make the example actually generate instead of only filling the inputs

gray

show sub-second encode in ms

compact UI, one-line header with measured cost, settings in accordion

vendor get_head_mask, removed from transformers 5

call post_init so transformers 5 sets all_tied_weights_keys

load 2023 checkpoints with weights_only=False for torch 2.6+

MiniGPT-4 v1 on Vicuna v0, patched for current transformers

initial commit