DeepSeek-R1-0528_w4a8_fp8

<p align="center"> <picture> <source media="(prefers-color-scheme: dark)" srcset="https://github.com/Tencent/AngelSlim/blob/main/docs/source/assets/logos/angelslim_logo_light.png?raw=true"> <img alt="AngelSlim" src="https://github.com/Tencent/AngelSlim/blob/main/docs/source/assets/logos/angelslim_logo.png?raw=true" width=55%> </picture> </p> <h3 align="center"> Dedicated to building a more intuitive, comprehensive, and efficient LLMs compression toolkit. </h3> <p align="center"> 📖 <a href="https://angelslim.readthedocs.io/">Documentation</a>&nbsp&nbsp | &nbsp&nbsp🤗 <a href="https://huggingface.co/AngelSlim">Hugging Face</a>&nbsp&nbsp | &nbsp&nbsp🤖 <a href="https://modelscope.cn/organization/AngelSlim">ModelScope</a>&nbsp&nbsp | &nbsp&nbsp💬 <a href="./docs/source/assets/angel_slim_wechat.png">WeChat</a> <br> </p>

Table of Contents

• Latest Updates
• Key Features
• Supported Models
• How to Use
  • Install AngelSlim
  • Quick Start
  • deployment & Evaluation
• Benchmark
• License
• Citation
• Technical Discussion

📣Latest Updates

• [25/07/04] We now support quantization for Hunyuan/Qwen2.5/Qwen3/DeepSeek-R1-Distill-Qwen and other models, including INT8/FP8/INT4 algorithms. We also opensource Qwen3-8B`s Eagle3 model weight.

Coming soon:

• Support W4A8 quantization for DeepSeek-R1.
• Support quantization for multimodal models like Qwen-VL.
• Release of new algorithm for speculative sampling.

🌟Key Features

• Highly Integrated: This toolkit integrates mainstream compression algorithms into a unified framework, offering developers one-click access with exceptional ease of use.
• Continuous Innovation: Beyond integrating widely-used industry algorithms, we are continuously researching better compression algorithms, which will be gradually open-sourced in the future.
• Performance-Driven: We continuously optimize end-to-end performance in model compression workflows and algorithm deployment, such as enabling quantization of models like Qwen3-235B and DeepSeek-R1 on a single GPU.

💼Supported Models

Quantization

Currently supports the following LLMs, including Hunyuan-Dense, Hunyuan-MoE, Qwen3-Dense, Qwen3-MoE, Qwen2.5, DeepSeek-R1 distilled Qwen models, and QwQ::

Speculative Decoding

The Eagle3 weights for the Qwen3-8B model are now available, with Eagle3 weights for other models in the Qwen3 series to be released soon.

🛎️How to Use

Install AngelSlim

We recommend using pip to install the latest stable version of AngelSlim:

pip install angelslim

Alternatively, you can clone the repository and install from source in editable mode:

cd AngelSlim && python setup.py install

For more detailed installation instructions, please refer to the Installation Documentation.

Quick Start

After installing AngelSlim, you can quickly start by running the following script to perform static FP8 quantization on the Qwen3-1.7B model:

• One-click Start

  python3 tools/run.py -c configs/qwen3/fp8_static/qwen3-1_7b_fp8_static.yaml
  

  This example will load the HuggingFace model and perform activation value calibration using the dataset specified in the config file, saving the quantized model weights.

• Code-based Start

  To perform dynamic FP8 quantization on Qwen3-1.7B:

  from angelslim.engine import Engine
  
  slim_engine = Engine()
  # Prepare model
  slim_engine.prepare_model(model_name="Qwen", model_path="Qwen/Qwen3-1.7B",)
  # Initialize compressor
  slim_engine.prepare_compressor("PTQ", default_method="fp8_dynamic")
  # Compress model
  slim_engine.run()
  # Save compressed model
  slim_engine.save("./output")
  

For more details, please refer to the Quick Start Documentation.

🖥️ Deployment and Testing

1. API Service Deployment

After specifying the quantized model path MODEL_PATH, you can deploy an OpenAI-compatible API service using the following LLMs inference frameworks:

vLLM

Use the following script to launch a vLLM server, recommended version vllm>=0.8.5.post1. For MOE INT8 quantized models, vllm>=0.9.0 is required.

bash deploy/run_vllm.sh $MODEL_PATH

SGLang

Use the following script to launch a SGLang server, recommended version sglang>=0.4.6.post1.

bash deploy/run_sglang.sh $MODEL_PATH

2. Service Invocation

Invoke requests via OpenAI's API format:

bash deploy/openai.sh $MODEL_PATH

3. Performance Evaluation

Evaluate the performance of quantized model using lm-evaluation-harness, recommended versionlm-eval>=0.4.8:

bash deploy/lm_eval.sh $MODEL_PATH

For more detaileds, please refer to the Deployment Documentation.

📈 Benchmark

Quantization

The performance test results for selected models are shown below. For the complete benchmark, refer to the Benchmark documentation

Hunyuan Series Models

Benchmark results for the Hunyuan-A13B-Instruct model with FP8 and INT4-GPTQ quantization algorithms on datasets including AIME 2024, GSM8K, BBH, and DROP:

Qwen3 Series Models

Benchmark results for Qwen3 series models with FP8-Static, FP8-Dynamic, INT4-GPTQ, and INT4-AWQ quantization algorithms on datasets including CEVAL, MMLU, GSM8K, and HUMANEVAL:

<table> <thead> <tr><th>Model</th><th>Quantization</th><th>CEVAL</th><th>MMLU</th><th>GSM8K</th><th>HUMANEVAL</th></tr> </thead> <tbody> <tr><td rowspan="4">Qwen3-0.6B</td><td>BF16</td><td>45.84</td><td>47.21</td><td>42.99</td><td>19.51</td></tr> <tr><td>FP8-Static</td><td>45.99</td><td>46.87</td><td>38.06</td><td>18.90</td></tr> <tr><td>FP8-Dynamic</td><td>45.99</td><td>46.93</td><td>38.29</td><td>20.73</td></tr> <tr><td>INT8-Dynamic</td><td>45.17</td><td>46.95</td><td>41.17</td><td>21.34</td></tr> <tr><td rowspan="6">Qwen3-8B</td><td>BF16</td><td>79.27</td><td>74.78</td><td>87.79</td><td>63.41</td></tr> <tr><td>FP8-Static</td><td>78.23</td><td>74.79</td><td>86.96</td><td>62.20</td></tr> <tr><td>FP8-Dynamic</td><td>78.45</td><td>74.75</td><td>87.64</td><td>62.80</td></tr> <tr><td>INT8-Dynamic</td><td>78.01</td><td>74.84</td><td>86.96</td><td>67.07</td></tr> <tr><td>INT4-GPTQ</td><td>77.19</td><td>73.26</td><td>86.43</td><td>62.20</td></tr> <tr><td>INT4-AWQ</td><td>76.15</td><td>73.59</td><td>86.96</td><td>63.41</td></tr> <tr><td rowspan="6">Qwen3-14B</td><td>BF16</td><td>83.06</td><td>78.90</td><td>88.40</td><td>55.49</td></tr> <tr><td>FP8-Static</td><td>82.62</td><td>78.57</td><td>89.46</td><td>57.32</td></tr> <tr><td>FP8-Dynamic</td><td>82.24</td><td>78.92</td><td>88.32</td><td>52.44</td></tr> <tr><td>INT8-Dynamic</td><td>81.87</td><td>78.13</td><td>86.28</td><td>56.10</td></tr> <tr><td>INT4-GPTQ</td><td>81.05</td><td>78.02</td><td>87.34</td><td>57.93</td></tr> <tr><td>INT4-AWQ</td><td>82.02</td><td>77.68</td><td>84.23</td><td>61.59</td></tr> <tr><td rowspan="5">Qwen3-32B</td><td>BF16</td><td>86.55</td><td>82.00</td><td>74.53</td><td>37.80</td></tr> <tr><td>FP8-Static</td><td>86.92</td><td>81.78</td><td>70.20</td><td>39.63</td></tr> <tr><td>FP8-Dynamic</td><td>86.55</td><td>81.89</td><td>70.43</td><td>38.41</td></tr> <tr><td>INT4-GPTQ</td><td>86.18</td><td>81.01</td><td>-</td><td>43.29</td></tr> <tr><td>INT4-AWQ</td><td>86.18</td><td>81.54</td><td>-</td><td>36.59</td></tr> <tr><td rowspan="4">Qwen3-30B-A3B</td><td>BF16</td><td>83.66</td><td>79.36</td><td>89.99</td><td>31.71</td></tr> <tr><td>FP8-Static</td><td>83.95</td><td>79.47</td><td>89.01</td><td>31.10</td></tr> <tr><td>FP8-Dynamic</td><td>84.10</td><td>79.40</td><td>89.16</td><td>32.93</td></tr> <tr><td>INT8-Dynamic</td><td>83.36</td><td>79.48</td><td>89.16</td><td>34.15</td></tr> <tr><td rowspan="4">Qwen3-235B-A22B</td><td>BF16</td><td>89.60</td><td>86.28</td><td>85.29</td><td>27.44</td></tr> <tr><td>FP8-Static</td><td>89.67</td><td>86.19</td><td>86.96</td><td>27.44</td></tr> <tr><td>FP8-Dynamic</td><td>89.67</td><td>86.18</td><td>85.22</td><td>28.05</td></tr> <tr><td>INT8-Dynamic</td><td>88.93</td><td>86.20</td><td>86.20</td><td>23.78</td></tr> <tr><td rowspan="5">QwQ-32B</td><td>BF16</td><td>85.74</td><td>82.03</td><td>73.31</td><td>42.68</td></tr> <tr><td>FP8-Static</td><td>85.44</td><td>81.91</td><td>75.36</td><td>42.68</td></tr> <tr><td>FP8-Dynamic</td><td>85.07</td><td>81.93</td><td>75.66</td><td>42.07</td></tr> <tr><td>INT4-GPTQ</td><td>84.03</td><td>81.26</td><td>68.23</td><td>45.73</td></tr> <tr><td>INT4-AWQ</td><td>83.58</td><td>81.01</td><td>68.69</td><td>43.29</td></tr> </tbody> </table>

Other Models

Benchmark results for other models with FP8-Static, FP8-Dynamic, INT4-GPTQ, and INT4-AWQ quantization algorithms on datasets including CEVAL, MMLU and GSM8K:

<table> <thead> <tr><th>Model</th><th>Quantization</th><th>CEVAL</th><th>MMLU</th><th>GSM8K</th></tr> </thead> <tbody> <tr><td rowspan="3">Qwen2.5-1.5B-Instruct</td><td>BF16</td><td>67.01</td><td>60.05</td><td>54.28</td></tr> <tr><td>FP8-Static</td><td>66.27</td><td>60.23</td><td>-</td></tr> <tr><td>FP8-Dynamic</td><td>66.79</td><td>60.08</td><td>51.71</td></tr> <tr><td rowspan="5">Qwen2.5-7B-Instruct</td><td>BF16</td><td>81.20</td><td>74.55</td><td>79.98</td></tr> <tr><td>FP8-Static</td><td>81.13</td><td>74.03</td><td>79.30</td></tr> <tr><td>FP8-Dynamic</td><td>80.31</td><td>74.07</td><td>79.00</td></tr> <tr><td>INT4-GPTQ</td><td>79.05</td><td>73.05</td><td>74.75</td></tr> <tr><td>INT4-AWQ</td><td>79.35</td><td>73.22</td><td>79.38</td></tr> <tr><td rowspan="5">Qwen2.5-32B-Instruct</td><td>BF16</td><td>87.30</td><td>83.21</td><td>81.73</td></tr> <tr><td>FP8-Static</td><td>87.59</td><td>83.08</td><td>81.58</td></tr> <tr><td>FP8-Dynamic</td><td>87.30</td><td>83.04</td><td>81.58</td></tr> <tr><td>INT4-GPTQ</td><td>86.70</td><td>82.45</td><td>82.03</td></tr> <tr><td>INT4-AWQ</td><td>87.00</td><td>82.64</td><td>-</td></tr> <tr><td rowspan="5">DeepSeek-R1-Distill-Qwen-7B</td><td>BF16</td><td>53.49</td><td>53.80</td><td>75.74</td></tr> <tr><td>FP8-Static</td><td>53.57</td><td>54.17</td><td>76.19</td></tr> <tr><td>FP8-Dynamic</td><td>52.97</td><td>54.13</td><td>74.15</td></tr> <tr><td>INT4-GPTQ</td><td>51.86</td><td>52.44</td><td>75.89</td></tr> <tr><td>INT4-AWQ</td><td>53.49</td><td>53.70</td><td>-</td></tr> <tr><td rowspan="5">DeepSeek-R1-Distill-Qwen-14B</td><td>BF16</td><td>77.71</td><td>74.28</td><td>85.67</td></tr> <tr><td>FP8-Static</td><td>77.56</td><td>74.66</td><td>86.73</td></tr> <tr><td>FP8-Dynamic</td><td>76.82</td><td>74.63</td><td>87.11</td></tr> <tr><td>INT4-GPTQ</td><td>74.29</td><td>72.37</td><td>84.61</td></tr> <tr><td>INT4-AWQ</td><td>74.81</td><td>73.00</td><td>86.05</td></tr> <tr><td rowspan="5">DeepSeek-R1-Distill-Qwen-32B</td><td>BF16</td><td>84.18</td><td>80.89</td><td>87.41</td></tr> <tr><td>FP8-Static</td><td>83.43</td><td>80.90</td><td>87.57</td></tr> <tr><td>FP8-Dynamic</td><td>83.73</td><td>81.10</td><td>86.43</td></tr> <tr><td>INT4-GPTQ</td><td>84.10</td><td>79.80</td><td>86.73</td></tr> <tr><td>INT4-AWQ</td><td>82.84</td><td>80.15</td><td>87.19</td></tr> </tbody> </table>

Speculative Decoding

Benchmark results for Qwen3 series models with Eagle3 speculative decoding algorithm on datasets including MT-bench, HunmanEval, GSM8K, and Alpaca:

Qwen3-8B

<table border="0"> <thead> <tr><th rowspan="3">Temperature</th><th rowspan="3">Method</th><th colspan="8">Datasets</th></tr> <tr><th colspan="2">MT-bench</th><th colspan="2">HumanEval</th><th colspan="2">GSM8K</th><th colspan="2">Alpaca</th></tr> <tr><th>Speedup</th><th>Accept length</th><th>Speedup</th><th>Accept length</th><th>Speedup</th><th>Accept length</th><th>Speedup</th><th>Accept length</th></tr> </thead> <tbody> <tr><td>T=0</td><td>Eagle3</td><td>2.63x</td><td>3.65</td><td>2.76x</td><td>3.85</td><td>2.82x</td><td>3.90</td><td>2.62x</td><td>3.48</td></tr> <tr><td>T=1</td><td>Eagle3</td><td>1.98x</td><td>2.75</td><td>2.25x</td><td>3.11</td><td>2.31x</td><td>3.15</td><td>2.10x</td><td>2.76</td></tr> </tbody> </table>

📝 Model License

The code for this project is open-sourced under the License for AngelSlim.

🔗 Citation

@software{AngelSlim2025,
    title={{AngelSlim}},
    author={Tencent AngelSlim Project Contributors},
    year={2025},
    month={6},
    url={https://github.com/Tencent/AngelSlim},
}

💬 Technical Discussion

• AngelSlim is continuously iterating and new features will be released soon. If you have any questions or suggestions, please open an issue on GitHub or join our WeChat technical discussion group.