alright, so 100k is still kinda the limit but that's fine
S C
100k isn't bad when atomic-agent is so efficient with it and also it reads super fast it seems
S C
hipfire does need its own bespoke model format though, which is a pain, but still the primary bother with it isn't really the format itself, they do provide a way to convert a safetensor to it, with quantization and everything, but the issue is I can't benefit from the progress made in quantization until that's ported over into this thing
Samu ／人◕ ‿‿ ◕人＼
you want to convert from some other fancy quantization to this quantization?
S C
more like I want that quantization to be readable by hipfire (which runs real fast on my card) llama is compatible with AMD but it's not nearly as fast
