amc-setup-calibration-stack

技能：启动 AutoMagicCalib 发布版容器

通过发布版容器设置 AutoMagicCalib 微服务和 UI：解析 AMC 检出目录、登录 NGC、按需下载 VGGT、配置 Docker Compose、启动服务并验证就绪状态。

前提条件

• 已安装 Docker 和 Docker Compose
• 已配置 NVIDIA Docker Runtime（用于 GPU 支持）
• 磁盘上有 auto-magic-calib 仓库。步骤 0b 会解析当前仓库、DeepStream tools/auto-magic-calib、DEEPSTREAM_REPO_ROOT 或 ~/auto-magic-calib；否则会在克隆 https://github.com/NVIDIA-AI-IOT/auto-magic-calib 前征求同意。
• 拥有可访问 NVIDIA 容器注册表的 NGC 账户
• 无需 sudo 即可运行 Docker；继续前用 docker ps 验证。

说明

步骤 0：验证 Docker 无需 sudo 即可运行

docker ps

• 如果成功 → 继续。
• 如果失败并显示“permission denied”→ 用户不在 docker 组中。请用户运行：

  sudo usermod -aG docker $USER && newgrp docker
  

  然后请用户确认 docker ps 可以正常运行，再继续。

Agent 提示：如果无法在 Agent 沙箱中运行 docker ps，请先让用户确认该命令可以正常运行（例如：“你能确认 docker ps 无需 sudo 即可运行吗？”），然后再继续。

步骤 0b：解析仓库检出目录

此技能需要 AMC 仓库资源（compose/、示例数据和 models/）。请先解析现有检出目录；克隆到 ~/auto-magic-calib 前征求同意。

REPO_URL="https://github.com/NVIDIA-AI-IOT/auto-magic-calib.git"
DEFAULT_CLONE_DIR="$HOME/auto-magic-calib"
CURRENT_GIT_ROOT="$(git rev-parse --show-toplevel 2>/dev/null || true)"

is_amc_checkout() {
  [ -n "$1" ] \
    && [ -f "$1/README.md" ] \
    && grep -q "AutoMagicCalib" "$1/README.md" 2>/dev/null \
    && [ -f "$1/compose/compose.yml" ] \
    && grep -q "auto-magic-calib-ms" "$1/compose/ms/compose.yml" 2>/dev/null \
    && grep -q "auto-magic-calib-ui" "$1/compose/ui/compose.yml" 2>/dev/null
}

REPO_ROOT=""
for candidate in \
  "$CURRENT_GIT_ROOT" \
  "${CURRENT_GIT_ROOT:+$CURRENT_GIT_ROOT/tools/auto-magic-calib}" \
  "${DEEPSTREAM_REPO_ROOT:+$DEEPSTREAM_REPO_ROOT/tools/auto-magic-calib}" \
  "$PWD/tools/auto-magic-calib" \
  "$DEFAULT_CLONE_DIR"; do
  if is_amc_checkout "$candidate"; then
    REPO_ROOT="$candidate"
    echo "✓ Using auto-magic-calib checkout: $REPO_ROOT"
    break
  fi
done

if [ -z "$REPO_ROOT" ]; then
  if [ -n "$CURRENT_GIT_ROOT" ] && [ -d "$CURRENT_GIT_ROOT/tools/auto-magic-calib" ]; then
    echo "Found $CURRENT_GIT_ROOT/tools/auto-magic-calib, but it is not an initialized AMC checkout."
    echo "If running from the DeepStream repository root:"
    echo "  git submodule update --init tools/auto-magic-calib"
  fi

  # Nothing usable on disk — STOP and ask the user for confirmation using the
  # host's question mechanism; if none is available, ask in chat and wait.
  # Do NOT clone silently from this block or clone over a tracked submodule path.
  echo "No usable auto-magic-calib checkout found. Ask the user for confirmation:"
  echo "  Clone $REPO_URL into $DEFAULT_CLONE_DIR? [y/N]"
  echo "On 'y' — run: git clone \"$REPO_URL\" \"$DEFAULT_CLONE_DIR\""
  exit 1
fi

cd "$REPO_ROOT"
export REPO_ROOT
echo "REPO_ROOT=$REPO_ROOT"

Agent 提示：切勿在未告知的情况下克隆。优先使用已初始化的 DeepStream tools/auto-magic-calib；不要克隆到该子模块路径。如果该路径存在但为空，请用户运行 git submodule update --init tools/auto-magic-calib。如果用户提供了其他 AMC 路径，请使用该路径。

步骤 0c：安装 Python venv（仅限新系统）

在全新系统上，可能尚未安装 pip 和 python3-venv。请先安装：

# Create a venv for HuggingFace CLI (project-local preferred)
REPO_DIR="$(git rev-parse --show-toplevel 2>/dev/null || pwd)"
HF_VENV="${REPO_DIR}/venv"
python3 -m venv "$HF_VENV" 2>/dev/null || {
  echo "ERROR: python3-venv not available." >&2
  echo "Install it manually: sudo apt install -y python3-venv python3-pip" >&2
  exit 1
}

# Install HuggingFace hub (needed for VGGT download)
"$HF_VENV/bin/pip" install --upgrade pip huggingface_hub

注意：如果已有包含 hf 的 venv，请跳过此步骤（检查仓库根目录下的 venv/bin/hf 或 ~/venv/amc/bin/hf）。

步骤 1：登录 NGC

使用主机的提问机制向用户索取 NGC API 密钥；如果没有可用的提问机制，请在聊天中询问并等待。然后运行：

echo "<NGC_API_KEY>" | docker login nvcr.io --username '$oauthtoken' --password-stdin
echo "✓ NGC authentication complete"

步骤 2：下载 VGGT 模型（如果尚未存在）

export REPO_ROOT=$(git rev-parse --show-toplevel)
cd "$REPO_ROOT"

if [ -f "models/vggt/vggt_1B_commercial.pt" ]; then
  echo "✓ VGGT model already present"
else
  echo "✗ VGGT model not found"
  echo "Options:"
  echo "  1. Continue without VGGT (AMC only - sufficient for most use cases)"
  echo "  2. Download VGGT model (~4.7GB, requires HuggingFace account)"
fi

下载 VGGT：请用户在 https://huggingface.co/facebook/VGGT-1B-Commercial 接受许可，并使用主机的提问机制索取读取令牌：https://huggingface.co/settings/tokens。通过 HF_TOKEN 传入令牌，以免其显示在 ps 输出中：

REPO_DIR="$(git rev-parse --show-toplevel 2>/dev/null || pwd)"
cd "$REPO_DIR"

# Find the HuggingFace CLI binary (named 'hf', not 'huggingface-cli')
HF_BIN="$(find "$REPO_DIR/venv" ~/venv/amc -name hf -type f 2>/dev/null | head -1)"
{ [ -z "$HF_BIN" ] || [ ! -x "$HF_BIN" ]; } && { echo "ERROR: hf binary not found or not executable; install the hf CLI (Step 0c) or set HF_BIN" >&2; exit 1; }

# Do NOT use --token on the command line (leaks via ps/argv). The HF CLI
# reads HF_TOKEN from the environment automatically.
HF_TOKEN="<HF_TOKEN>" "$HF_BIN" download facebook/VGGT-1B-Commercial \
  --local-dir models/vggt/

# Verify
ls -lh models/vggt/vggt_1B_commercial.pt
# Should show ~4.7GB file

重要：在对 models 目录设置 chown 1000:1000 之前下载模型——下载期间当前用户需要写入权限。下载完成后，在步骤 4 中设置权限。

步骤 3：配置 Compose 环境变量

Compose 环境文件用于控制端口和路径。启动前请更新该文件：

cd $REPO_ROOT/compose

# Find available backend port (8000-8009)
for port in {8000..8009}; do
  if ! lsof -Pi :$port -sTCP:LISTEN -t >/dev/null 2>&1; then
    MS_PORT=$port
    echo "Using backend port: $MS_PORT"
    break
  fi
done
[ -z "$MS_PORT" ] && { echo "ERROR: no free backend port in 8000-8009; free one or widen the range." >&2; exit 1; }

# Find available UI port (5000-5009)
for port in {5000..5009}; do
  if ! lsof -Pi :$port -sTCP:LISTEN -t >/dev/null 2>&1; then
    UI_PORT=$port
    echo "Using UI port: $UI_PORT"
    break
  fi
done
[ -z "$UI_PORT" ] && { echo "ERROR: no free UI port in 5000-5009; free one or widen the range." >&2; exit 1; }

# Get host IP
HOST_IP=$(hostname -I | awk '{print $1}')
echo "Host IP: $HOST_IP"

# Preserve existing keys and restrict permissions on the Compose environment file.
COMPOSE_ENV_BASENAME="env"
ENV_FILE=".${COMPOSE_ENV_BASENAME}"
if [ -f "$ENV_FILE" ]; then
  BACKUP="${ENV_FILE}.bak.$(date +%s)"
  cp "$ENV_FILE" "$BACKUP"
  chmod 600 "$BACKUP"
fi
touch "$ENV_FILE"
chmod 600 "$ENV_FILE"
set_env_key() {
  local k="$1" v="$2"
  if grep -qE "^${k}=" "$ENV_FILE"; then
    sed -i "s|^${k}=.*|${k}=${v}|" "$ENV_FILE"
  else
    echo "${k}=${v}" >> "$ENV_FILE"
  fi
}
set_env_key AUTO_MAGIC_CALIB_MS_PORT "${MS_PORT}"
set_env_key AUTO_MAGIC_CALIB_UI_PORT "${UI_PORT}"
set_env_key PROJECT_DIR "../../projects"
set_env_key MODEL_DIR "../../models"
set_env_key HOST_IP "${HOST_IP}"

# Keep timestamped Compose environment backups out of git.
GITIGNORE="$REPO_ROOT/.gitignore"
touch "$GITIGNORE"
BACKUP_PATTERN="compose/${ENV_FILE}.bak.*"
grep -qxF "$BACKUP_PATTERN" "$GITIGNORE" || echo "$BACKUP_PATTERN" >> "$GITIGNORE"

echo "✓ Compose environment file updated"
cat "$ENV_FILE"

重要：HOST_IP 必须是机器的网络 IP（不能是 localhost），这样 UI 容器才能通过浏览器访问后端。

可选：仅当 VGGT 模型挂载在非默认容器路径时，才设置 VGGT_MODEL_PATH；MS 容器中的默认路径为 /tmp/vggt_model/vggt_1B_commercial.pt。

RTSP 标定的可选操作：启动后使用 skills/amc-run-rtsp-calibration/SKILL.md。该技能会验证 VIOS 是否可访问；必要时使用临时 compose 覆盖文件重新启动微服务，以导出 VIOS_BASE_URL，且不会更改已签入的 compose 文件。

步骤 4：设置目录权限

容器以 UID/GID 1000 运行。projects 和 models 目录必须归该 UID 所有，容器才能正常读写：

cd "$REPO_ROOT"

# Create projects directory if it doesn't exist
mkdir -p projects

# Set ownership (required for containers to write calibration outputs).
# Do this AFTER VGGT download is complete (current user needs write access during download).
# Get explicit user confirmation before running sudo chown — it recursively changes
# ownership of $REPO_ROOT/projects and $REPO_ROOT/models to UID/GID 1000.
[ -d projects ] && [ -d models ] || {
  echo "ERROR: expected projects/ and models/ under $REPO_ROOT" >&2; exit 1;
}
echo "About to chown -R 1000:1000 on:"
echo "  $REPO_ROOT/projects"
echo "  $REPO_ROOT/models"
echo "(required because containers run as UID 1000). Confirm before proceeding."
sudo chown 1000:1000 -R projects
sudo chown 1000:1000 -R models

echo "✓ Permissions set"

步骤 5：启动服务

拉取镜像前，如果步骤 1 中通过身份验证的 NGC 密钥无法实际访问发布镜像，应立即失败；否则 docker compose up 可能会在完成部分操作后因 401/403 错误而中止。

cd $REPO_ROOT/compose

# Fail-fast image-access check: confirm the NGC key can reach every release
# image BEFORE pulling. `docker manifest inspect` checks registry access without
# downloading layers, and the image list is read from the resolved compose so it
# tracks the release tag automatically.
IMAGES=$(docker compose config --images | sort -u)
[ -z "$IMAGES" ] && { echo "ERROR: no images resolved from compose — check the Compose environment settings and chosen profile." >&2; exit 1; }
for img in $IMAGES; do
  echo "Checking access: $img"
  if ! docker manifest inspect "$img" >/dev/null 2>&1; then
    echo "NGC login succeeded, but this key cannot access the required image:" >&2
    echo "  $img" >&2
    echo "Provide an NGC key with access to this image's namespace, then re-run Step 1 (login) and retry." >&2
    exit 1
  fi
done

# Start all services (images pulled automatically on first run)
docker compose up -d

# Check containers are running
docker compose ps

具体镜像标签会随发布版本变化；请从当前启用的 compose 文件中读取标签，不要硬编码版本号。

步骤 6：确认服务正在运行

# Read ports from the Compose environment file.
COMPOSE_ENV_BASENAME="env"
COMPOSE_ENV_FILE="$REPO_ROOT/compose/.${COMPOSE_ENV_BASENAME}"
MS_PORT=$(grep AUTO_MAGIC_CALIB_MS_PORT "$COMPOSE_ENV_FILE" | cut -d= -f2)
UI_PORT=$(grep AUTO_MAGIC_CALIB_UI_PORT "$COMPOSE_ENV_FILE" | cut -d= -f2)
HOST_IP=$(grep HOST_IP "$COMPOSE_ENV_FILE" | cut -d= -f2)

# Wait for microservice readiness. Cold image pulls or first startup can need
# extra time after `docker compose up -d` returns.
READY_URL="http://localhost:${MS_PORT}/v1/ready"
echo "Waiting for microservice readiness at ${READY_URL} ..."
ready_response=""
for attempt in $(seq 1 24); do
  if ready_response=$(curl -fsS --max-time 5 "${READY_URL}" 2>/dev/null) && \
     echo "${ready_response}" | grep -q '"code"[[:space:]]*:[[:space:]]*0'; then
    echo "Microservice ready: ${ready_response}"
    break
  fi
  if [ "${attempt}" -lt 24 ]; then
    printf "  [%02d/24] Microservice not ready yet; retrying in 5s...\n" "${attempt}"
    sleep 5
  fi
done

if ! echo "${ready_response}" | grep -q '"code"[[:space:]]*:[[:space:]]*0'; then
  echo "ERROR: microservice did not report ready within 120 seconds: ${READY_URL}" >&2
  echo "Check status and logs:" >&2
  echo "  cd ${REPO_ROOT}/compose && docker compose ps" >&2
  echo "  cd ${REPO_ROOT}/compose && docker compose logs auto-magic-calib-ms" >&2
  exit 1
fi

# Check UI is serving
UI_STATUS=$(curl -s -o /dev/null -w "%{http_code}" --max-time 5 "http://localhost:${UI_PORT}")
if [ "${UI_STATUS}" != "200" ]; then
  echo "ERROR: Web UI returned HTTP ${UI_STATUS}; check docker compose ps and UI logs." >&2
  exit 1
fi
echo "Web UI ready: HTTP ${UI_STATUS}"

echo "Microservice: http://${HOST_IP}:${MS_PORT}"
echo "Web UI:       http://${HOST_IP}:${UI_PORT}"

成功标准

• docker compose ps 显示 MS 和 UI 容器状态为 Up；MS 应处于健康状态。
• /v1/ready 返回 code:0，且步骤 6 会打印微服务和 UI 的 URL。
• 可以通过浏览器访问 http://<HOST_IP>:<AUTO_MAGIC_CALIB_UI_PORT>。
• 项目保存在 $REPO_ROOT/projects/ 下。

故障排查

问题	解决方法
Docker 权限被拒绝	请用户运行 sudo usermod -aG docker $USER && newgrp docker，然后重试 docker ps。
docker login 被拒绝	请提供当前有效的 NGC 密钥，然后重新登录。
无法访问所需镜像	该密钥没有镜像命名空间的访问权限；请提供具有访问权限的密钥，然后重试步骤 1 和步骤 5。
缺少 python3 -m venv、pip 或 hf	安装 python3-venv/python3-pip；HF 二进制文件名为 hf。
VGGT 权限错误	在执行 chown 1000:1000 之前下载 VGGT；如需恢复，请还原 models/ 的用户所有权并重新下载。
端口被占用	在 8000-8009 中选择一个空闲的 MS 端口，并在 5000-5009 中选择一个空闲的 UI 端口，然后更新 Compose 环境文件。
就绪状态超时或容器已退出	运行 cd $REPO_ROOT/compose && docker compose ps，并检查 docker compose logs auto-magic-calib-ms。
项目/模型权限被拒绝	仅针对 projects/ 或 models/ 重新执行步骤 4。
UI 无法连接后端	确认 Compose 环境文件中的 HOST_IP 是机器的网络 IP，而不是 localhost。
GPU 不可用	使用 docker run --rm --runtime=nvidia --gpus all ubuntu:20.04 nvidia-smi 验证 NVIDIA 运行时。

常见修复方法：

cd $REPO_ROOT/compose

# View logs
docker compose logs -f

# View logs for specific service
docker compose logs -f auto-magic-calib-ms

# Restart all services
docker compose restart

# Stop and remove containers
docker compose down

# Update Compose environment settings and relaunch
docker compose up -d

停止服务

cd $REPO_ROOT/compose

# Stop all services (containers removed, data persisted)
docker compose down

# Stop and remove volumes
docker compose down -v

相关技能

• skills/amc-run-sample-calibration/SKILL.md - 使用随附的示例数据集对运行中的服务栈进行健全性检查
• skills/amc-run-video-calibration/SKILL.md - 通过 REST API 使用自己的预录制 MP4 文件进行校准
• skills/amc-run-rtsp-calibration/SKILL.md - 通过 VIOS 捕获实时 RTSP 流进行校准