meituan/DeepSeek-R1-Block-INT8

1---2license: mit3library_name: transformers4---5 6# Block-wise INT8 DeepSeek-R17 8The INT8 data type is both friendly and efficient for most hardware platforms.9 10**We provide a block-wise INT8 weight for DeepSeek-R1.**11 12In benchmarking, we observe **no accuracy loss** and up to **33\%** performance enhancement.13 14Thanks to our merged [PULL REQUEST](https://github.com/sgl-project/sglang/pull/3730), [SGLang](https://github.com/sgl-project/sglang/tree/main) is now support the block-wise INT8 quantization operation.15 16## 1. Benchmarking Result (detailed in [PULL REQUEST](https://github.com/sgl-project/sglang/pull/3730)):17| Model  | Config | Accuracy (GSM8K) | Accuracy (MMLU) | Output Throughput(qps=128) |18|--------|--------|-------------------|----------------|------------------------------|19| BF16 R1 | A100\*32  | 95.5              | 87.1           | 3342.29                       |20| INT8 R1 | (A100\*16)x2 | **95.8**              | **87.1**           | 4450.02 **(+33%)**                |21 22## 2. Quantization Process23 24We apply INT8 quantization to the BF16 checkpoints.25 26The quantization scales are determined by dividing the block-wise maximum of element values by the INT8 type maximum.27 28To generate this weight, run the provided script in the ``./inference`` directory:29 30``31python3 bf16_cast_block_int8.py --input-bf16-hf-path /path/to/bf16-weights/ --output-int8-hf-path /path/to/save-int8-weight/32``33 34## 3. Trouble Shooting35 36Before inference, you should confirm the "quantization_config" of `config.json` in `/path/to/save-int8-weight/`  should be:37 38``39"quantization_config": {40  "activation_scheme": "dynamic",41  "quant_method": "blockwise_int8",42  "weight_block_size": [43    128,44    12845  ]46}47``48 49---50 51# DeepSeek-R152<!-- markdownlint-disable first-line-h1 -->53<!-- markdownlint-disable html -->54<!-- markdownlint-disable no-duplicate-header -->55 56<div align="center">57  <img src="https://github.com/deepseek-ai/DeepSeek-V2/blob/main/figures/logo.svg?raw=true" width="60%" alt="DeepSeek-V3" />58</div>59<hr>60<div align="center" style="line-height: 1;">61  <a href="https://www.deepseek.com/" target="_blank" style="margin: 2px;">62    <img alt="Homepage" src="https://github.com/deepseek-ai/DeepSeek-V2/blob/main/figures/badge.svg?raw=true" style="display: inline-block; vertical-align: middle;"/>63  </a>64  <a href="https://chat.deepseek.com/" target="_blank" style="margin: 2px;">65    <img alt="Chat" src="https://img.shields.io/badge/🤖%20Chat-DeepSeek%20R1-536af5?color=536af5&logoColor=white" style="display: inline-block; vertical-align: middle;"/>66  </a>67  <a href="https://huggingface.co/deepseek-ai" target="_blank" style="margin: 2px;">68    <img alt="Hugging Face" src="https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-DeepSeek%20AI-ffc107?color=ffc107&logoColor=white" style="display: inline-block; vertical-align: middle;"/>69  </a>70</div>71 72<div align="center" style="line-height: 1;">73  <a href="https://discord.gg/Tc7c45Zzu5" target="_blank" style="margin: 2px;">74    <img alt="Discord" src="https://img.shields.io/badge/Discord-DeepSeek%20AI-7289da?logo=discord&logoColor=white&color=7289da" style="display: inline-block; vertical-align: middle;"/>75  </a>76  <a href="https://github.com/deepseek-ai/DeepSeek-V2/blob/main/figures/qr.jpeg?raw=true" target="_blank" style="margin: 2px;">77    <img alt="Wechat" src="https://img.shields.io/badge/WeChat-DeepSeek%20AI-brightgreen?logo=wechat&logoColor=white" style="display: inline-block; vertical-align: middle;"/>78  </a>79  <a href="https://twitter.com/deepseek_ai" target="_blank" style="margin: 2px;">80    <img alt="Twitter Follow" src="https://img.shields.io/badge/Twitter-deepseek_ai-white?logo=x&logoColor=white" style="display: inline-block; vertical-align: middle;"/>81  </a>82</div>83 84<div align="center" style="line-height: 1;">85  <a href="https://github.com/deepseek-ai/DeepSeek-R1/blob/main/LICENSE" style="margin: 2px;">86    <img alt="License" src="https://img.shields.io/badge/License-MIT-f5de53?&color=f5de53" style="display: inline-block; vertical-align: middle;"/>87  </a>88</div>89 90 91<p align="center">92  <a href="https://github.com/deepseek-ai/DeepSeek-R1/blob/main/DeepSeek_R1.pdf"><b>Paper Link</b>👁️</a>93</p>94 95## 1. Introduction96 97We introduce our first-generation reasoning models, DeepSeek-R1-Zero and DeepSeek-R1. 98DeepSeek-R1-Zero, a model trained via large-scale reinforcement learning (RL) without supervised fine-tuning (SFT) as a preliminary step, demonstrated remarkable performance on reasoning.99With RL, DeepSeek-R1-Zero naturally emerged with numerous powerful and interesting reasoning behaviors.100However, DeepSeek-R1-Zero encounters challenges such as endless repetition, poor readability, and language mixing. To address these issues and further enhance reasoning performance,101we introduce DeepSeek-R1, which incorporates cold-start data before RL.102DeepSeek-R1 achieves performance comparable to OpenAI-o1 across math, code, and reasoning tasks. 103To support the research community, we have open-sourced DeepSeek-R1-Zero, DeepSeek-R1, and six dense models distilled from DeepSeek-R1 based on Llama and Qwen. DeepSeek-R1-Distill-Qwen-32B outperforms OpenAI-o1-mini across various benchmarks, achieving new state-of-the-art results for dense models.104 105**NOTE: Before running DeepSeek-R1 series models locally, we kindly recommend reviewing the [Usage Recommendation](#usage-recommendations) section.**106 107<p align="center">108  <img width="80%" src="figures/benchmark.jpg">109</p>110 111## 2. Model Summary112 113---114 115**Post-Training: Large-Scale Reinforcement Learning on the Base Model**116 117-  We directly apply reinforcement learning (RL) to the base model without relying on supervised fine-tuning (SFT) as a preliminary step. This approach allows the model to explore chain-of-thought (CoT) for solving complex problems, resulting in the development of DeepSeek-R1-Zero. DeepSeek-R1-Zero demonstrates capabilities such as self-verification, reflection, and generating long CoTs, marking a significant milestone for the research community. Notably, it is the first open research to validate that reasoning capabilities of LLMs can be incentivized purely through RL, without the need for SFT. This breakthrough paves the way for future advancements in this area.118 119-   We introduce our pipeline to develop DeepSeek-R1. The pipeline incorporates two RL stages aimed at discovering improved reasoning patterns and aligning with human preferences, as well as two SFT stages that serve as the seed for the model's reasoning and non-reasoning capabilities.120    We believe the pipeline will benefit the industry by creating better models. 121 122---123 124**Distillation: Smaller Models Can Be Powerful Too**125 126-  We demonstrate that the reasoning patterns of larger models can be distilled into smaller models, resulting in better performance compared to the reasoning patterns discovered through RL on small models. The open source DeepSeek-R1, as well as its API, will benefit the research community to distill better smaller models in the future. 127- Using the reasoning data generated by DeepSeek-R1, we fine-tuned several dense models that are widely used in the research community. The evaluation results demonstrate that the distilled smaller dense models perform exceptionally well on benchmarks. We open-source distilled 1.5B, 7B, 8B, 14B, 32B, and 70B checkpoints based on Qwen2.5 and Llama3 series to the community.128 129## 3. Model Downloads130 131### DeepSeek-R1 Models132 133<div align="center">134 135| **Model** | **#Total Params** | **#Activated Params** | **Context Length** | **Download** |136| :------------: | :------------: | :------------: | :------------: | :------------: |137| DeepSeek-R1-Zero | 671B | 37B | 128K   | [🤗 HuggingFace](https://huggingface.co/deepseek-ai/DeepSeek-R1-Zero)   |138| DeepSeek-R1   | 671B | 37B |  128K   | [🤗 HuggingFace](https://huggingface.co/deepseek-ai/DeepSeek-R1)   |139 140</div>141 142DeepSeek-R1-Zero & DeepSeek-R1 are trained based on DeepSeek-V3-Base. 143For more details regarding the model architecture, please refer to [DeepSeek-V3](https://github.com/deepseek-ai/DeepSeek-V3) repository.144 145### DeepSeek-R1-Distill Models146 147<div align="center">148 149| **Model** | **Base Model** | **Download** |150| :------------: | :------------: | :------------: |151| DeepSeek-R1-Distill-Qwen-1.5B  | [Qwen2.5-Math-1.5B](https://huggingface.co/Qwen/Qwen2.5-Math-1.5B) | [🤗 HuggingFace](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B)   |152| DeepSeek-R1-Distill-Qwen-7B  | [Qwen2.5-Math-7B](https://huggingface.co/Qwen/Qwen2.5-Math-7B) | [🤗 HuggingFace](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-7B)   |153| DeepSeek-R1-Distill-Llama-8B  | [Llama-3.1-8B](https://huggingface.co/meta-llama/Llama-3.1-8B) | [🤗 HuggingFace](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-8B)   |154| DeepSeek-R1-Distill-Qwen-14B   | [Qwen2.5-14B](https://huggingface.co/Qwen/Qwen2.5-14B) | [🤗 HuggingFace](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B)   |155|DeepSeek-R1-Distill-Qwen-32B  | [Qwen2.5-32B](https://huggingface.co/Qwen/Qwen2.5-32B) | [🤗 HuggingFace](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B)   |156| DeepSeek-R1-Distill-Llama-70B  | [Llama-3.3-70B-Instruct](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct) | [🤗 HuggingFace](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-70B)   |157 158</div>159 160DeepSeek-R1-Distill models are fine-tuned based on open-source models, using samples generated by DeepSeek-R1.161We slightly change their configs and tokenizers. Please use our setting to run these models.162 163## 4. Evaluation Results164 165### DeepSeek-R1-Evaluation166 For all our models, the maximum generation length is set to 32,768 tokens. For benchmarks requiring sampling, we use a temperature of $0.6$, a top-p value of $0.95$, and generate 64 responses per query to estimate pass@1.167<div align="center">168 169 170| Category | Benchmark (Metric) | Claude-3.5-Sonnet-1022 | GPT-4o 0513 | DeepSeek V3 | OpenAI o1-mini | OpenAI o1-1217 | DeepSeek R1 |171|----------|-------------------|----------------------|------------|--------------|----------------|------------|--------------|172| | Architecture | - | - | MoE | - | - | MoE |173| | # Activated Params | - | - | 37B | - | - | 37B |174| | # Total Params | - | - | 671B | - | - | 671B |175| English | MMLU (Pass@1) | 88.3 | 87.2 | 88.5 | 85.2 | **91.8** | 90.8 |176| | MMLU-Redux (EM) | 88.9 | 88.0 | 89.1 | 86.7 | - | **92.9** |177| | MMLU-Pro (EM) | 78.0 | 72.6 | 75.9 | 80.3 | - | **84.0** |178| | DROP (3-shot F1) | 88.3 | 83.7 | 91.6 | 83.9 | 90.2 | **92.2** |179| | IF-Eval (Prompt Strict) | **86.5** | 84.3 | 86.1 | 84.8 | - | 83.3 |180| | GPQA-Diamond (Pass@1) | 65.0 | 49.9 | 59.1 | 60.0 | **75.7** | 71.5 |181| | SimpleQA (Correct) | 28.4 | 38.2 | 24.9 | 7.0 | **47.0** | 30.1 |182| | FRAMES (Acc.) | 72.5 | 80.5 | 73.3 | 76.9 | - | **82.5** |183| | AlpacaEval2.0 (LC-winrate) | 52.0 | 51.1 | 70.0 | 57.8 | - | **87.6** |184| | ArenaHard (GPT-4-1106) | 85.2 | 80.4 | 85.5 | 92.0 | - | **92.3** |185| Code | LiveCodeBench (Pass@1-COT) | 33.8 | 34.2 | - | 53.8 | 63.4 | **65.9** |186| | Codeforces (Percentile) | 20.3 | 23.6 | 58.7 | 93.4 | **96.6** | 96.3 |187| | Codeforces (Rating) | 717 | 759 | 1134 | 1820 | **2061** | 2029 |188| | SWE Verified (Resolved) | **50.8** | 38.8 | 42.0 | 41.6 | 48.9 | 49.2 |189| | Aider-Polyglot (Acc.) | 45.3 | 16.0 | 49.6 | 32.9 | **61.7** | 53.3 |190| Math | AIME 2024 (Pass@1) | 16.0 | 9.3 | 39.2 | 63.6 | 79.2 | **79.8** |191| | MATH-500 (Pass@1) | 78.3 | 74.6 | 90.2 | 90.0 | 96.4 | **97.3** |192| | CNMO 2024 (Pass@1) | 13.1 | 10.8 | 43.2 | 67.6 | - | **78.8** |193| Chinese | CLUEWSC (EM) | 85.4 | 87.9 | 90.9 | 89.9 | - | **92.8** |194| | C-Eval (EM) | 76.7 | 76.0 | 86.5 | 68.9 | - | **91.8** |195| | C-SimpleQA (Correct) | 55.4 | 58.7 | **68.0** | 40.3 | - | 63.7 |196 197</div>198 199 200### Distilled Model Evaluation201 202 203<div align="center">204 205| Model                                    | AIME 2024 pass@1 | AIME 2024 cons@64 | MATH-500 pass@1 | GPQA Diamond pass@1 | LiveCodeBench pass@1 | CodeForces rating |206|------------------------------------------|------------------|-------------------|-----------------|----------------------|----------------------|-------------------|207| GPT-4o-0513                          | 9.3              | 13.4              | 74.6            | 49.9                 | 32.9                 | 759               |208| Claude-3.5-Sonnet-1022             | 16.0             | 26.7                 | 78.3            | 65.0                 | 38.9                 | 717               |209| o1-mini                              | 63.6             | 80.0              | 90.0            | 60.0                 | 53.8                 | **1820**          |210| QwQ-32B-Preview                              | 44.0             | 60.0                 | 90.6            | 54.5               | 41.9                 | 1316              |211| DeepSeek-R1-Distill-Qwen-1.5B       | 28.9             | 52.7              | 83.9            | 33.8                 | 16.9                 | 954               |212| DeepSeek-R1-Distill-Qwen-7B          | 55.5             | 83.3              | 92.8            | 49.1                 | 37.6                 | 1189              |213| DeepSeek-R1-Distill-Qwen-14B         | 69.7             | 80.0              | 93.9            | 59.1                 | 53.1                 | 1481              |214| DeepSeek-R1-Distill-Qwen-32B        | **72.6**         | 83.3              | 94.3            | 62.1                 | 57.2                 | 1691              |215| DeepSeek-R1-Distill-Llama-8B         | 50.4             | 80.0              | 89.1            | 49.0                 | 39.6                 | 1205              |216| DeepSeek-R1-Distill-Llama-70B        | 70.0             | **86.7**          | **94.5**        | **65.2**             | **57.5**             | 1633              |217 218</div>219 220 221## 5. Chat Website & API Platform222You can chat with DeepSeek-R1 on DeepSeek's official website: [chat.deepseek.com](https://chat.deepseek.com), and switch on the button "DeepThink"223 224We also provide OpenAI-Compatible API at DeepSeek Platform: [platform.deepseek.com](https://platform.deepseek.com/)225 226## 6. How to Run Locally227 228### DeepSeek-R1 Models229 230Please visit [DeepSeek-V3](https://github.com/deepseek-ai/DeepSeek-V3) repo for more information about running DeepSeek-R1 locally.231 232**NOTE: Hugging Face's Transformers has not been directly supported yet.**233 234### DeepSeek-R1-Distill Models235 236DeepSeek-R1-Distill models can be utilized in the same manner as Qwen or Llama models.237 238For instance, you can easily start a service using [vLLM](https://github.com/vllm-project/vllm):239 240```shell241vllm serve deepseek-ai/DeepSeek-R1-Distill-Qwen-32B --tensor-parallel-size 2 --max-model-len 32768 --enforce-eager242```243 244You can also easily start a service using [SGLang](https://github.com/sgl-project/sglang)245 246```bash247python3 -m sglang.launch_server --model deepseek-ai/DeepSeek-R1-Distill-Qwen-32B --trust-remote-code --tp 2248```249 250### Usage Recommendations251 252**We recommend adhering to the following configurations when utilizing the DeepSeek-R1 series models, including benchmarking, to achieve the expected performance:**253 2541. Set the temperature within the range of 0.5-0.7 (0.6 is recommended) to prevent endless repetitions or incoherent outputs.2552. **Avoid adding a system prompt; all instructions should be contained within the user prompt.**2563. For mathematical problems, it is advisable to include a directive in your prompt such as: "Please reason step by step, and put your final answer within \boxed{}."2574. When evaluating model performance, it is recommended to conduct multiple tests and average the results.258 259Additionally, we have observed that the DeepSeek-R1 series models tend to bypass thinking pattern (i.e., outputting "\<think\>\n\n\</think\>") when responding to certain queries, which can adversely affect the model's performance.260**To ensure that the model engages in thorough reasoning, we recommend enforcing the model to initiate its response with "\<think\>\n" at the beginning of every output.**261 262## 7. License263This code repository and the model weights are licensed under the [MIT License](https://github.com/deepseek-ai/DeepSeek-R1/blob/main/LICENSE).264DeepSeek-R1 series support commercial use, allow for any modifications and derivative works, including, but not limited to, distillation for training other LLMs. Please note that:265- DeepSeek-R1-Distill-Qwen-1.5B, DeepSeek-R1-Distill-Qwen-7B, DeepSeek-R1-Distill-Qwen-14B and DeepSeek-R1-Distill-Qwen-32B are derived from [Qwen-2.5 series](https://github.com/QwenLM/Qwen2.5), which are originally licensed under [Apache 2.0 License](https://huggingface.co/Qwen/Qwen2.5-1.5B/blob/main/LICENSE), and now finetuned with 800k samples curated with DeepSeek-R1.266- DeepSeek-R1-Distill-Llama-8B is derived from Llama3.1-8B-Base and is originally licensed under [llama3.1 license](https://huggingface.co/meta-llama/Llama-3.1-8B/blob/main/LICENSE).267- DeepSeek-R1-Distill-Llama-70B is derived from Llama3.3-70B-Instruct and is originally licensed under [llama3.3 license](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct/blob/main/LICENSE).268 269## 8. Citation270```271@misc{deepseekai2025deepseekr1incentivizingreasoningcapability,272      title={DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning}, 273      author={DeepSeek-AI and Daya Guo and Dejian Yang and Haowei Zhang and Junxiao Song and Ruoyu Zhang and Runxin Xu and Qihao Zhu and Shirong Ma and Peiyi Wang and Xiao Bi and Xiaokang Zhang and Xingkai Yu and Yu Wu and Z. F. Wu and Zhibin Gou and Zhihong Shao and Zhuoshu Li and Ziyi Gao and Aixin Liu and Bing Xue and Bingxuan Wang and Bochao Wu and Bei Feng and Chengda Lu and Chenggang Zhao and Chengqi Deng and Chenyu Zhang and Chong Ruan and Damai Dai and Deli Chen and Dongjie Ji and Erhang Li and Fangyun Lin and Fucong Dai and Fuli Luo and Guangbo Hao and Guanting Chen and Guowei Li and H. Zhang and Han Bao and Hanwei Xu and Haocheng Wang and Honghui Ding and Huajian Xin and Huazuo Gao and Hui Qu and Hui Li and Jianzhong Guo and Jiashi Li and Jiawei Wang and Jingchang Chen and Jingyang Yuan and Junjie Qiu and Junlong Li and J. L. Cai and Jiaqi Ni and Jian Liang and Jin Chen and Kai Dong and Kai Hu and Kaige Gao and Kang Guan and Kexin Huang and Kuai Yu and Lean Wang and Lecong Zhang and Liang Zhao and Litong Wang and Liyue Zhang and Lei Xu and Leyi Xia and Mingchuan Zhang and Minghua Zhang and Minghui Tang and Meng Li and Miaojun Wang and Mingming Li and Ning Tian and Panpan Huang and Peng Zhang and Qiancheng Wang and Qinyu Chen and Qiushi Du and Ruiqi Ge and Ruisong Zhang and Ruizhe Pan and Runji Wang and R. J. Chen and R. L. Jin and Ruyi Chen and Shanghao Lu and Shangyan Zhou and Shanhuang Chen and Shengfeng Ye and Shiyu Wang and Shuiping Yu and Shunfeng Zhou and Shuting Pan and S. S. Li and Shuang Zhou and Shaoqing Wu and Shengfeng Ye and Tao Yun and Tian Pei and Tianyu Sun and T. Wang and Wangding Zeng and Wanjia Zhao and Wen Liu and Wenfeng Liang and Wenjun Gao and Wenqin Yu and Wentao Zhang and W. L. Xiao and Wei An and Xiaodong Liu and Xiaohan Wang and Xiaokang Chen and Xiaotao Nie and Xin Cheng and Xin Liu and Xin Xie and Xingchao Liu and Xinyu Yang and Xinyuan Li and Xuecheng Su and Xuheng Lin and X. Q. Li and Xiangyue Jin and Xiaojin Shen and Xiaosha Chen and Xiaowen Sun and Xiaoxiang Wang and Xinnan Song and Xinyi Zhou and Xianzu Wang and Xinxia Shan and Y. K. Li and Y. Q. Wang and Y. X. Wei and Yang Zhang and Yanhong Xu and Yao Li and Yao Zhao and Yaofeng Sun and Yaohui Wang and Yi Yu and Yichao Zhang and Yifan Shi and Yiliang Xiong and Ying He and Yishi Piao and Yisong Wang and Yixuan Tan and Yiyang Ma and Yiyuan Liu and Yongqiang Guo and Yuan Ou and Yuduan Wang and Yue Gong and Yuheng Zou and Yujia He and Yunfan Xiong and Yuxiang Luo and Yuxiang You and Yuxuan Liu and Yuyang Zhou and Y. X. Zhu and Yanhong Xu and Yanping Huang and Yaohui Li and Yi Zheng and Yuchen Zhu and Yunxian Ma and Ying Tang and Yukun Zha and Yuting Yan and Z. Z. Ren and Zehui Ren and Zhangli Sha and Zhe Fu and Zhean Xu and Zhenda Xie and Zhengyan Zhang and Zhewen Hao and Zhicheng Ma and Zhigang Yan and Zhiyu Wu and Zihui Gu and Zijia Zhu and Zijun Liu and Zilin Li and Ziwei Xie and Ziyang Song and Zizheng Pan and Zhen Huang and Zhipeng Xu and Zhongyu Zhang and Zhen Zhang},274      year={2025},275      eprint={2501.12948},276      archivePrefix={arXiv},277      primaryClass={cs.CL},278      url={https://arxiv.org/abs/2501.12948}, 279}280 281```282 283## 9. Contact284If you have any questions, please raise an issue or contact us at [service@deepseek.com](service@deepseek.com).285