Anonymous
8/6/2025, 1:47:59 AM
No.106156902
>>106156802
not for me, since I use it as a backend. uh... at this point, just their antislop and familiarity with the launch args personally. I started using it because they offered a binary before llama.cpp as far as I remember, and I was having issues with nvcc at the time and compiling for cublas kept fucking up. the antislop is logit bias with extra steps, but the extra steps are nifty and the last PR I found for llama.cpp about it was years ago and basically said it would be totally incompatible. not sure how kobold did it but I don't see why llama.cpp couldn't just copy their implementation, but what do I know (not much)
not for me, since I use it as a backend. uh... at this point, just their antislop and familiarity with the launch args personally. I started using it because they offered a binary before llama.cpp as far as I remember, and I was having issues with nvcc at the time and compiling for cublas kept fucking up. the antislop is logit bias with extra steps, but the extra steps are nifty and the last PR I found for llama.cpp about it was years ago and basically said it would be totally incompatible. not sure how kobold did it but I don't see why llama.cpp couldn't just copy their implementation, but what do I know (not much)