["/bin/bash" "/opt/nim/start_server.sh"]

BASE_IMAGE=nvcr.io/nvstaging/nim/china-vllm-basecontainer-stage3@sha256:a2596c7f08e17c6f91bdafec7fc0c50ee8448bdc32acae998b1dc3828dc74952 TELEMETRY_MODE=0 TELEMETRY_ENDPOINT=0 PIP_INDEX_URL=https://pypi.org/simple MANIFEST_FILE=model_manifest_with_checksum.yaml printf '%s\n' '#!/usr/bin/env bash' 'set -eu' 'export PYTHONPATH="/opt/nim:${PYTHONPATH:-}"' '# NVBug 6434380: NIM_SERVER_PORT must act as the public API port. nimlib' '# get_http_api_port already falls back to it, but other boot-chain' '# components consult NIM_HTTP_API_PORT only — alias it here, before ANY' '# process starts, so every consumer sees one consistent port. An explicit' '# NIM_HTTP_API_PORT still wins.' 'if [ -n "${NIM_SERVER_PORT:-}" ]; then case "${NIM_HTTP_API_PORT:-}" in ""|8000) export NIM_HTTP_API_PORT="$NIM_SERVER_PORT";; esac; fi' '# Video opt-in convention (FFmpeg-free mandate): if the user mounted FFmpeg' '# shared libs at /opt/ffmpeg8, PREPEND them (never replace — CUDA paths must' '# survive) so the torchcodec backend can dlopen them. No env needed by users.' 'if [ -d /opt/ffmpeg8 ]; then export LD_LIBRARY_PATH="/opt/ffmpeg8${LD_LIBRARY_PATH:+:$LD_LIBRARY_PATH}"; fi' '# NVBug 6428290 reopen-1: the serve path prints the startup banner (via nimlib),' '# but arg-driven subcommands exec directly and skip it. Print the banner for' '# list-model-profiles so QA sees the same banner (stdout, before the profile list).' 'if [ "$#" -gt 0 ]; then' '  case "$1" in list-model-profiles) [ -f /opt/nim/banner.txt ] && cat /opt/nim/banner.txt || true;; esac' '  exec "$@"' 'fi' 'exec start_server' > "$SERVER_START_SCRIPT_PATH" &&
  chmod a+rx "$SERVER_START_SCRIPT_PATH"

SERVER_START_SCRIPT_PATH=/opt/nim/start_server.sh

[8000/tcp]

BASE_IMAGE=nvcr.io/nvstaging/nim/china-vllm-basecontainer-stage3@sha256:a2596c7f08e17c6f91bdafec7fc0c50ee8448bdc32acae998b1dc3828dc74952 TELEMETRY_MODE=0 TELEMETRY_ENDPOINT=0 PIP_INDEX_URL=https://pypi.org/simple MANIFEST_FILE=model_manifest_with_checksum.yaml set -eux; for r in fp8-mtp fp8-dflash nvfp4-mtp nvfp4_fast-mtp; do test -f "/opt/nim/$r.yaml"; done; test ! -f /opt/nim/nvfp4-dflash.yaml; test ! -f /opt/nim/nvfp4_fast-dflash.yaml; grep -qx 'max_num_seqs: 1' /opt/nim/nvfp4-mtp.yaml; grep -qx 'gpu_memory_utilization: 0.6' /opt/nim/nvfp4-mtp.yaml; grep -qx 'max_num_seqs: 1' /opt/nim/nvfp4_fast-mtp.yaml; grep -qx 'gpu_memory_utilization: 0.6' /opt/nim/nvfp4_fast-mtp.yaml; python3 -c "import yaml; m=yaml.safe_load(open('/opt/nim/etc/default/model_manifest.yaml')); assert len(m['profiles']) == 4; assert {p['tags']['precision'] + '-' + p['tags']['recipe'] for p in m['profiles']} == {'fp8-mtp', 'fp8-dflash', 'nvfp4-mtp', 'nvfp4_fast-mtp'}"

config_files/*.yaml /opt/nim/

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

MANIFEST_FILE=model_manifest_with_checksum.yaml

nvs:1000

BASE_IMAGE=nvcr.io/nvstaging/nim/china-vllm-basecontainer-stage3@sha256:a2596c7f08e17c6f91bdafec7fc0c50ee8448bdc32acae998b1dc3828dc74952 TELEMETRY_MODE=0 TELEMETRY_ENDPOINT=0 PIP_INDEX_URL=https://pypi.org/simple set -eux; mkdir -p /etc/ld.so.conf.d; touch /etc/ld.so.conf.d/nvidia.conf; for p in /usr/local/nvidia/lib /usr/local/nvidia/lib64 /usr/local/cuda/lib64; do grep -qxF "$p" /etc/ld.so.conf.d/nvidia.conf ||
  echo "$p" >> /etc/ld.so.conf.d/nvidia.conf; done; if command -v ldconfig >/dev/null 2>&1; then ldconfig; fi