Unclipped ggufs
Hey @HauhauCS , I was downloading your uncensored Gemma 4-4B, and since you don't post safetensors, I looked at the Q8 and it has clipped f16 tensors in a lot of places, and like Gemma 3, they are trained at bf16, and clipping both in conversion and quantization is a big problem most people don't realize yet. So actually in my testing a bf16 trained model is better off with a clean gguf conversion to bf16 then quanted to Q8 then to be quanted or converted to fp16. Does way less damage, so can you do a bf16 conversion to gguf and post it or a Q8P with bf16 tensors and not fp16? Thanks. Oh the fp16 vision projector too, but I can go grab that from Bartowski, but just for anyone reading this with anything from 2020 or newer 30 series card, stick to bf16 which is 95 plus percent of people and amd rnda 3 and 4 and apple mls support bf16 too.