flux:different results on different machines.

Hi！
During the debugging phase, I've noticed that training the [FLUX-Controlnet-Inpainting](https://github.com/alimama-creative/FLUX-Controlnet-Inpainting) model yields different results on different machines.

Specifically, on one machine, everything trains fine,
but on another, I'm getting an error that says "Gradient for module.xxx contains NaN!".

I've double-checked and confirmed that my code is exactly the same on both machines, and use the same seed in set_seed().

Has anyone else encountered this same issue? I'm curious to know if there are any potential causes for this discrepancy in training results across different machines. Any insights or suggestions would be greatly appreciated.

Start with the linked FLUX-Controlnet-Inpainting training setup and compare the two machines, including the environment and the point where the NaN gradient appears. Reproduce the discrepancy with the same seed; done means identifying a machine-dependent cause or a reliably reproducible failure and documenting the relevant conditions.

Vom Indexierungsmodell aus dem Issue-Text verfasst.