Bazzite, ROCM and AI image generation
I changed my GPU recently and did a fresh reinstall of Bazzite. I've been playing around with AI stuff, since I couldn't before (Used to have a 4gb RX 570 and then an 8gb 5700XT)
Using Podman, I created a container for Ollama and OpenWebUI. These work great. I downloaded Pinokio's Appimage and played around with some TTS tools and everything seemed to be working fine as well.
Two days ago I decided to dip my toes in AI image generation and tools. I do graphic design, mostly vector graphics for CNC, and one of the most common tasks is getting a crappy logo from the client (usually a mobile screenshot off Canva) and I need to clean it up and trace it (usually manually) on InkScape.
I could not get any tools to work properly. I tried several approaches, and I'm at a loss of where to even begin.
Stuff I tried:
InvokeAI via Pinokio
Doesn't work. No matter the settings I add to invoke.yaml or the env vars I pass, it keeps falling back to CPU mode
InvokeAI via Distrobox
Doesn't work. For some reason Invoke insists on installing non-rocm compatible pytorch versions.
InvokeAI via Podman (using AMD's ROCm containers as a base)
Doesn't work. Same as before, it refuses to use the correct libraries, specifically bitsandbytes
InvokeAI via Python UV (manual install)
This kinda worked. but it was super unstable and I was getting constant OOM errors, even when rocm-smi showed that VRAM was at ~60% usage.
ComfyUI via Podman (https://github.com/corundex/ComfyUI-ROCm?tab=readme-ov-file)
This image was specifically created for rocm support, tested with my exact same GPU model (RX 9070XT 16gb), yet it kept crashing as soon as I started a workflow.
Researching on this topic is becoming a slog, I was wondering if someone here has successfully installed and used any of these tools.
I'm really interested in InvokeAI, since it would make everyday tasks (cleanup, redraws, flatten) a lot easier and quicker, and seems to have a pretty decent workflow
Using Podman, I created a container for Ollama and OpenWebUI. These work great. I downloaded Pinokio's Appimage and played around with some TTS tools and everything seemed to be working fine as well.
Two days ago I decided to dip my toes in AI image generation and tools. I do graphic design, mostly vector graphics for CNC, and one of the most common tasks is getting a crappy logo from the client (usually a mobile screenshot off Canva) and I need to clean it up and trace it (usually manually) on InkScape.
I could not get any tools to work properly. I tried several approaches, and I'm at a loss of where to even begin.
Stuff I tried:
InvokeAI via Pinokio
Doesn't work. No matter the settings I add to invoke.yaml or the env vars I pass, it keeps falling back to CPU mode
InvokeAI via Distrobox
Doesn't work. For some reason Invoke insists on installing non-rocm compatible pytorch versions.
InvokeAI via Podman (using AMD's ROCm containers as a base)
Doesn't work. Same as before, it refuses to use the correct libraries, specifically bitsandbytes
InvokeAI via Python UV (manual install)
This kinda worked. but it was super unstable and I was getting constant OOM errors, even when rocm-smi showed that VRAM was at ~60% usage.
ComfyUI via Podman (https://github.com/corundex/ComfyUI-ROCm?tab=readme-ov-file)
This image was specifically created for rocm support, tested with my exact same GPU model (RX 9070XT 16gb), yet it kept crashing as soon as I started a workflow.
Researching on this topic is becoming a slog, I was wondering if someone here has successfully installed and used any of these tools.
I'm really interested in InvokeAI, since it would make everyday tasks (cleanup, redraws, flatten) a lot easier and quicker, and seems to have a pretty decent workflow
GitHub
ComfyUI with AMD ROCm support for GPU-accelerated AI image generation on AMD RX 6000/7000+ GPUs - corundex/ComfyUI-ROCm
