Solution to the memory issues in the official demo

[hubert0527]

There is a typo here:

FLUX-Controlnet-Inpainting/main.py

Line 17 in 7c00862

"black-forest-labs/FLUX.1-dev", subfolder='transformer', torch_dytpe=torch.bfloat16

The fix
The fix is to replace torch_dytpe with torch_dtype.
Before the fix, the model ignores the data type specification, and loads the flux transformer at float32, which requires gigantic memory.
After the fix, I can inference at 28.5 GB memory.

Potentially related to #3, #23, #24

Misc
As a side note for those who wants more memory saving, somehow pipe.enable_xformers_memory_efficient_attention() has conflict with FLUX+controlnet (reports a shape mismatch error in #9.
On the other hand, I haven't tested pipe.enable_model_cpu_offload() yet.

[aycaecemgul]

Thank you! I was frusturated with this!