["/bin/bash" "-c" "$SERVER_START_SCRIPT_PATH"]

RUN |3 TELEMETRY_MODE=0 TELEMETRY_ENDPOINT=0 PIP_INDEX_URL=https://urm.nvidia.com/artifactory/api/pypi/nv-shared-pypi/simple /bin/sh -c touch $SERVER_START_SCRIPT_PATH &&
  chmod a+rx $SERVER_START_SCRIPT_PATH &&
  cat > $SERVER_START_SCRIPT_PATH <<-EOF #!/usr/bin/env bash set -eu # SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved. # SPDX-License-Identifier: LicenseRef-NvidiaProprietary # L40 GPU fix: nimlib maps "L40" to "l4" due to substring matching, so we force the L40S # TensorRT profile. L40 and L40S share the same Ada Lovelace architecture (sm_89). if [ -z "\${NIM_MODEL_PROFILE+x}" ]; then GPU_NAME=\$(nvidia-smi --query-gpu=name --format=csv,noheader 2>/dev/null | head -1 ||
  echo "") if echo "\$GPU_NAME" | grep -qi "L40S"; then export PIPELINE_GPU_SKU=\${PIPELINE_GPU_SKU:-l40s} export NIM_MODEL_PROFILE=461089b6381f768a0b7e5f3a6e0e5080128ec95fbeff9849b4b2523c4c74e24f echo "Detected NVIDIA L40S GPU - using L40S TensorRT profile" elif echo "\$GPU_NAME" | grep -qi "L40"; then export PIPELINE_GPU_SKU=\${PIPELINE_GPU_SKU:-l40s} export NIM_MODEL_PROFILE=461089b6381f768a0b7e5f3a6e0e5080128ec95fbeff9849b4b2523c4c74e24f echo "Detected NVIDIA L40 GPU - using L40S TensorRT profile (same architecture)" fi fi # Test CUDA IPC functionality if [ -z "\${USE_CUDA_IPC+x}" ]; then if python3 "$NIM_DIR_PATH/test_cuda_ipc.py" >/dev/null; then export USE_CUDA_IPC=1 else export USE_CUDA_IPC=0 fi else export USE_CUDA_IPC fi echo "Using CUDA IPC: \${USE_CUDA_IPC}" # MPS: set pipe directory BEFORE starting daemon so daemon and clients use same path if [ "\${ENABLE_CUDA_MPS:-0}" = "1" ]; then export CUDA_MPS_PIPE_DIRECTORY=\${CUDA_MPS_PIPE_DIRECTORY:-/tmp/nvidia-pipe} if output=\$(nvidia-cuda-mps-control -d 2>&1); then echo "Success: MPS available (pipe dir: \$CUDA_MPS_PIPE_DIRECTORY)" ps -ef | grep mps # Disable CUDA IPC under MPS: cudaEventSynchronize() deadlocks under MPS export USE_CUDA_IPC=0 echo "MPS enabled: Forcing USE_CUDA_IPC=0" else echo "MPS not available" fi else echo "CUDA MPS not enabled (ENABLE_CUDA_MPS=${ENABLE_CUDA_MPS:-0})" fi start_server EOF # buildkit

SERVER_START_SCRIPT_PATH=/opt/nim/start_server.sh

model_manifest.yaml /opt/nim/etc/default/model_manifest.yaml

nvs:1000

TELEMETRY_MODE=0 TELEMETRY_ENDPOINT=0 PIP_INDEX_URL=https://urm.nvidia.com/artifactory/api/pypi/nv-shared-pypi/simple [ -d /etc/ld.so.conf.d ] ||
  mkdir -p /etc/ld.so.conf.d &&
  [ -f /etc/ld.so.conf.d/nvidia.conf ] ||
  touch /etc/ld.so.conf.d/nvidia.conf &&
  echo "/usr/local/nvidia/lib" >> /etc/ld.so.conf.d/nvidia.conf &&
  echo "/usr/local/nvidia/lib64" >> /etc/ld.so.conf.d/nvidia.conf &&
  echo "/usr/local/cuda/lib64" >> /etc/ld.so.conf.d/nvidia.conf &&
  if command -v ldconfig >/dev/null 2>&1; then ldconfig; fi

TELEMETRY_MODE=0 TELEMETRY_ENDPOINT=0 PIP_INDEX_URL=https://urm.nvidia.com/artifactory/api/pypi/nv-shared-pypi/simple groupadd --gid 1000 --non-unique nvs &&
  useradd --create-home --shell /usr/sbin/nologin --uid 1000 --non-unique --gid 1000 nvs &&
  chown 1000.1000 $NIM_DIR_PATH

./test_cuda_ipc.py /opt/nim

./LICENSE ./VERSION ./Acknowledgements.txt /opt/nim

./metrics_helper.py /opt/nim