| This is a high quality quantization of Yi-VL-6B and of the visual transformer. | |
| Q5_K is almost equal to fp16 in inference and Q6_K is basically the same (not as reliably tested for visual transformers but I assume it's equal to language models) | |
| You currently need to apply this PR to make it work: https://github.com/ggerganov/llama.cpp/pull/5093 - this adds the additional normalization steps into the projection | |
| I do not like this model, it's hallucinating more than anything else based on llava. |