So I made an image classifier API as that is the best way for me to package moondream2 as it insists on being in python. Used grok to help me make it and was surprised it was only ~40 lines using fastapi.
But it didn't work!
The simple example i had created the model, loaded the image and ran questions against it, but grok insisted on separating the embeddings and just could not fix the errors that kept happening until i reached my daily limits.
Also tried qwen-coder 30b locally and it created similar code and it also did not work!
So i manually looked at this frankencode and quickly solved the issue and the thing actually works!
I am even so versed in this shit (python venv/packaging;) that it works flawlessly with an AMD card on linux.
