FlagRelease/DeepSeek-R1-FlagOS-Iluvatar-INT8

Introduction

DeepSeek-R1-FlagOS-Iluvatar-INT8 provides an all-in-one deployment solution, enabling execution of DeepSeek-R1 on Iluvatar GPUs. As the first-generation release for the ILUVATAR-BI150 series, this package delivers three key features:

1. Comprehensive Integration:
   • Integrated with FlagScale (https://github.com/FlagOpen/FlagScale).
   • Open-source inference execution code, preconfigured with all necessary software and hardware settings.
   • Verified model files, available on ModelScope (Model Link).
   • Pre-built Docker image for rapid deployment on Iluvatar.
2. INT8 Checkpoints:
   • INT8 checkpoints dequantized from the official DeepSeek-R1 FP8 model to ensure enhanced inference performance.
3. Consistency Validation:
   • Evaluation tests verifying consistency of results between Nvidia H100 and Iluvatar.

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 kernels—approximately 70%—to run the DeepSeek-R1 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, including fused MoE.

INT8 Dequantization

We provide dequantized model weights in bfloat16 to run DeepSeek-R1 on Iluvatar GPUs, along with adapted configuration files and tokenizer.

Bundle Download

	Usage	Iluvatar
Basic Image	basic software environment that supports model running	docker pull flagrelease-registry.cn-beijing.cr.aliyuncs.com/flagrelease/flagrelease:deepseek-flagos-iluvatar
Model	model weight and configuration files	https://www.modelscope.cn/models/FlagRelease/DeepSeek-R1-FlagOS-Iluvatar-INT8

Evaluation Results

Benchmark Result

Metrics	DeepSeek-R1-H100-CUDA	DeepSeek-R1-FlagOS-Iluvatar-INT8
GSM8K (EM)	95.75	95.53
MMLU (Acc.)	85.34	82.16
CEVAL	89.00	80.31
AIME 2024 (Pass@1)	76.66	TBD

How to Run Locally

📌 Getting Started

Environment Setup

# install FlagScale
git clone https://github.com/FlagOpen/FlagScale.git
cd FlagScale
pip install .

# download image and ckpt
flagscale pull --image flagrelease-registry.cn-beijing.cr.aliyuncs.com/flagrelease/flagrelease:deepseek-flagos-iluvatar --ckpt https://www.modelscope.cn/models/FlagRelease/DeepSeek-R1-FlagOS-Iluvatar-INT8.git --ckpt-path <CKPT_PATH>

# build and enter the container
docker run -dit -v <CKPT_PATH>:<CKPT_PATH> --pid=host --ipc=host --privileged --cap-add=ALL --network=host --name=flagrelease_bi flagrelease-registry.cn-beijing.cr.aliyuncs.com/flagrelease/flagrelease:deepseek-flagos-iluvatar
docker exec -it flagrelease_bi /bin/bash

Download and install FlagGems

git clone https://github.com/FlagOpen/FlagGems.git
cd FlagGems
git checkout deepseek_release_iluvatar
# no additional dependencies since they are already handled in the Docker environment
pip install ./ --no-deps
cd ../

Download FlagScale and unpatch the vendor's code to build vllm

git clone https://github.com/FlagOpen/FlagScale.git
cd FlagScale
python tools/patch/unpatch.py --device-type bi_V150 --commit-id fe6c78c6 --key-path <KEY_PATH> --dir build # The key needs to be obtained from the vendor
cd build/bi_V150/FlagScale/vllm
bash clean_vllm.sh;bash build_vllm.sh;bash install_vllm.sh
cd ../

Serve

# config the deepseek_r1 yaml
build/bi_V150/FlagScale/
├── examples/
│   └── deepseek_r1/
│       └── config_deepseek_r1.yaml # set hostfile
│       └── serve/
│           └── deepseek_r1.yaml # set model parameters and server port

# install flagscale
pip install .

# serve
flagscale serve deepseek_r1

Contributing

We warmly welcome global developers to join us:

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

📞 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.