知识卡片
权重内存只是推理占用下限
参数数乘每参数字节数可估算原始权重存储,但推理还需要缓存、激活、临时缓冲与量化元数据等。训练另含梯度和优化器状态;不能只用7B×2字节就保证14GB设备可稳定运行所有长度和并发。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
已有反馈 · 0 条
还没有人提交反馈。
在整书笔记中的位置
- · 按任务适配模型
原书出处 · 1 处
- · 查看原书依据
32-bit floating point representation, each parameter occupying 4 bytes of storage. Float16 16-bit floating point representation. Only 5 bits are reserved for the exponent as opposed to 8 bits in Float32. This means that using Float16 comes with overflow/underflow problems for very large and small numbers. bfloat16 (BF16) 16-bit floating point representation. Just like Float32, 8 bits are reserved for the exponent, thus alleviating the
参考来源
- 位置:《Designing Large Language Model Applications》第05章《按任务适配模型》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch05.html)
- 结论依据:摘录中关于权重内存只是推理占用下限的定义、比较或条件支撑卡片;原书表述不严谨的范围已在覆盖记录逐项限定,不把案例结果外推为普遍结论。
- 原始内容:32-bit floating point representation, each parameter occupying 4 bytes of storage. Float16 16-bit floating point representation. Only 5 bits are reserved for the exponent as opposed to 8 bits in Float32. This means that using Float16 comes with overflow/underflow problems for very large and small numbers. bfloat16 (BF16) 16-bit floating point representation. Just like Float32, 8 bits are reserved for the exponent, thus alleviating the