Can not load shared library: stable-diffusion_hipblas.dll #29
Comments
|
Hi there! |
|
In-app console: stablediffusiongui_diffuser.log: [edit] |
As i see, it is logged into multiple places, so that's can cause duplicates. This path 'D:\AI\stablediffusion\sdgui' it was set in the installer itself when you installed? Or you moved the installed folder into this new place? Please try to reinstall and use the default path what is in the installer. I don't think this is the problem, but it's worth a try. Please check out the task manager, if you have a running stablediffusion* process. If yes, please stop it all. Then please start the GUI with the following parameter: This will disable the external process handling, so the process will not start. You need to start manually with the following cmd:
May escaping required here. If you started, then please check out the output (if have) and the stablediffusiongui_diffuser.log. And there is an app.log in the "Data folder" |
|
The path was set in installer, I keep most software out of C drive. There was no process in task manager. I've tried running with the cmd Output: The app.log in the appdata (i missed that location) contains nothing more than what I already pasted: I will try to reinstall into default location shortly and let you know if that changes anything. [update] [update2] So maybe the problem is not that the _hipblas.dll does not load but that it crashes on init somehow? But my GPU should be able to run with HIP and ROCm. Koboldcpp runs fine with it using rocm build, offloading the computation to GPU. |
|
It's weird. Not the GPU the problem. The original error is comming from here, which means, the lib can not be loaded because something missing for it. (maybe some runtime lib missing on win10 what is available on 11, but i have no proof if it can run on win11 too :( ) So, today i want to release a new version, where i replaced hipblas with vulkan binaries. Maybe the next release will work for you. |
I think I've found something. I've investigated, using the ldd command from Git Bash and the output is enlightening: So rocblas.dll and hipblas.dll are missing. But I installed HIP SDK from AMD, do I need something more? Not sure what provides those. [update] I've googled and CUBLAS_STATUS_INVALID_VALUE shows up when there is a version mismatch apparently, but I am way out of my depth here. Why is it even CUDA, shouldn't it use ROCm? In any case, thank you for your help so far. I really like what you are building here. Any ideas how to solve the latest issue? Did I install wrong HIP SDK version? |
|
Nice catch! As i wrote earlier i cant test ROCM and HIPBLAS builds, but what i know is the ROCM version which is 5.5.0. At least the _hipbals.dll is built with this version. |
|
That was it. Installed HIP SDK 5.5.1, updated path, and now it works like a charm. Maybe a Wiki or readme entry for the HIP SDK version requirement would be nice to add. In any case, thanks for your support, now it works wonderfully! <3 |
Now in the new release there is no more ROCM / HIP. |
Hi.
1st of all: very promising project. Not having to deal with npm to use SD will be a godsend.
I do seem to have an issue with generating anything tho. When queueing a job, I immediatly get "Generation error: External process stopped". The diffuser.log has "[EXTPROCESS] Can not load shared library: D:\AI\stablediffusion\sdgui\stable-diffusion_hipblas.dll" message in it but not any more details.
The dll is present at that path. Tried running as admin, no change. I have Radeon RX6800. Running Win10.
Not sure what might be the issue.
[edit]
I've also tried running the GUI with -avx512, it starts generation and is able to do the "model hash" step, but then fails as well with "Generation error: External process stopped." Error, however the diffuser.log is empty in that case.
The text was updated successfully, but these errors were encountered: