feat: include run.yaml in the container image (#2005)

As part of the build process, we now include the generated run.yaml
(based of the provided build configuration file) into the container. We
updated the entrypoint to use this run configuration as well.

Given this simple distribution configuration:

```
# build.yaml
version: '2'
distribution_spec:
  description: Use (an external) Ollama server for running LLM inference
  providers:
    inference:
    - remote::ollama
    vector_io:
    - inline::faiss
    safety:
    - inline::llama-guard
    agents:
    - inline::meta-reference
    telemetry:
    - inline::meta-reference
    eval:
    - inline::meta-reference
    datasetio:
    - remote::huggingface
    - inline::localfs
    scoring:
    - inline::basic
    - inline::llm-as-judge
    - inline::braintrust
    tool_runtime:
    - remote::brave-search
    - remote::tavily-search
    - inline::code-interpreter
    - inline::rag-runtime
    - remote::model-context-protocol
    - remote::wolfram-alpha
  container_image: "registry.access.redhat.com/ubi9"
image_type: container
image_name: test
```

Build it:
```
llama stack build --config build.yaml
```

Run it:

```
podman run --rm \
         -p 8321:8321 \
         -e OLLAMA_URL=http://host.containers.internal:11434 \
         --name llama-stack-server \
         localhost/leseb-test:0.2.2
```

Signed-off-by: Sébastien Han <seb@redhat.com>

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}")

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 = [

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 <template_or_config> <image_name> <container_base> <pip_dependencies> [<special_pip_deps>]" >&2
+  echo "Usage: $0 <template_or_config> <image_name> <container_base> <pip_dependencies> [<run_config>] [<special_pip_deps>]" >&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!"