Does this means it's not recommended to use for now untill TheROCK fixes it?
Generating images with ComfyUI using the triton backend doesn't take a lot of time, probably 1min per image for 1mpx, 7 minutes for 2mpx images.
Thanks — that's a useful data point, and it answers your own question: yes, it's usable now, and your numbers are better than the ones on this card.
This card measures the torch + MIOpen path. The cost there is a missing MIOpen tuning database for gfx1151 — /opt/rocm/share/miopen/db/gfx1151_20.HIP.fdb.txt is unreadable, and that install ships pretuned DBs for gfx908 / gfx90a / gfx942 and none for gfx1151. So every convolution shape is auto-tuned the first time it is seen. That is why VAE decode was 825.6 s cold and 20.5 s warm — a 40× swing with nothing about the model or the code changed. The transformer is GEMM work and unaffected, which is why the sampler rate was stable across both runs.
Triton sidesteps that entirely, because it generates and compiles its own kernels rather than dispatching into MIOpen's convolution library — so there is no tuning database to be missing. That is consistent with your ~1 min at 1 MP against my 6.8 min warm.
So: not "wait for TheROCK" — use the triton backend, or accept one slow warm-up on the MIOpen path and reuse the process.
I'd like to add your numbers to the card, with attribution, if you're happy for me to. What ComfyUI backend/flags and ROCm version are you on?