I also meet the similar issue. Use python3.9, TensorFlow-macos 2.9.0, metal0.4. When train a YOLOV3. the memory will raise to 80G