diff --git a/.github/workflows/providers-build.yml b/.github/workflows/providers-build.yml index 117c8b6d2..23257d7dc 100644 --- a/.github/workflows/providers-build.yml +++ b/.github/workflows/providers-build.yml @@ -107,3 +107,41 @@ jobs: - name: Build a single provider run: | USE_COPY_NOT_MOUNT=true LLAMA_STACK_DIR=. uv run llama stack build --image-type venv --image-name test --providers inference=remote::ollama + + build-custom-container-distribution: + runs-on: ubuntu-latest + steps: + - name: Checkout repository + uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2 + + - name: Set up Python + uses: actions/setup-python@8d9ed9ac5c53483de85588cdf95a591a75ab9f55 # v5.5.0 + with: + python-version: '3.10' + + - name: Install uv + uses: astral-sh/setup-uv@0c5e2b8115b80b4c7c5ddf6ffdd634974642d182 # v5.4.1 + with: + python-version: "3.10" + + - name: Install LlamaStack + run: | + uv venv + source .venv/bin/activate + uv pip install -e . + + - name: Build a single provider + run: | + yq -i '.image_type = "container"' llama_stack/templates/dev/build.yaml + yq -i '.image_name = "test"' llama_stack/templates/dev/build.yaml + USE_COPY_NOT_MOUNT=true LLAMA_STACK_DIR=. uv run llama stack build --config llama_stack/templates/dev/build.yaml + + - name: Inspect the container image entrypoint + run: | + IMAGE_ID=$(docker images --format "{{.Repository}}:{{.Tag}}" | head -n 1) + entrypoint=$(docker inspect --format '{{ .Config.Entrypoint }}' $IMAGE_ID) + echo "Entrypoint: $entrypoint" + if [ "$entrypoint" != "[python -m llama_stack.distribution.server.server --config /app/run.yaml]" ]; then + echo "Entrypoint is not correct" + exit 1 + fi diff --git a/llama_stack/cli/stack/_build.py b/llama_stack/cli/stack/_build.py index 26c09af4e..80ab0631b 100644 --- a/llama_stack/cli/stack/_build.py +++ b/llama_stack/cli/stack/_build.py @@ -317,11 +317,15 @@ def _generate_run_config( to_write = json.loads(run_config.model_dump_json()) f.write(yaml.dump(to_write, sort_keys=False)) - # this path is only invoked when no template is provided - cprint( - f"You can now run your stack with `llama stack run {run_config_file}`", - color="green", - ) + # Only print this message for non-container builds since it will be displayed before the + # container is built + # For non-container builds, the run.yaml is generated at the very end of the build process so it + # makes sense to display this message + if build_config.image_type != LlamaStackImageType.CONTAINER.value: + cprint( + f"You can now run your stack with `llama stack run {run_config_file}`", + color="green", + ) return run_config_file @@ -355,6 +359,13 @@ def _run_stack_build_command_from_build_config( build_file_path = build_dir / f"{image_name}-build.yaml" os.makedirs(build_dir, exist_ok=True) + run_config_file = None + # Generate the run.yaml so it can be included in the container image with the proper entrypoint + # Only do this if we're building a container image and we're not using a template + if build_config.image_type == LlamaStackImageType.CONTAINER.value and not template_name and config_path: + cprint("Generating run.yaml file", color="green") + run_config_file = _generate_run_config(build_config, build_dir, image_name) + with open(build_file_path, "w") as f: to_write = json.loads(build_config.model_dump_json()) f.write(yaml.dump(to_write, sort_keys=False)) @@ -364,6 +375,7 @@ def _run_stack_build_command_from_build_config( build_file_path, image_name, template_or_config=template_name or config_path or str(build_file_path), + run_config=run_config_file, ) if return_code != 0: raise RuntimeError(f"Failed to build image {image_name}") diff --git a/llama_stack/distribution/build.py b/llama_stack/distribution/build.py index 5b61ae081..9664449f3 100644 --- a/llama_stack/distribution/build.py +++ b/llama_stack/distribution/build.py @@ -93,6 +93,7 @@ def build_image( build_file_path: Path, image_name: str, template_or_config: str, + run_config: str | None = None, ): container_base = build_config.distribution_spec.container_image or "python:3.10-slim" @@ -108,6 +109,11 @@ def build_image( container_base, " ".join(normal_deps), ] + + # When building from a config file (not a template), include the run config path in the + # build arguments + if run_config is not None: + args.append(run_config) elif build_config.image_type == LlamaStackImageType.CONDA.value: script = str(importlib.resources.files("llama_stack") / "distribution/build_conda_env.sh") args = [ diff --git a/llama_stack/distribution/build_container.sh b/llama_stack/distribution/build_container.sh index fb4780432..ad316d45e 100755 --- a/llama_stack/distribution/build_container.sh +++ b/llama_stack/distribution/build_container.sh @@ -19,12 +19,16 @@ UV_HTTP_TIMEOUT=${UV_HTTP_TIMEOUT:-500} # mounting is not supported by docker buildx, so we use COPY instead USE_COPY_NOT_MOUNT=${USE_COPY_NOT_MOUNT:-} +# Path to the run.yaml file in the container +RUN_CONFIG_PATH=/app/run.yaml + +BUILD_CONTEXT_DIR=$(pwd) + if [ "$#" -lt 4 ]; then # This only works for templates - echo "Usage: $0 []" >&2 + echo "Usage: $0 [] []" >&2 exit 1 fi - set -euo pipefail template_or_config="$1" @@ -35,8 +39,27 @@ container_base="$1" shift pip_dependencies="$1" shift -special_pip_deps="${1:-}" +# Handle optional arguments +run_config="" +special_pip_deps="" + +# Check if there are more arguments +# The logics is becoming cumbersom, we should refactor it if we can do better +if [ $# -gt 0 ]; then + # Check if the argument ends with .yaml + if [[ "$1" == *.yaml ]]; then + run_config="$1" + shift + # If there's another argument after .yaml, it must be special_pip_deps + if [ $# -gt 0 ]; then + special_pip_deps="$1" + fi + else + # If it's not .yaml, it must be special_pip_deps + special_pip_deps="$1" + fi +fi # Define color codes RED='\033[0;31m' @@ -75,7 +98,7 @@ WORKDIR /app # We install the Python 3.11 dev headers and build tools so that any # C‑extension wheels (e.g. polyleven, faiss‑cpu) can compile successfully. -RUN dnf -y update && dnf install -y iputils net-tools wget \ +RUN dnf -y update && dnf install -y iputils git net-tools wget \ vim-minimal python3.11 python3.11-pip python3.11-wheel \ python3.11-setuptools python3.11-devel gcc make && \ ln -s /bin/pip3.11 /bin/pip && ln -s /bin/python3.11 /bin/python && dnf clean all @@ -119,6 +142,45 @@ EOF done fi +# Function to get Python command +get_python_cmd() { + if is_command_available python; then + echo "python" + elif is_command_available python3; then + echo "python3" + else + echo "Error: Neither python nor python3 is installed. Please install Python to continue." >&2 + exit 1 + fi +} + +if [ -n "$run_config" ]; then + # Copy the run config to the build context since it's an absolute path + cp "$run_config" "$BUILD_CONTEXT_DIR/run.yaml" + add_to_container << EOF +COPY run.yaml $RUN_CONFIG_PATH +EOF + + # Parse the run.yaml configuration to identify external provider directories + # If external providers are specified, copy their directory to the container + # and update the configuration to reference the new container path + python_cmd=$(get_python_cmd) + external_providers_dir=$($python_cmd -c "import yaml; config = yaml.safe_load(open('$run_config')); print(config.get('external_providers_dir') or '')") + if [ -n "$external_providers_dir" ]; then + echo "Copying external providers directory: $external_providers_dir" + add_to_container << EOF +COPY $external_providers_dir /app/providers.d +EOF + # Edit the run.yaml file to change the external_providers_dir to /app/providers.d + if [ "$(uname)" = "Darwin" ]; then + sed -i.bak -e 's|external_providers_dir:.*|external_providers_dir: /app/providers.d|' "$BUILD_CONTEXT_DIR/run.yaml" + rm -f "$BUILD_CONTEXT_DIR/run.yaml.bak" + else + sed -i 's|external_providers_dir:.*|external_providers_dir: /app/providers.d|' "$BUILD_CONTEXT_DIR/run.yaml" + fi + fi +fi + stack_mount="/app/llama-stack-source" client_mount="/app/llama-stack-client-source" @@ -178,15 +240,16 @@ fi RUN pip uninstall -y uv EOF -# if template_or_config ends with .yaml, it is not a template and we should not use the --template flag -if [[ "$template_or_config" != *.yaml ]]; then +# If a run config is provided, we use the --config flag +if [[ -n "$run_config" ]]; then + add_to_container << EOF +ENTRYPOINT ["python", "-m", "llama_stack.distribution.server.server", "--config", "$RUN_CONFIG_PATH"] +EOF +# If a template is provided (not a yaml file), we use the --template flag +elif [[ "$template_or_config" != *.yaml ]]; then add_to_container << EOF ENTRYPOINT ["python", "-m", "llama_stack.distribution.server.server", "--template", "$template_or_config"] EOF -else - add_to_container << EOF -ENTRYPOINT ["python", "-m", "llama_stack.distribution.server.server"] -EOF fi # Add other require item commands genearic to all containers @@ -258,9 +321,10 @@ $CONTAINER_BINARY build \ "${CLI_ARGS[@]}" \ -t "$image_tag" \ -f "$TEMP_DIR/Containerfile" \ - "." + "$BUILD_CONTEXT_DIR" # clean up tmp/configs +rm -f "$BUILD_CONTEXT_DIR/run.yaml" set +x echo "Success!" diff --git a/tests/unit/distribution/test_build_path.py b/tests/unit/distribution/test_build_path.py index a913bd88b..555cdda4a 100644 --- a/tests/unit/distribution/test_build_path.py +++ b/tests/unit/distribution/test_build_path.py @@ -16,8 +16,9 @@ from llama_stack.distribution.utils.image_types import LlamaStackImageType def test_container_build_passes_path(monkeypatch, tmp_path): called_with = {} - def spy_build_image(cfg, build_file_path, image_name, template_or_config): + def spy_build_image(cfg, build_file_path, image_name, template_or_config, run_config=None): called_with["path"] = template_or_config + called_with["run_config"] = run_config return 0 monkeypatch.setattr( @@ -36,3 +37,4 @@ def test_container_build_passes_path(monkeypatch, tmp_path): assert "path" in called_with assert isinstance(called_with["path"], str) assert Path(called_with["path"]).exists() + assert called_with["run_config"] is None