FROM nvidia/cuda:12.1.1-devel-ubuntu22.04
ARG MODEL
ARG IMGMODEL
ARG WHISPERMODEL
ARG MMPROJ
ARG MODEL_NAME
ARG ADDITIONAL
RUN mkdir /opt/koboldcpp
RUN apt update && apt install git build-essential libopenblas-dev wget python3-pip -y
RUN git clone https://github.com/lostruins/koboldcpp /opt/koboldcpp
WORKDIR /opt/koboldcpp
COPY default.json /opt/koboldcpp/default.json
RUN make -j$(nproc) LLAMA_OPENBLAS=1 LLAMA_CUBLAS=1 LLAMA_PORTABLE=1 LLAMA_COLAB=1
RUN wget -O model.ggml $MODEL || true
RUN wget -O imgmodel.ggml $IMGMODEL || true
RUN wget -O mmproj.ggml $MMPROJ || true
RUN wget -O whispermodel.ggml $WHISPERMODEL || true
CMD /bin/python3 ./koboldcpp.py --model model.ggml --whispermodel whispermodel.ggml --sdmodel imgmodel.ggml --sdthreads 4 --sdquant --sdclamped --mmproj mmproj.ggml $ADDITIONAL --port 7860 --hordemodelname $MODEL_NAME --hordemaxctx 1 --hordegenlen 1 --preloadstory default.json --ignoremissing