#ratio

A decoder-only model has 6.7 billion parameters in total, 32 layers and $d_{model} = 4096$. You apply LoRA with rank $r = 16$ to all four attention projections $W_q, W_k, W_v, W_o$ (each $4096 \times 4096$) in every layer. What percentage of the model's parameter count is trainable?