Phr00t, Any Chance to Try Pruned Qwen with 14B Params?
#115
by
qpqpqpqpqpqp
- opened
I'm think this breaks LORA compatibility since layers won't align anymore.
I ran it with Lightning
It might not crash as some layers probably still do align, but I suspect quality is reduced... especially with NSFW LORAs (on top of quality loss due to pruning). I think using a GGUF would be better than this...
I know, but the pruned model is faster for me and needs less memory. It takes longer to generate with the 20B model