Implementation:Mlc ai Mlc llm Package

Knowledge Sources	MLC-LLM
Domains	Deep_Learning, Mobile_Deployment
Last Updated	2026-02-09 00:00 GMT

Overview

Concrete tool for packaging compiled model libraries with platform-specific bindings into deployable mobile bundles provided by MLC-LLM.

Description

The package function in python/mlc_llm/interface/package.py is the top-level Python entry point for the MLC-LLM mobile packaging pipeline. It reads the mlc-package-config.json configuration file, compiles model libraries for the specified device target, validates the compiled libraries, builds the platform-specific bindings, and produces a complete output directory ready for integration into an iOS or Android application project.

The function orchestrates four major sub-operations:

build_model_library -- Iterates over each model entry in the configuration, downloads model artifacts from Hugging Face if needed, JIT-compiles model libraries for the target device, optionally copies model weights to the bundle directory, and generates the mlc-app-config.json runtime configuration file.
validate_model_lib -- Combines all compiled model libraries into a single static archive and validates that each model's TVM FFI symbols are present in the global symbol table.
build_android_binding -- (Android only) Moves the model library to the build directory, invokes prepare_libs.py to build the JNI bindings, and assembles the mlc4j Gradle package with source files, build configuration, and runtime assets.
build_iphone_binding -- (iOS only) Invokes prepare_libs.sh to build the static libraries and copies them to the output directory.

Usage

Use the package function when:

Building a complete mobile deployment package from a mlc-package-config.json configuration
Automating the end-to-end compilation and packaging pipeline in CI/CD
Producing the output artifacts needed to integrate MLC-LLM into an Xcode or Android Studio project

Code Reference

Source Location

Repository: MLC-LLM
File: python/mlc_llm/interface/package.py (Lines 323-373)

Signature

def package(
    package_config_path: Path,
    mlc_llm_source_dir: Path,
    output: Path,
) -> None:
    """Python entrypoint of package.

    Parameters
    ----------
    package_config_path : Path
        Path to the mlc-package-config.json file.
    mlc_llm_source_dir : Path
        Path to the MLC-LLM source directory.
    output : Path
        Path to the output directory for compiled libraries,
        bundled weights, and platform bindings.
    """

I/O Contract

Inputs

Name	Type	Required	Description
`package_config_path`	Path	Yes	Path to the `mlc-package-config.json` file specifying device, models, and compilation options
`mlc_llm_source_dir`	Path	Yes	Path to the root of the MLC-LLM source tree, used to locate platform-specific build scripts (`ios/prepare_libs.sh`, `android/mlc4j/prepare_libs.py`)
`output`	Path	Yes	Path to the output directory where compiled libraries, bundled weights, and platform bindings will be written

Outputs

Name	Type	Description
`output/bundle/mlc-app-config.json`	JSON file	Runtime configuration mapping model IDs to compiled library names and model URLs or local paths
`output/lib/libmodel_iphone.a`	Static library	Combined static archive of all compiled model libraries for iOS (when device is `iphone`)
`output/lib/libmodel_android.a`	Static library	Combined static archive of all compiled model libraries for Android (when device is `android`)
`output/lib/*.a` (iOS)	Static libraries	MLC-LLM and TVM runtime static libraries for iOS
`output/lib/mlc4j/` (Android)	Directory	Complete mlc4j Gradle package containing JNI source, build configuration, compiled shared libraries, and runtime assets
`output/bundle/<model_id>/`	Directory	Model weight files for models with `bundle_weight: true`

Usage Examples

Basic Usage (Python API)

from pathlib import Path
from mlc_llm.interface.package import package

package(
    package_config_path=Path("ios/MLCChat/mlc-package-config.json"),
    mlc_llm_source_dir=Path("/path/to/mlc-llm"),
    output=Path("dist"),
)

Command-Line Usage

# Package for iOS
mlc_llm package \
    --package-config ios/MLCChat/mlc-package-config.json \
    --mlc-llm-source-dir /path/to/mlc-llm \
    --output dist/

# Package for Android
mlc_llm package \
    --package-config android/MLCChat/mlc-package-config.json \
    --mlc-llm-source-dir /path/to/mlc-llm \
    --output dist/

Integration with Xcode (iOS)

# After running mlc_llm package, copy the output into the Xcode project
cp dist/lib/*.a ios/MLCChat/lib/
cp -r dist/bundle/ ios/MLCChat/bundle/
# Open the Xcode project and build
open ios/MLCChat.xcodeproj

Integration with Android Studio

# After running mlc_llm package, the mlc4j library is ready
# Copy the mlc4j output into the Android project dependencies
cp -r dist/lib/mlc4j/ android/MLCChat/mlc4j/
# Build the Android project
cd android/MLCChat && ./gradlew assembleRelease

Related Pages

Implements Principle

Principle:Mlc_ai_Mlc_llm_Model_Library_Packaging

Environment Links

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment