A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 10, 17:45 UTC
[docker] install cuda13 version of lmcache and nixl (#30913)
Signed-off-by: Qidong Su <soodoshll@gmail.com>
Qidong Su committed
Jan 5, 2026 at 15:50 UTC
af1b07b0c57fab2b65f55a0c14ced3186c23a63f
1 file changed
+20
-1
docker/Dockerfile
+20
-1
@@ -615,6 +615,7 @@ RUN mv vllm src/vllm
615
FROM vllm-base AS vllm-openai-base
616
ARG TARGETPLATFORM
617
ARG INSTALL_KV_CONNECTORS=false
618
+ARG CUDA_VERSION
619
620
ARG PIP_INDEX_URL UV_INDEX_URL
621
ARG PIP_EXTRA_INDEX_URL UV_EXTRA_INDEX_URL
@@ -626,8 +627,26 @@ ENV UV_HTTP_TIMEOUT=500
627
# install kv_connectors if requested
628
RUN --mount=type=cache,target=/root/.cache/uv \
629
--mount=type=bind,source=requirements/kv_connectors.txt,target=/tmp/kv_connectors.txt,ro \
630
+ CUDA_MAJOR="${CUDA_VERSION%%.*}"; \
631
+ CUDA_VERSION_DASH=$(echo $CUDA_VERSION | cut -d. -f1,2 | tr '.' '-'); \
632
+ CUDA_HOME=/usr/local/cuda; \
633
+ # lmcache requires explicit specifying CUDA_HOME
634
+ BUILD_PKGS="libcusparse-dev-${CUDA_VERSION_DASH} \
635
+ libcublas-dev-${CUDA_VERSION_DASH} \
636
+ libcusolver-dev-${CUDA_VERSION_DASH}"; \
637
if [ "$INSTALL_KV_CONNECTORS" = "true" ]; then \
630
- uv pip install --system -r /tmp/kv_connectors.txt || true; \
638
+ if [ "$CUDA_MAJOR" -ge 13 ]; then \
639
+ uv pip install --system nixl-cu13; \
640
+ fi; \
641
+ uv pip install --system -r /tmp/kv_connectors.txt --no-build || ( \
642
+ # if the above fails, install from source
643
+ apt-get update -y && \
644
+ apt-get install -y --no-install-recommends ${BUILD_PKGS} && \
645
+ uv pip install --system -r /tmp/kv_connectors.txt --no-build-isolation && \
646
+ apt-get purge -y ${BUILD_PKGS} && \
647
+ # clean up -dev packages, keep runtime libraries
648
+ rm -rf /var/lib/apt/lists/* \
649
+ ); \
650
fi
651
652
ENV VLLM_USAGE_SOURCE production-docker-image