diff --git a/scripts/ci/ci_install_deepep.sh b/scripts/ci/ci_install_deepep.sh index d46b2c13f..738ac7612 100755 --- a/scripts/ci/ci_install_deepep.sh +++ b/scripts/ci/ci_install_deepep.sh @@ -46,9 +46,6 @@ if [ ! -e "$LIB_PATH/libmlx5.so" ]; then fi apt-get update && apt-get install -y libfabric-dev -# Install NVSHMEM -pip install nvidia-nvshmem-cu12==3.4.5 --force-reinstall - # Install DeepEP DEEPEP_DIR=/root/.cache/deepep rm -rf ${DEEPEP_DIR} diff --git a/scripts/ci/ci_install_dependency.sh b/scripts/ci/ci_install_dependency.sh index cac70576b..d2370e18d 100755 --- a/scripts/ci/ci_install_dependency.sh +++ b/scripts/ci/ci_install_dependency.sh @@ -134,6 +134,10 @@ if [ "$IS_BLACKWELL" != "1" ]; then git clone --branch v0.5 --depth 1 https://github.com/EvolvingLMMs-Lab/lmms-eval.git $PIP_CMD install -e lmms-eval/ $PIP_INSTALL_SUFFIX fi + +# DeepEP depends on nvshmem 3.4.5 +$PIP_CMD install nvidia-nvshmem-cu12==3.4.5 --force-reinstall $PIP_INSTALL_SUFFIX + # Cudnn with version less than 9.16.0.29 will cause performance regression on Conv3D kernel $PIP_CMD install nvidia-cudnn-cu12==9.16.0.29 --force-reinstall $PIP_INSTALL_SUFFIX $PIP_CMD uninstall xformers || true