[CI/Build][Hardware][NVIDIA] Add public CUDA 13.4 Rubin build path (#54640)

Signed-off-by: Meenakshi Venkataraman <meenakshiv@nvidia.com>
Signed-off-by: meena-at-work <80416898+meena-at-work@users.noreply.github.com>
Signed-off-by: Shang Wang <shangw@nvidia.com>
Signed-off-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
Co-authored-by: Shang Wang <shangw@nvidia.com>
Co-authored-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
This commit is contained in:
meena-at-work
2026-09-09 04:53:12 +00:00
committed by GitHub
co-authored by Shang Wang Tyler Michael Smith
parent bf7eee9d7b
commit 385dce36bc
3 changed files with 63 additions and 33 deletions
+28 -8
View File
@@ -797,6 +797,9 @@ RUN CUDA_VERSION_DASH=$(echo $CUDA_VERSION | cut -d. -f1,2 | tr '.' '-') && \
cuda-nvcc-${CUDA_VERSION_DASH} \
cuda-cudart-${CUDA_VERSION_DASH} \
cuda-nvrtc-${CUDA_VERSION_DASH} \
# FlashInfer's runtime JIT-compiled kernels (e.g. trtllm_fmha_kernel_launcher.cu)
# #include <nvrtc.h>, which only the -dev package ships.
cuda-nvrtc-dev-${CUDA_VERSION_DASH} \
cuda-cuobjdump-${CUDA_VERSION_DASH} \
libcurand-dev-${CUDA_VERSION_DASH} \
libcublas-dev-${CUDA_VERSION_DASH} \
@@ -806,10 +809,21 @@ RUN CUDA_VERSION_DASH=$(echo $CUDA_VERSION | cut -d. -f1,2 | tr '.' '-') && \
numactl && \
rm -rf /var/lib/apt/lists/*
# Ensures nvcc/ptxas (installed above, needed at runtime for FlashInfer/
# DeepGEMM/EP kernel JIT compilation) are reachable on PATH. Harmless no-op
# if already set by FINAL_BASE_IMAGE.
ENV CUDA_HOME=/usr/local/cuda
ENV PATH="${CUDA_HOME}/bin:${PATH}"
ENV LD_LIBRARY_PATH="${CUDA_HOME}/lib64:${LD_LIBRARY_PATH}"
# Apply the CUDA 13.4-only libcurand compatibility link described in the build
# stage. CUDA 13.5 uses its native libcurand.so.11. DeepEP expects NCCL directly
# under its configured /usr prefix, while Ubuntu installs it in a multiarch
# subdirectory.
# subdirectory - but that symlink is only possible when NCCL is already
# present (e.g. nvcr.io/nvidia/cuda-dl-base bundles it; a bare-CUDA image
# doesn't, and gets it later via the standard nvidia-nccl-cu* pip package,
# UV_OVERRIDE below). Check reality (whether libnccl.so.2 is actually found)
# and tolerate not finding it instead of failing the build.
RUN if [ "${INSTALL_RUBIN_PRERELEASE}" = "true" ]; then \
CUDA_MAJOR_MINOR=$(echo "${CUDA_VERSION}" | cut -d. -f1,2) \
&& if [ "${CUDA_MAJOR_MINOR}" = "13.4" ]; then \
@@ -827,10 +841,13 @@ RUN if [ "${INSTALL_RUBIN_PRERELEASE}" = "true" ]; then \
echo "Unsupported Rubin CUDA version: ${CUDA_MAJOR_MINOR}" >&2; \
exit 1; \
fi \
&& NCCL_LIBRARY=$(find /usr/lib -name libnccl.so.2 -print -quit) \
&& test -n "${NCCL_LIBRARY}" \
&& ln -sf "${NCCL_LIBRARY}" /usr/lib/libnccl.so.2 \
&& ln -sf libnccl.so.2 /usr/lib/libnccl.so \
&& if NCCL_LIBRARY=$(find /usr/lib -name libnccl.so.2 -print -quit) \
&& [ -n "${NCCL_LIBRARY}" ]; then \
if [ "${NCCL_LIBRARY}" != "/usr/lib/libnccl.so.2" ]; then \
ln -sf "${NCCL_LIBRARY}" /usr/lib/libnccl.so.2; \
fi \
&& ln -sf libnccl.so.2 /usr/lib/libnccl.so; \
fi \
&& ldconfig; \
fi
@@ -961,10 +978,13 @@ ARG GDRCOPY_CUDA_VERSION=12.8
ARG GDRCOPY_OS_VERSION=Ubuntu24_04
ARG TARGETPLATFORM
COPY tools/install_gdrcopy.sh /tmp/install_gdrcopy.sh
# Some Rubin FINAL_BASE_IMAGE choices (internal Rubin CUDA images,
# nvcr.io/nvidia/cuda-dl-base) already provide GDRCopy; others don't, so it
# falls through to the normal install below. Check reality (ldconfig) rather
# than inferring from which FINAL_BASE_IMAGE flag is set.
RUN set -eux; \
if [ "${INSTALL_RUBIN_PRERELEASE}" = "true" ]; then \
# The Rubin CUDA image already provides GDRCopy. \
ldconfig -p | grep -q 'libgdrapi\.so'; \
if [ "${INSTALL_RUBIN_PRERELEASE}" = "true" ] && ldconfig -p | grep -q 'libgdrapi\.so'; then \
true; \
else \
case "${TARGETPLATFORM}" in \
linux/arm64) UUARCH="aarch64" ;; \