DeepSeek-R1-AWQ-W4AFP8

This model is a W4AFP8 quantized DeepSeek-R1 with AWQ quantizaton.
Releated PR：https://github.com/sgl-project/sglang/pull/8573
Releated Project: https://github.com/TMElyralab/sglang/tree/lyra_w4afp8

Benchmark

Test configuration: input/output len = 1000/1000, qps=64, max_concurrency=64, num_prompt=128
Device: H20 * 8
Compared to the original model：

• bs=64，input/output throughput has increased by 56%.
• bs=128，input/output throughput has increased by 125%.

============ Serving Benchmark Result ============
Backend:                                 sglang       
Max request concurrency:                 64        
Successful requests:                     128       
Benchmark duration (s):                  105.50    
Total input tokens:                      128000    
Total generated tokens:                  128000    
Total generated tokens (retokenized):    127551    
Request throughput (req/s):              1.21      
Input token throughput (tok/s):          1213.24   
Output token throughput (tok/s):         1213.24   
Total token throughput (tok/s):          2426.49   
Concurrency:                             63.97     
----------------End-to-End Latency----------------
Mean E2E Latency (ms):                   52728.31  
Median E2E Latency (ms):                 52728.33  
---------------Time to First Token----------------
Mean TTFT (ms):                          5444.26   
Median TTFT (ms):                        5425.69   
P99 TTFT (ms):                           8768.54   
---------------Inter-Token Latency----------------
Mean ITL (ms):                           47.33     
Median ITL (ms):                         44.18     
P95 ITL (ms):                            46.58     
P99 ITL (ms):                            46.76     
Max ITL (ms):                            7819.3
==================================================

How To Use

Mirror：lmsysorg/sglang:v0.4.6.post5-cu124
or lmsysorg/sglang:v0.5.1.post5-cu126 (cuda12.6 env need to update ptxas to 12.8 on Hopper. reference)

# 1. Repo clone
git clone https://github.com/TMElyralab/sglang && cd sglang
git checkout lyra_w4afp8

# 2. SGLang install
pip install uv cmake
pip install -e "python[all]"

# 3. Recompile sgl-kernel
cd sgl-kernel & make build

# 4. Run SGLang
python3 -m sglang.launch_server --model-path /path/to/DeepSeek-R1-AWQ-W4AFP8 --tp 8 --trust-remote-code --host 0.0.0.0 --port 8000 --mem-fraction-static 0.9 --quantization w4a8_machete --cuda-graph-max-bs 128 --max-running-requests 128

Citation

We are TMElyralab, the Acceleration Team from Tencent Music Entertainment (TME).

@Misc{TMElyralab_2025,
  author =       {Sa Xiao, Mian Peng, Haoxiong Su, Kangjian Wu, Bin Wu, Yibo Lu, Qiwen Mao, Wenjiang Zhou},
  howpublished = {\url{https://github.com/TMElyralab}},
  year =         {2025}
}

---

---

Paper Link👁️

1. Introduction

We introduce our first-generation reasoning models, DeepSeek-R1-Zero and DeepSeek-R1. DeepSeek-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. With RL, DeepSeek-R1-Zero naturally emerged with numerous powerful and interesting reasoning behaviors. However, DeepSeek-R1-Zero encounters challenges such as endless repetition, poor readability, and language mixing. To address these issues and further enhance reasoning performance, we introduce DeepSeek-R1, which incorporates cold-start data before RL. DeepSeek-R1 achieves performance comparable to OpenAI-o1 across math, code, and reasoning tasks. To 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.

NOTE: Before running DeepSeek-R1 series models locally, we kindly recommend reviewing the Usage Recommendation section.

2. Model Summary

---

Post-Training: Large-Scale Reinforcement Learning on the Base Model

• 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.

• 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. We believe the pipeline will benefit the industry by creating better models.

---

Distillation: Smaller Models Can Be Powerful Too

• 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.
• 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.

3. Model Downloads

DeepSeek-R1 Models

DeepSeek-R1-Zero & DeepSeek-R1 are trained based on DeepSeek-V3-Base. For more details regarding the model architecture, please refer to DeepSeek-V3 repository.

DeepSeek-R1-Distill Models

DeepSeek-R1-Distill models are fine-tuned based on open-source models, using samples generated by DeepSeek-R1. We slightly change their configs and tokenizers. Please use our setting to run these models.

4. Evaluation Results

DeepSeek-R1-Evaluation

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.

Distilled Model Evaluation

5. Chat Website & API Platform

You can chat with DeepSeek-R1 on DeepSeek's official website: chat.deepseek.com, and switch on the button "DeepThink"

We also provide OpenAI-Compatible API at DeepSeek Platform: platform.deepseek.com

6. How to Run Locally

DeepSeek-R1 Models

Please visit DeepSeek-V3 repo for more information about running DeepSeek-R1 locally.

NOTE: Hugging Face's Transformers has not been directly supported yet.

DeepSeek-R1-Distill Models

DeepSeek-R1-Distill models can be utilized in the same manner as Qwen or Llama models.

For instance, you can easily start a service using vLLM:

vllm serve deepseek-ai/DeepSeek-R1-Distill-Qwen-32B --tensor-parallel-size 2 --max-model-len 32768 --enforce-eager

You can also easily start a service using SGLang

python3 -m sglang.launch_server --model deepseek-ai/DeepSeek-R1-Distill-Qwen-32B --trust-remote-code --tp 2

Usage Recommendations

We recommend adhering to the following configurations when utilizing the DeepSeek-R1 series models, including benchmarking, to achieve the expected performance:

1. Set the temperature within the range of 0.5-0.7 (0.6 is recommended) to prevent endless repetitions or incoherent outputs.
2. Avoid adding a system prompt; all instructions should be contained within the user prompt.
3. 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{}."
4. When evaluating model performance, it is recommended to conduct multiple tests and average the results.

Additionally, 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. 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.

7. License

This code repository and the model weights are licensed under the MIT License. DeepSeek-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:

• 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, which are originally licensed under Apache 2.0 License, and now finetuned with 800k samples curated with DeepSeek-R1.
• DeepSeek-R1-Distill-Llama-8B is derived from Llama3.1-8B-Base and is originally licensed under llama3.1 license.
• DeepSeek-R1-Distill-Llama-70B is derived from Llama3.3-70B-Instruct and is originally licensed under llama3.3 license.

8. Citation

@misc{deepseekai2025deepseekr1incentivizingreasoningcapability,
      title={DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning}, 
      author={DeepSeek-AI},
      year={2025},
      eprint={2501.12948},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2501.12948}, 
}

9. Contact

If you have any questions, please raise an issue or contact us at service@deepseek.com.