Runtime Error on Tiled VAE Encode Node

[alexcetea]

Hi, when trying to use the Tiled VAE Encode Node I get the following error:

VAEEncodeTiled_TiledDiffusion Input type (float) and bias type (struct c10::BFloat16) should be the same

Attached workflow. archviz_ai_basic.json

How can I fix this?

[stlasVK]

With the help of ChatGPT, I solved the issue by adding the --fp32-vae flag to the run_nvidia_gpu.bat file. The updated command looks like this:

.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --fp32-vae pause

This forces the VAE to run in full precision (float32) mode, resolving the compatibility issue I was facing.

Note: Running VAE in full precision (float32) may increase GPU VRAM usage compared to half precision (float16). If you have limited VRAM, this might affect performance or lead to out-of-memory errors when handling large workflows or high-resolution images.

[dolphin-electric]

Thank you this helped me, I was going nuts. I don't understand how it could suddenly happen from nowhere.. I hope a fix is comming