Estimate the VRAM needed for model weights and inference from parameter count, precision bytes and activation overhead, to choose deployment hardware.
训练还需优化器状态(通常 2–4 倍权重)。