error: input types 'tensor<2x1280xf32>' and 'tensor<1280xf16>' are not broadcast compatible

[tvsj]

Device Environment
macOS m2Max
python3.12

[kijai]

I think on Mac you have to set the model dtype to fp32 too. This didn't work at all on MPS before, I did get it running on fp32 for someone else, but I can't really further work on that, it's an issue with the original code.

[tvsj]

I think on Mac you have to set the model dtype to fp32 too. This didn't work at all on MPS before, I did get it running on fp32 for someone else, but I can't really further work on that, it's an issue with the original code.

thanks.I tried to change dtype to FP32. The memory usage is too high. My device has only 32GB of memory and cannot support the program.

then i export PYTORCH_MPS_HIGH_WATERMARK_RATIO=0.0
The program did not report an error, but there was no progress change in this state for a long time.

Finally, something to celebrate happened!The progress bar is moving!
A rough estimate of 9% of the progress took at least 10 minutes

[KannManMachen]

I can confirm that setting the "dtype" to "fp32" for these notes did the trick:

• SUPIR Model Loader (v2)
• SUPIR First Stage (Denoiser)
• SUPIR Encode

Used hardware:

• Apple M2 Max
• 12x CPU / 38x GPU
• 64 GB RAM

Required time for upscaling/enhancing from 512 px to 1536 px:

• first run: 180 seconds
• second run: 140 seconds