- IQ4F : IQ4_XS feed-forawrd (IQ4_NL for ffn_down due to shape constraints)
- Q8A : Q8_0 attention, Q8_0 output, Q8_0 embeds
- Q8SH : Q8_0 shared experts
Readable speeds on a 24GiB GPU + 64GB RAM w/ long context
Readable speeds on a 24GiB GPU + 64GB RAM w/ long context
有文件?提交种子。
提交种子