[XFORMERS]: xFormers can't load C++/CUDA extensions. xFormers was built for:
PyTorch 1.13.1+cu117 with CUDA 1107 (you have 1.13.1+cpu)
Python 3.10.9 (you have 3.10.9)
Please reinstall xformers (see https://github.com/facebookresearch/xformers#installing-xformers)
Memory-efficient attention, SwiGLU, sparse and more won't be available.
Set XFORMERS_MORE_DETAILS=1 for more details
u:\ai\invokeai.venv\lib\site-packages\torchvision\io\image.py:13: UserWarning: Failed to load image Python extension: [WinError 127] The specified procedure could not be found
warn(f"Failed to load image Python extension: {e}")

More and more errors after I followed the instructions. I am not a python person, or a pytorch person, or a conda person, or a xformers person......Man!

mossy fossil Feb 4, 2023, 10:25 AM

#

Now what?

lilac chasm Feb 4, 2023, 11:18 PM

#

This should be fixed by installing 2.3.0-rc4. Up on the releases page as of a couple of hours ago.

mossy fossil Feb 5, 2023, 7:07 AM

#

Hi! Thank you for your reply. I did an update using "update.bat" it actually downgraded my pytorch and others. (I wish I had grabbed the screen) and still in 2.3.0-a0. So I simply downloaded the 2.3.0-rc4. But there is no install file. Do I just copy and overwrite everything in the original InvokeAI folder? Thank you.

lilac chasm Feb 7, 2023, 5:53 PM

#

Download https://github.com/invoke-ai/InvokeAI/files/10671597/InvokeAI-installer-v2.3.0-rc6.zip, unzip, run install.bat and when it asks you whether it is OK to reuse the existing invokeai directory say "yes".

mossy fossil Feb 7, 2023, 11:14 PM

#

Hello! Yes I did earlier when I learned that rc6 was out. Very nice indeed! Solved my xformer issue, implemented the Negative prompts box. Could be even better if we could collapse the box entirely to save space because I believe we don't really look at it after a while. LOL. But one disappointment for me is that, it renders passes 17 images where it would crash with "CUDA out of memory" problem, got me excited, but crashed at the 25th image. Oh well, It's an improvement. Also the loading of a 7gb model took it nearly 800++ sec. wow.......I thought it hanged.

lilac chasm Feb 7, 2023, 11:30 PM

#

That’s not good. How much RAM and VRAM have you got? Do you see the VRAM usage climbing?

#

Try putting —free_gpu_mem into invokeai.init

#

@sage quartz this looks like another memory issue. Do you have insight ? Seems rather extreme.

sage quartz Feb 7, 2023, 11:34 PM

#

lilac chasm <@983345600911249418> this looks like another memory issue. Do you have insight...

I can't think of anything off the top of my head... if xformers is being used, then it's bypassing the diffusers attention code entirely. @mossy fossil Can you try running with --no-xformers and see if it gets you past that step? And also share your machine & GPU specs.

lilac chasm Feb 8, 2023, 1:23 AM

#

The other thing to try is to put —max_loaded_models =1 into the init file. This prevents memory caching

sage quartz Feb 8, 2023, 2:20 AM

#

(without a space)

mossy fossil Feb 8, 2023, 7:25 AM

#

Hi! Thanks for all the replies!! My system is a Ryzen 5 2600, 16GB RAM, GTX1070 8GB VRAM. In the lower-end category. I entered the magic keywords into invokeai.init, ordered 40 images of 512X768, 30 steps each on DDIM. Undisturbed. It crashed at the 24th image. LOL!
with this:

torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 78.00 MiB (GPU 0; 8.00 GiB total capacity; 7.08 GiB already allocated; 0 bytes free; 7.33 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

I am able render 300 images with another diffusor (Easy Diffusion) so I guess my system is....well... still alright, I think? LOL.

Anyway I still love InvokeAI. LOL!

sage quartz Feb 8, 2023, 11:38 AM

#

Can you try the latest release candidate?

mossy fossil Feb 8, 2023, 3:22 PM

#

Hi! I am already running RC-6, so is there a RC-7? Where is it? I lost the link to the latest RC releases. LOL. Thanks!!!

sage quartz Feb 8, 2023, 5:08 PM

#

Oh, not yet... I'm not sure if the fix is a fix but the final RC or 2.3.0 will have the latest slicing and memory freeing that may help.

#

If it doesn't after 2.3.0 is out, please file a bug report so we can investigate.

mossy fossil Feb 11, 2023, 10:34 PM

#

Hello! I have updated to the latest 2.3.0 (official release) (12-Feb2023).
<>First launch: Crashed the system, had to reboot.
<>Second launch: Unable to load a 7.5gb model - CUDA out of memory (took a long time)
<>Third launch : loaded a 2gb model - batch generation 30 images (512X768) 25 steps - Crashed at the 23rd Image - CUDA out of memory. As usual...

Oh well.....LOL.

sage quartz Feb 12, 2023, 2:15 AM

#

You can't fit a 7.5GB model plus all of what InvokeAI needs to run in VRAM.

#

Are you running with --precision float16 as a command-line argument (or in invokeai.init)?

#This is driving me nuts