llama.cpp b4821 crashes on macos 15.2 when loading q6_k quants over 45gb

tested llama.cpp b4821 on macos 15.2 with m2 max 96gb and it crashes when loading llama 3.3 70b q6_k (48.2gb). the crash happens during model load, not inference.... metal backend just segfaults with no useful error message. q5_k_m (43.1gb) loads fine. q4_k_m (38.6gb) loads fine. only q6_k fails. anyone else seeing this or is it just my setup....