Need help with AMD GPU

[ElPenguin]

OS : Rocky Linux 9.6
CPU : AMD R7 3800x
GPU : AMD RX6600
RAM : 32Go

Hey !

I'm trying to run the app on leftover hardware but I can't.

I installed the application, I launched the rocm update script, I have the AMD drivers installed on my system. The problem is that every time I run a test with just a simple word for image generation, I have this error :
./webui.sh: line 45: 1843 Segmentation fault (core dumped) HF_HUB_DISABLE_XET=1 AUDIOCRAFT_CACHE_DIR='./models/Audiocraft/' python3 webui.py

Can someone help me?

[Woolverine94]

Hello @ElPenguin,

Thanks for your feedback and your interest in biniou.

Can you confirm that you are using ROCm 5.6 as specified here ?

There's nearly zero chances you can benefit from AMD GPU acceleration if you're not using this specific version of ROCm, as biniou is using a specific Pytorch version compiled for it.

[ElPenguin]

Hey !

I can't install ROCm 5.6.
I try on Rocky 9.6/8.10, Ubuntu 22/24 and there is always a system compatibility issue.
I will try with older versions and Debian 12..

[Woolverine94]

Hello @ElPenguin,

Thank you very much for the detailed feedback !

It makes sense that everything failed except the Chatbot, as mentioned above.

For the record, previous link to install ROCm 5.6 on ubuntu was for 7900xtx and 6900xt, i don't know if it's my 6600 the problem and I try on a laptop with rtx2070 and it work

The Compatibility matrix for ROCm doesn't show versions prior to 6.0, so it's difficult to say if your hardware is compatible. Nevertheless, and if this could help, I've found this post on reddit that makes me realize that the installation procedure you follow uses ROCm 5.6.1, not ROCm 5.6, but I'm not sure it's relevant in your case.

PS : On 2070 mobile, every time I do an OOM, VRAM is not release, i need to restart biniou

Interesting behavior, that I didn't identified yet : OOM is supposed to kill the biniou process, so you have to restart it anyway.
It could probably be fixed by duplicating at the beginning of an inference the clean_ram() function that was invoked as the last step of a module. I will try to investigate and fix this.

If you're interested in testing it, you can try that for the Stable Diffusion module :

diff --git a/ressources/txt2img_sd.py b/ressources/txt2img_sd.py
index 1b19dc2..4c1aca2 100755
--- a/ressources/txt2img_sd.py
+++ b/ressources/txt2img_sd.py
@@ -193,7 +193,7 @@ def image_txt2img_sd(
     ):
 
     print(">>>[Stable Diffusion 🖼️ ]: starting module")
-
+    clean_ram()
     modelid_txt2img_sd = model_cleaner_sd(modelid_txt2img_sd)
     lora_model_txt2img_sd = model_cleaner_lora(lora_model_txt2img_sd)
     lora_model2_txt2img_sd = model_cleaner_lora(lora_model2_txt2img_sd)

It should clean RAM and VRAM before launching an inference, even after an OOM.

[ElPenguin]

Hi @Woolverine94

The Compatibility matrix for ROCm doesn't show versions prior to 6.0, so it's difficult to say if your hardware is compatible. Nevertheless, and if this could help, I've found this post on reddit that makes me realize that the installation procedure you follow uses ROCm 5.6.1, not ROCm 5.6, but I'm not sure it's relevant in your case.

Interesting... the matrix shows the GFX1030 die but not the 1032 (includes the RX 6600). Maybe this one isn't compatible.

If you're interested in testing it, you can try that for the Stable Diffusion module

I'm a beginner at AI, I'll probably do more testing at the end of the year. Nvidia is going to release graphics cards with a lot of RAM, and my goal is to find an alternative to cloud video generation.

One last quick question: I saw a project that allows RAM to be converted into virtual VRAM when the entire GPU VRAM is used. Is it possible to have an equivalent for this project?
https://github.com/pollockjj/ComfyUI-MultiGPU?tab=readme-ov-file

EDIT : and multi GPU support ? :3

[Woolverine94]

Hello @ElPenguin,

I'm a beginner at AI, I'll probably do more testing at the end of the year. Nvidia is going to release graphics cards with a lot of RAM, and my goal is to find an alternative to cloud video generation.

Honestly, biniou is probably not the good tool for such a goal : video modules are pretty basics and are more suitable for low quality animated gif than full quality long HD videos. But you can still try them and see if they suits your need !

One last quick question: I saw a project that allows RAM to be converted into virtual VRAM when the entire GPU VRAM is used. Is it possible to have an equivalent for this project?

In fact, biniou is already using CPU Offloading and VAE slicing by default. They're probably less effective than ComfyUI-MultiGPU, but from what I experimented, they're already pretty effective.

EDIT : and multi GPU support ? :3

Unlikely, mostly because I don't have access to multi-gpu nvidia hardware to add it and also because it's a feature that must be supported by the model.

[ElPenguin]

@Woolverine94

Thanks for all these answers. In any case, I still think I'll use Biniou, even in full CPU mode. It's a complete tool for most small tasks.

[Woolverine94]

@ElPenguin
Thanks to your for your tests and feedback !