llama.cpp metal backend crashes past 28k context on m2 max, cuda works fine at 64k

tested on llama.cpp b4729 with llama 3.3 70b q4_k_m.... metal backend on m2 max crashes around 28.4k tokens with "metal buffer allocation failed", same model on cuda (4090) handles 64k context clean is this a known metal limitation or is there a config flag im missing