FlagRelease/Moonlight-16B-A3B-hygon-FlagOS

Introduction

Moonlight-16B-A3B is a high-performance large language model optimized for the Hygon DCU platform. Built on a Mixture of Experts (MoE) architecture, the model features a total of 16 billion parameters with approximately 3 billion active parameters per inference, striking an optimal balance between high performance and efficient inference throughput. Deeply optimized for Hygon GPU hardware, Moonlight-16B-A3B supports the vLLM inference framework, making it well-suited for large-scale deployment scenarios.

Integrated Deployment

• —Out-of-the-box inference scripts with pre-configured hardware and software parameters
• —Released FlagOS-Hygon container image supporting deployment within minutes

Consistency Validation

• —Rigorously evaluated through benchmark testing: Performance and results from the FlagOS software stack are compared against native stacks on multiple public benchmarks.

Evaluation Results

Benchmark Result

User Guide

Environment Setup

Operation Steps

Download FlagOS Image

docker pull harbor.baai.ac.cn/external-cooperation/moonlight-16b-a3b-hygon-tree_0.5.0_hcu3.0-gems_5.0.2-vllm_0.13.0-plugin_0.1.1-cx_none-python_3.10.12-torch_2.9.0-das.opt1.dtk2604.20260206.g275d08c2-pcp_hygon-dpu_hygon-x86_64-driver_1.11.0:202607071400

Download Open-source Model Weights

pip install modelscope

modelscope download \
  --model FlagRelease/Moonlight-16B-A3B-hygon-FlagOS \
  --local_dir /data/Moonlight-16B-A3B-hygon-FlagOS

Start the Container

docker run -itd \
  --name Moonlight-16B-A3B-hygon-FlagOS \
  --device=/dev/kfd \
  --device=/dev/mkfd \
  --device=/dev/dri \
  --group-add video \
  -v /opt/hyhal:/opt/hyhal \
  --ipc=host \
  --ulimit memlock=-1 \
  --ulimit stack=67108864 \
  --network host \
  -v /data/Moonlight-16B-A3B-hygon-FlagOS:/data/Moonlight-16B-A3B-hygon-FlagOS \
  harbor.baai.ac.cn/external-cooperation/moonlight-16b-a3b-hygon-tree_0.5.0_hcu3.0-gems_5.0.2-vllm_0.13.0-plugin_0.1.1-cx_none-python_3.10.12-torch_2.9.0-das.opt1.dtk2604.20260206.g275d08c2-pcp_hygon-dpu_hygon-x86_64-driver_1.11.0:202607071400 \
  sleep infinity

Enter the Container

docker exec -it Moonlight-16B-A3B-hygon-FlagOS /bin/bash

Start the Server

export VLLM_FL_FLAGOS_BLACKLIST="mul,copy_"
export HIP_VISIBLE_DEVICES=4,5
export VLLM_PLUGINS=fl
export TRITON_ALL_BLOCKS_PARALLEL=1
export USE_FLAGGEMS=1
nohup vllm serve /data/Moonlight-16B-A3B-hygon-FlagOS \
    --served-model-name Moonlight-16B-A3B-hygon-FlagOS \
    --port 8003 \
    --trust-remote-code \
    --max-model-len 8192 \
    --gpu-memory-utilization 0.9 \
    --tensor-parallel-size 2 \
    --enforce-eager \
    > flagos_server.log 2>&1 &

Service Invocation

Invocation Script

curl http://localhost:8003/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Moonlight-16B-A3B-hygon-FlagOS",
    "messages": [{"role": "user", "content": "hello"}]
  }'

AnythingLLM Integration Guide

1. Download & Install

• —Visit the official site: https://anythingllm.com/
• —Choose the appropriate version for your OS (Windows/macOS/Linux)
• —Follow the installation wizard to complete the setup

2. Configuration

• —Launch AnythingLLM
• —Open settings (bottom left, fourth tab)
• —Configure core LLM parameters:
• —API URL: http://localhost:8003/v1
• —Model: moonlight-flagos
• —Click "Save Settings" to apply changes

3. Model Interaction

• —After model loading is complete:
• —Click "New Conversation"
• —Enter your question (e.g., "Explain the basics of quantum computing")
• —Click the send button to get a response

Technical Overview

FlagOS is a fully open-source system software stack designed to unify the "model–system–chip" layers and foster an open, collaborative ecosystem. It enables a "develop once, run anywhere" workflow across diverse AI accelerators, unlocking hardware performance, eliminating fragmentation among vendor-specific software stacks, and substantially lowering the cost of porting and maintaining AI workloads.

With core technologies such as FlagScale, together with vllm-plugin-fl, distributed training/inference framework, FlagGems universal operator library, FlagCX communication library, and FlagTree unified compiler, the FlagRelease platform leverages the FlagOS stack to automatically produce and release various combinations of <chip + open-source model>.

This enables efficient and automated model migration across diverse chips, opening a new chapter for large model deployment and application.

FlagGems

FlagGems is a high-performance, generic operator library implemented in Triton language. It is built on a collection of backend-neutral kernels that aims to accelerate LLM training and inference across diverse hardware platforms.

FlagTree

FlagTree is an open-source unified compiler for multiple AI chips. It provides unified compilation capabilities across multiple backends and rapidly implements single-repository multi-backend support.

FlagScale and vllm-plugin-fl

FlagScale is a comprehensive toolkit designed to support the entire lifecycle of large models. It integrates capabilities from Megatron-LM and vLLM to provide an end-to-end solution for training and inference.

vllm-plugin-fl is a vLLM plugin built on the FlagOS unified multi-chip backend.

FlagCX

FlagCX is a scalable and adaptive cross-chip communication library for distributed AI workloads.

FlagEval Evaluation Framework

FlagEval is a comprehensive evaluation system and open platform for large models. It supports large-scale benchmark evaluation across NLP, CV, Audio, and Multimodal tasks.

Contributing

We warmly welcome global developers to join us:

1. 1.Submit Issues to report problems
2. 2.Create Pull Requests to contribute code
3. 3.Improve technical documentation
4. 4.Expand hardware adaptation support

License

The model weights are derived from moonshotai/Moonlight-16B-A3B and are open-sourced under the Apache License 2.0: https://www.apache.org/licenses/LICENSE-2.0.txt