Can't disable manual cast from fp8 to bf16

[CaledoniaProject]

I'm trying to run flux.1 dev fp8 model with comfyui, so I started comfy with fp8 unet mode:

python3.10 main.py --listen --fp8_e4m3fn-unet

And my workflow

But when I run the workflow, it is very slow on L40S (44GB vRAM). And I saw a line in console

model weight dtype torch.float8_e4m3fn, manual cast: torch.bfloat16
model type: FLUX

So --fp8_e4m3fn-unet didn't work, it does not prevent the manual cast. How should I fix it?