Inconsistent performance between flux and SDXL with precision half and all-in-fp32 on AMD RDNA2

Nobody has claimed this yet.

I can get massive speedups on my AMD 6800xt for flux by enabling --all-in-fp32 rather than --precision half (on the nf4 model, still need to try with others). For SDXL, this has the opposite effect in pretty much the same magnitude (almost half as fast).

Flux with --precision half: 12-14s/it
Flux with --all-in-fp32: 7.something s/it
SDXL with --precision half 1.55 it/s or 0.64s/it
SDXL with --all-in-fp32 1.1s/it

(All for a 1152x896 image with >20 steps).

I still want test.with other flux models than the nf4, but it's very unexpected (the nf4 performance is in line.with what I got from fp8 a while back)

Am I taking crazy pills? And if that has to be for some reason, is there a way to switch the precision without restarting forge?

Start by reproducing the reported Flux NF4 and SDXL runs on an AMD 6800 XT with --precision half and --all-in-fp32 at 1152x896 for more than 20 steps. Compare the measured iterations per second, then determine why the precision options behave differently and whether switching precision without restarting Forge is possible.

Written by the indexing model from the issue text.