automatic(fp16 lora) will fry image if fp16 flux unet used

not sure if this is a normal usecase.
use 22g flux fp16 unet, fp8 t5,
lora is https://civitai.com/models/180891/flat-color-anime?modelVersionId=738709
when using automatic(fp16 lora) in diffusion low bits, the image generated will get fried.
if i using "automatic" in diffusion low bits, the image will be good, but it's slow than above.

Reproduce the report with the 22G Flux fp16 UNet, fp8 T5, the linked Flat Color Anime LoRA, and both Diffusion low-bits modes. Start by tracing the code path for “automatic(fp16 lora)” and compare it with “automatic”; done means the fp16-LoRA path produces an undistorted image without the slower-mode regression.

Written by the indexing model from the issue text.