[AMD][ROCm][gfx1103] Big things coming from my tiny package.

"Claude, the workflow does not run on my computer due to X, fix it, make no mistakes."

For me the issue was Krea2 not working at all with a LoRA on an AMD based mini pc with integrated gpu, which was weird because you'd figure with 64GB of total memory to play with that it could figure out where to put what to make it happen. Well what if it was just trying to do too much and wasn't optimized to run on the AMD pipeline? I think that was the case.

After an hour of subagents digging through the codebase the clanker dropped this on me:

"There's no way it just dug through all of that, wrote an entirely new attention mechanism with CLI flag and stopped it from consuming so much RAM." Slash me goes to sleep expecting to wake up to an OOM hyprland desktop recovery...

To my surprise it ran without issue generating wallpaper scale images. And no CPU pegged to 100% like when enabling dynamic vram on ROCm.

So my message to you, the end user struggling to get ComfyUI to work on their particular hardware, is to submit your problem to an agentic coding harness like Claude Code using the included AGENTS.md file in the ComfyUI root dir. No money? No problem. Use OpenCode, install it and select one of the free models using the /models command. Tell it to make any code changes on a testing git branch, if it doesn't work then revert. If it does? 2 E Z

https://github.com/user-attachments/assets/3ac09f46-4073-450e-9812-313f8987599f

Start with the ComfyUI root `AGENTS.md`, then identify the code and tests for the attention mechanism and CLI flag mentioned in the issue. The report describes a change made for one AMD ROCm system but does not specify a requested change or a reproducible test; clarify the desired outcome and how to verify it before starting.

Written by the indexing model from the issue text.