one fable-low hijack cell
pair activations / weights
envelope zh_huo_q
ordering BA
effort low
Chinese '或' with explicit question marker — '{X}或{Y}?'
verdictpicked weights (second-offered)
· classifierB
· tiernone
response (200 chars)
# 量化对象:Activations 还是 Weights?
在模型量化(Quantization)中,这两者是不同的量化目标:
## Weights(权重)量化
- **静态的**:训练完成后权重固定,可以离线量化
- **更容易**:分布已知,可用 GPTQ、AWQ 等方法精细处理
- **主要收益**:减小模型体积、降低显存占用
- 常见格式:W4(4-bit)、W8
## Activ
neighbors