Introduction
MiniCPM_o_2.6-FlagOS-NVIDIA provides an all-in-one deployment solution, enabling execution of MiniCPM_o_2.6 on NVIDIA GPUs. As the first-generation release for the NVIDIA-H100, this package delivers two key features:
- Comprehensive Integration:
- Integrated with FlagScale (https://github.com/FlagOpen/FlagScale).
- Open-source inference execution code, preconfigured with all necessary software and hardware settings.
- Pre-built Docker image for rapid deployment on NVIDIA-H100.
- Consistency Validation:
- Evaluation tests verifying consistency of results between the official and ours.
Technical Summary
Serving Engine
We use FlagScale as the serving engine to improve the portability of distributed inference.
FlagScale is an end-to-end framework for large models across multiple chips, maximizing computational resource efficiency while ensuring model effectiveness. It ensures both ease of use and high performance for users when deploying models across different chip architectures:
- One-Click Service Deployment: FlagScale provides a unified and simple command execution mechanism, allowing users to fast deploy services seamlessly across various hardware platforms using the same command. This significantly reduces the entry barrier and enhances user experience.
- Automated Deployment Optimization: FlagScale automatically optimizes distributed parallel strategies based on the computational capabilities of different AI chips, ensuring optimal resource allocation and efficient utilization, thereby improving overall deployment performance.
- Automatic Operator Library Switching: Leveraging FlagScale's unified Runner mechanism and deep integration with FlagGems, users can seamlessly switch to the FlagGems operator library for inference by simply adding environment variables in the configuration file.
Triton Support
We validate the execution of DeepSeed-R1 model with a Triton-based operator library as a PyTorch alternative.
We use a variety of Triton-implemented operation kernelsโapproximately 70%โto run the MiniCPM_o_2.6 model. These kernels come from two main sources:
Most Triton kernels are provided by FlagGems (https://github.com/FlagOpen/FlagGems). You can enable FlagGems kernels by setting the environment variable USE_FLAGGEMS. For more details, please refer to the "How to Run Locally" section.
Also included are Triton kernels from vLLM.
Bundle Download
Usage | Nvidia | |
---|---|---|
Basic Image | basic software environment that supports model running | 'docker pull flagrelease-registry.cn-beijing.cr.aliyuncs.com/flagrelease/flagrelease:deepseek-flagos-nvidia |
Evaluation Results
Benchmark Result
Metrics | MiniCPM_o_2.6-H100-CUDA | MiniCPM_o_2.6-H100-FlagOS |
---|---|---|
mmmu_val | 48.11 | 48.33 |
math_vision_test | 22.89 | 22.30 |
ocrbench_test | 85.80 | 85.70 |
blink_val | 54.87 | 55.81 |
mmmvet_v2 | 57.66 | 59.03 |
mmmu_pro_vision_test | 70.46 | 69.77 |
mmmu_pro_standard_test | 30.46 | 30.81 |
cmmmu_val | 39.33 | 39.33 |
cii_bench_test | 50.07 | 50.33 |
How to Run Locally
๐ Getting Started
Download open-source weights
pip install modelscope
modelscope download --model <Model Name> --local_dir <Cache Path>
Download the FlagOS image
docker pull <IMAGE>
Start the inference service
docker run -itd --name flagrelease_nv --privileged --gpus all --net=host --ipc=host --device=/dev/infiniband --shm-size 512g --ulimit memlock=-1 -v <CKPT_PATH>:<CKPT_PATH> flagrelease-registry.cn-beijing.cr.aliyuncs.com/flagrelease/flagrelease:deepseek-flagos-nvidia /bin/bash
docker exec -it flagrelease_nv /bin/bash
conda activate flagscale-inference
Download and install FlagGems
git clone https://github.com/FlagOpen/FlagGems.git
cd FlagGems
pip install .
cd ../
Modify the configuration
cd FlagScale/examples/minicpm_o_2.6/conf
# Modify the configuration in config_minicpm_o_2.6.yaml
defaults:
- _self_
- serve: minicpm_o_2.6
experiment:
exp_name: minicpm_o_2.6
exp_dir: outputs/${experiment.exp_name}
task:
type: serve
deploy:
use_fs_serve: false
runner:
ssh_port: 22
envs:
CUDA_DEVICE_MAX_CONNECTIONS: 1
cmds:
before_start: source /root/miniconda3/bin/activate flagscale-inference && export USE_FLAGGEMS=1
action: run
hydra:
run:
dir: ${experiment.exp_dir}/hydra
cd FlagScale/examples/minicpm_o_2.6/conf/serve
# Modify the configuration in minicpm_o_2.6.yaml
- serve_id: vllm_model
engine: vllm
engine_args:
model: /models/MiniCPM_o_2 # path of weight of DeepSeek-R1-Distill-Qwen-32B
served_model_name: minicpmo26-flagos
tensor_parallel_size: 1
pipeline_parallel_size: 1
gpu_memory_utilization: 0.9
max_num_seqs: 256
limit_mm_per_prompt: image=18
port: 9010
trust_remote_code: true
enable_chunked_prefill: true
# install flagscale
cd FlagScale/
pip install .
#ใVerifiable on a single machineใ
Serve
flagscale serve <Model>
Contributing
We warmly welcome global developers to join us:
- Submit Issues to report problems
- Create Pull Requests to contribute code
- Improve technical documentation
- Expand hardware adaptation support
๐ Contact Us
Scan the QR code below to add our WeChat group
send "FlagRelease"
License
This project and related model weights are licensed under the MIT License.