llama.cpp b4729 batch size over 512 tanks throughput on 4090

tested llama 3.3 70b q4_k_m on llama.cpp b4729 with batch sizes 128, 256, 512, 1024, 2048 on single 4090. throughput peaks at batch 512 (31.2 tok/s) then drops hard at 1024 (18.7 tok/s) and crashes with cuda oom at 2048. expected throughput to keep climbing with batch size but it just tanks past 512.... anyone know if this is gpu memory bandwidth limit or something in llama.cpp implementation