Issues on Mac: fp8 and fp16 don't work because dtype is not compatible and fp32 crashes memory

[haiggoh]

fp8_unet gives me this error: Trying to convert Float8_e4m3fn to the MPS backend but it does not have support for that type.

dtype auto (I think this defaults to fp16?) gives me this error: "compute_index_ranges_weights" not implemented for 'Half'

dtype fp32 seems to work maybe? But only with PYTORCH_MPS_HIGH_WATERMARK_RATIO=0.0 because of memory issues and so far I only got noise, but I keep trying.

If anyone knows how to fix any one of these by any chance I would love to see it!