Fix CUDA driver linking in llama.cpp image

platform/docker/llama-cpp/Dockerfile

@@ -8,6 +8,13 @@ ARG CUDA_ARCHITECTURES="86;120"
 RUN apt-get update && apt-get install -y --no-install-recommends \
     ca-certificates cmake git libcurl4-openssl-dev ninja-build pkg-config && \
     rm -rf /var/lib/apt/lists/*
+# CUDA's devel image keeps libcuda.so.1 in its compatibility directory, which
+# is not part of the default linker search path. llama.cpp's shared CUDA
+# backend links successfully without this, but the final server link then
+# cannot resolve CUDA driver symbols such as cuMemCreate. Register it for the
+# build stage; at runtime the NVIDIA container runtime supplies the host driver.
+RUN printf '%s\n' /usr/local/cuda/compat >/etc/ld.so.conf.d/cuda-compat.conf && \
+    ldconfig
 RUN test -n "$LLAMA_CPP_COMMIT"
 RUN git clone --filter=blob:none https://github.com/ggml-org/llama.cpp /src/llama.cpp && \
     git -C /src/llama.cpp checkout "$LLAMA_CPP_COMMIT"