Closed Harbitos closed 4 days ago
You fixed gguf problem but it looks like you haven't installed HIP SDK correctly.
Its needed if you want to use --zluda
and don't use --zluda and --directml at the same time in your commands
correct commands are --zluda
or --directml
without "use"
If you want to use ZLUDA you have to remember your CPU must support PCIe™ atomics. For RX580 you probably need build your own RocblasLibs. Here is the link with whole guide: https://github.com/vladmandic/automatic/wiki/ZLUDA
BTW: If you don't want to do this then you can just use --directml
command
Steps to fix: Go into venv/scripts
open CMD
. Type activate
then pip install gguf
. Run.bat
If you want to use ZLUDA you have to remember your CPU must support PCIe™ atomics. For RX580 you probably need build your own RocblasLibs. Here is the link with whole guide: https://github.com/vladmandic/automatic/wiki/ZLUDA
BTW: If you don't want to do this then you can just use
--directml
command
SD started working with --directml, generated a couple of images, but also gave errors a couple of times: the first - nothing was written, restarting SD helped, the second - resource deadlock would occur:resource deadlock would occur, restarting helped too.
And which is better, ZLUDA or --derectml? A friend told me that directml is more stable than ZLUDA. If ZLUDA is better, then please explain as simply as possible how to install it, otherwise at one point it is difficult for me to create RocblasLibs libraries.
As far as i know ZLUDA is faster and better but i can't test it myself. You have everything in the link i paste how to make ZLUDA work. There is nothing more to it you just need CPU supporting PCIe™ atomics.
This is the whole guide:
RX580 is not on the list so you need to build your own RocblasLibs BUT i saw once somewhere ppl already made libs for RX580 so if you will look for it in google you should be able to find
EDIT: If you will stay with --directml
i will also suggest you to use --no-half-vae --no-half
this commands
Steps to fix: Go into
venv/scripts
openCMD
. Typeactivate
thenpip install gguf
. Run.bat
He fixed it. I believe the method you mentioned doesn't work because Forge has its own integrated GGUF module but there is bug when its not visible. Downloading GGUF with pip install gguf
will download wrong version
As far as i know ZLUDA is faster and better but i can't test it myself. You have everything in the link i paste how to make ZLUDA work. There is nothing more to it you just need CPU supporting PCIe™ atomics.
This is the whole guide:
RX580 is not on the list so you need to build your own RocblasLibs BUT i saw once somewhere ppl already made libs for RX580 so if you will look for it in google you should be able to find
Drop the link, I can't find it.
Drop the link, I can't find it.
Link to what ? To the guide ? I don't want to be mean to you or anything but if you can't find the link from few comments above which you even quoted then just give up and use DirectML :D
Unless you talk about RocblasLibs for RX580 then i don't have a link
Steps to fix: Go into
venv/scripts
openCMD
. Typeactivate
thenpip install gguf
. Run.bat
i had the ModuleNotFoundError: No module named 'gguf'
error. i edited launch.py as in the original comment, then i did what this comment said, and that fixed it
I have a different situation, windows10, amd RX580 graphics card, Intel Xeon processor, the latest version of Git and Python 3.10.6. The initial installation of stable-diffusion-webui-amdgpu-forge by lshqqytiger returns an error:
If I added in launch.py
– gives me the following error:
Even when adding in launch.py:
– leads to the same error as mentioned above about NVIDIA.
Adding
--use-directml
--use-zluda
to COMMANDLINE_ARGS results in the same error with NVIDIA. I read most of the issues on the Internet about gguf or NVIDIA, but I didn't find any useful information. Deleting the "venv" folder did not help. Reinstalled several times, installed viagit clone https://github.com/lshqqytiger/stable-diffusion-webui-amdgpu-forge.git
Please help me fix this. I don't know much about programming, so if you need to add some code, write on which line and after that. Screenshots are attached launch.py , in webui-user, nothing is written in COMMANDLINE_ARGS.