ARG BASE_IMAGE FROM ${BASE_IMAGE} COPY plugin /opt/dsv4_vision_src COPY patch_dspark_skip_bias_vl.py /tmp/patch_dspark_skip_bias_vl.py COPY patch_mm_hc_streams.py /tmp/patch_mm_hc_streams.py RUN python3 -m pip install --no-cache-dir --no-deps /opt/dsv4_vision_src \ && python3 /tmp/patch_dspark_skip_bias_vl.py \ && python3 /tmp/patch_mm_hc_streams.py ENV VLLM_PLUGINS=dsv4_vision ENV TORCH_CUDA_ARCH_LIST=12.1a ENV FLASHINFER_CUDA_ARCH_LIST=12.1a ENTRYPOINT ["vllm", "serve"]