Llama-vision Attentions. Find top attended patch ids

[claudiopisa]

I have been analysing the Llama-3.2InferanceAnalysis.ipynb code you posted. At some point in the code the top attented patch are plotted over the image:

#Assume 14x14 patch size → 16×16 grid → 256 total patches
patch_size = 14 
top_attended_patch_ids = [85,86,69,70,53,54,101,102,117,118]  # Replace with your actual patch indices

#Draw rectangles around top patches#
patch_id = 0
for row in range(0, 224, patch_size):
    for col in range(0, 224, patch_size):
        if patch_id in top_attended_patch_ids:
            draw.rectangle([col, row, col + patch_size, row + patch_size], outline="red", width=2)
            draw.text((col + 2, row + 2), str(patch_id), fill="red")
        patch_id += 1

Could you please help me understand how did you find the values of the top attended patch ids ? Because the heatmaps plotted have on the x-axis the embeddings values of the patches, so i don't understand how to retrieve the original patch id values