Error while deserializing header: HeaderTooSmall
Downloaded the sd3_medium.safetensors model file, along with the 2 clip models and the t5xxl_fp8_e4m3fn text encoder.
With the basic workflow all the paths in the loader nodes are specified correctly, but I get the above error when trying to run it when it reaches the CLIP triple loader node.
Doing a git pull instead enough, you also need to:
pip install --upgrade torch torchvision torchaudio -r requirements.txt
I found the problem.
t5xxl_fp8_e4m3fn requires a higher version of torch (2.3.1 is available).
But updating torch seems a bit troublesome, and the corresponding cu for 2.3.1 is cu121.
So I recommend using t5xxl_fp16.
Probably no point in posting this since I'm probably the only person who experienced this issue in hindsight but eh, whatever.
Reason trying to load the FP8 encoder failed is because the download appears to have gotten corrupted. Didn't notice until I checked the file I had and the repository and noticed the size was off by a few hundred mb.
Guessing it happened because the servers were strained with people rushing to download SD3 since that's never happened before.
After redownloading the issue has been resolved.