Calc

1b = 十亿 = 1e9

>>> 1e9/1024/1024/1024
0.9313225746154785

例: FP32 7B参数量:7 * 0.93 * (32/8)=26.04GB

REF

vLLM推理性能鉴赏 - 知乎 (zhihu.com)

huggingface.co