Model loaded twice?

[mvsoom]

Why is the model loaded twice? I'm curious.

Once in 8 bit:

loopvlm/generate.py

Line 302 in e15de9b

model = _load_model(checkpoint_path, device, precision, use_tp)

Once in bfloat16:

loopvlm/generate.py

Line 321 in e15de9b

_model = PaliGemmaForConditionalGeneration.from_pretrained(