Update README.md

README.md

@@ -276,3 +276,489 @@ lm_eval \
     </tr>
   </tbody>
 </table>
+
+## Inference Performance
+
+
+This model achieves up to 3.4x speedup in single-stream deployment and up to 2.0x speedup in multi-stream asynchronous deployment, depending on hardware and use-case scenario.
+The following performance benchmarks were conducted with [vLLM](https://docs.vllm.ai/en/latest/) version 0.7.2, and [GuideLLM](https://github.com/neuralmagic/guidellm).
+
+<details>
+<summary>Benchmarking Command</summary>
+
+```
+guidellm --model neuralmagic/DeepSeek-R1-Distill-Qwen-32B-quantized.w4a16 --target "http://localhost:8000/v1" --data-type emulated --data "prompt_tokens=<prompt_tokens>,generated_tokens=<generated_tokens>" --max seconds 360 --backend aiohttp_server
+```
+</details>
+
+### Single-stream performance (measured with vLLM version 0.7.2)
+<table>
+  <thead>
+    <tr>
+      <th></th>
+      <th></th>
+      <th></th>
+      <th></th>
+      <th style="text-align: center;" colspan="2" >Instruction Following<br>256 / 128</th>
+      <th style="text-align: center;" colspan="2" >Multi-turn Chat<br>512 / 256</th>
+      <th style="text-align: center;" colspan="2" >Docstring Generation<br>768 / 128</th>
+      <th style="text-align: center;" colspan="2" >RAG<br>1024 / 128</th>
+      <th style="text-align: center;" colspan="2" >Code Completion<br>256 / 1024</th>
+      <th style="text-align: center;" colspan="2" >Code Fixing<br>1024 / 1024</th>
+      <th style="text-align: center;" colspan="2" >Large Summarization<br>4096 / 512</th>
+      <th style="text-align: center;" colspan="2" >Large RAG<br>10240 / 1536</th>
+    </tr>
+    <tr>
+      <th>GPU class</th>
+      <th>Number of GPUs</th>
+      <th>Model</th>
+      <th>Average cost reduction</th>
+      <th>Latency (s)</th>
+      <th>QPD</th>
+      <th>Latency (s)</th>
+      <th>QPD</th>
+      <th>Latency (s)</th>
+      <th>QPD</th>
+      <th>Latency (s)</th>
+      <th>QPD</th>
+      <th>Latency (s)</th>
+      <th>QPD</th>
+      <th>Latency (s)</th>
+      <th>QPD</th>
+      <th>Latency (s)</th>
+      <th>QPD</th>
+      <th>Latency (s)</th>
+      <th>QPD</th>
+    </tr>
+  </thead>
+  <tbody style="text-align: center" >
+    <tr>
+      <th rowspan="3" valign="top">A6000</th>
+      <td>2</td>
+      <th>deepseek-ai/DeepSeek-R1-Distill-Qwen-32B</th>
+      <td>---</td>
+      <td>6.3</td>
+      <td>359</td>
+      <td>12.8</td>
+      <td>176</td>
+      <td>6.5</td>
+      <td>347</td>
+      <td>6.6</td>
+      <td>342</td>
+      <td>49.9</td>
+      <td>45</td>
+      <td>50.8</td>
+      <td>44</td>
+      <td>26.6</td>
+      <td>85</td>
+      <td>83.4</td>
+      <td>27</td>
+    </tr>
+    <tr>
+      <td>1</td>
+      <th>neuralmagic/DeepSeek-R1-Distill-Qwen-32B-quantized.w8a8</th>
+      <td>1.81</td>
+      <td>6.9</td>
+      <td>648</td>
+      <td>13.8</td>
+      <td>325</td>
+      <td>7.2</td>
+      <td>629</td>
+      <td>7.2</td>
+      <td>622</td>
+      <td>54.8</td>
+      <td>82</td>
+      <td>55.6</td>
+      <td>81</td>
+      <td>30.0</td>
+      <td>150</td>
+      <td>94.8</td>
+      <td>47</td>
+    </tr>
+    <tr>
+      <td>1</td>
+      <th>neuralmagic/DeepSeek-R1-Distill-Qwen-32B-quantized.w4a16</th>
+      <td>3.07</td>
+      <td>3.9</td>
+      <td>1168</td>
+      <td>7.8</td>
+      <td>580</td>
+      <td>4.3</td>
+      <td>1041</td>
+      <td>4.6</td>
+      <td>975</td>
+      <td>29.7</td>
+      <td>151</td>
+      <td>30.9</td>
+      <td>146</td>
+      <td>19.3</td>
+      <td>233</td>
+      <td>61.4</td>
+      <td>73</td>
+    </tr>
+    <tr>
+      <th rowspan="3" valign="top">A100</th>
+      <td>1</td>
+      <th>deepseek-ai/DeepSeek-R1-Distill-Qwen-32B</th>
+      <td>---</td>
+      <td>5.6</td>
+      <td>361</td>
+      <td>11.1</td>
+      <td>180</td>
+      <td>5.7</td>
+      <td>350</td>
+      <td>5.8</td>
+      <td>347</td>
+      <td>44.0</td>
+      <td>46</td>
+      <td>44.7</td>
+      <td>45</td>
+      <td>23.6</td>
+      <td>85</td>
+      <td>73.7</td>
+      <td>27</td>
+    </tr>
+    <tr>
+      <td>1</td>
+      <th>neuralmagic/DeepSeek-R1-Distill-Qwen-32B-quantized.w8a8</th>
+      <td>1.50</td>
+      <td>3.7</td>
+      <td>547</td>
+      <td>7.3</td>
+      <td>275</td>
+      <td>3.8</td>
+      <td>536</td>
+      <td>3.8</td>
+      <td>528</td>
+      <td>29.0</td>
+      <td>69</td>
+      <td>29.5</td>
+      <td>68</td>
+      <td>15.7</td>
+      <td>128</td>
+      <td>53.1</td>
+      <td>38</td>
+    </tr>
+    <tr>
+      <td>1</td>
+      <th>neuralmagic/DeepSeek-R1-Distill-Qwen-32B-quantized.w4a16</th>
+      <td>2.30</td>
+      <td>2.2</td>
+      <td>894</td>
+      <td>4.5</td>
+      <td>449</td>
+      <td>2.4</td>
+      <td>831</td>
+      <td>2.5</td>
+      <td>798</td>
+      <td>17.4</td>
+      <td>116</td>
+      <td>18.0</td>
+      <td>112</td>
+      <td>10.5</td>
+      <td>191</td>
+      <td>49.5</td>
+      <td>41</td>
+    </tr>
+    <tr>
+      <th rowspan="3" valign="top">H100</th>
+      <td>1</td>
+      <th>deepseek-ai/DeepSeek-R1-Distill-Qwen-32B</th>
+      <td>---</td>
+      <td>3.3</td>
+      <td>327</td>
+      <td>6.7</td>
+      <td>163</td>
+      <td>3.4</td>
+      <td>320</td>
+      <td>3.4</td>
+      <td>317</td>
+      <td>26.6</td>
+      <td>41</td>
+      <td>26.9</td>
+      <td>41</td>
+      <td>14.3</td>
+      <td>77</td>
+      <td>47.8</td>
+      <td>23</td>
+    </tr>
+    <tr>
+      <td>1</td>
+      <th>neuralmagic/DeepSeek-R1-Distill-Qwen-32B-FP8-dynamic</th>
+      <td>1.52</td>
+      <td>2.2</td>
+      <td>503</td>
+      <td>4.3</td>
+      <td>252</td>
+      <td>2.2</td>
+      <td>490</td>
+      <td>2.3</td>
+      <td>485</td>
+      <td>17.3</td>
+      <td>63</td>
+      <td>17.5</td>
+      <td>63</td>
+      <td>9.5</td>
+      <td>116</td>
+      <td>33.4</td>
+      <td>33</td>
+    </tr>
+    <tr>
+      <td>1</td>
+      <th>neuralmagic/DeepSeek-R1-Distill-Qwen-32B-quantized.w4a16</th>
+      <td>1.61</td>
+      <td>2.1</td>
+      <td>532</td>
+      <td>4.1</td>
+      <td>268</td>
+      <td>2.1</td>
+      <td>516</td>
+      <td>2.1</td>
+      <td>513</td>
+      <td>16.1</td>
+      <td>68</td>
+      <td>16.5</td>
+      <td>66</td>
+      <td>9.1</td>
+      <td>120</td>
+      <td>31.9</td>
+      <td>34</td>
+    </tr>
+  </tbody>
+</table>
+
+**Use case profiles: prompt tokens / generation tokens
+
+**QPD: Queries per dollar, based on on-demand cost at [Lambda Labs](https://lambdalabs.com/service/gpu-cloud) (observed on 2/18/2025).
+
+
+### Multi-stream asynchronous performance (measured with vLLM version 0.7.2)
+<table>
+  <thead>
+    <tr>
+      <th></th>
+      <th></th>
+      <th></th>
+      <th style="text-align: center;" colspan="2" >Instruction Following<br>256 / 128</th>
+      <th style="text-align: center;" colspan="2" >Multi-turn Chat<br>512 / 256</th>
+      <th style="text-align: center;" colspan="2" >Docstring Generation<br>768 / 128</th>
+      <th style="text-align: center;" colspan="2" >RAG<br>1024 / 128</th>
+      <th style="text-align: center;" colspan="2" >Code Completion<br>256 / 1024</th>
+      <th style="text-align: center;" colspan="2" >Code Fixing<br>1024 / 1024</th>
+      <th style="text-align: center;" colspan="2" >Large Summarization<br>4096 / 512</th>
+      <th style="text-align: center;" colspan="2" >Large RAG<br>10240 / 1536</th>
+    </tr>
+    <tr>
+      <th>Hardware</th>
+      <th>Model</th>
+      <th>Average cost reduction</th>
+      <th>Maximum throughput (QPS)</th>
+      <th>QPD</th>
+      <th>Maximum throughput (QPS)</th>
+      <th>QPD</th>
+      <th>Maximum throughput (QPS)</th>
+      <th>QPD</th>
+      <th>Maximum throughput (QPS)</th>
+      <th>QPD</th>
+      <th>Maximum throughput (QPS)</th>
+      <th>QPD</th>
+      <th>Maximum throughput (QPS)</th>
+      <th>QPD</th>
+      <th>Maximum throughput (QPS)</th>
+      <th>QPD</th>
+      <th>Maximum throughput (QPS)</th>
+      <th>QPD</th>
+    </tr>
+  </thead>
+  <tbody style="text-align: center" >
+    <tr>
+      <th rowspan="3" valign="top">A6000x2</th>
+      <th>deepseek-ai/DeepSeek-R1-Distill-Qwen-32B</th>
+      <td>---</td>
+      <td>6.2</td>
+      <td>13940</td>
+      <td>1.9</td>
+      <td>4348</td>
+      <td>2.7</td>
+      <td>6153</td>
+      <td>2.1</td>
+      <td>4778</td>
+      <td>0.6</td>
+      <td>1382</td>
+      <td>0.4</td>
+      <td>930</td>
+      <td>0.3</td>
+      <td>685</td>
+      <td>0.1</td>
+      <td>124</td>
+    </tr>
+    <tr>
+      <th>neuralmagic/DeepSeek-R1-Distill-Qwen-32B-quantized.w8a8</th>
+      <td>1.80</td>
+      <td>8.7</td>
+      <td>19492</td>
+      <td>4.2</td>
+      <td>9474</td>
+      <td>4.1</td>
+      <td>9290</td>
+      <td>3.0</td>
+      <td>6802</td>
+      <td>1.2</td>
+      <td>2734</td>
+      <td>0.9</td>
+      <td>1962</td>
+      <td>0.5</td>
+      <td>1177</td>
+      <td>0.1</td>
+      <td>254</td>
+    </tr>
+    <tr>
+      <th>neuralmagic/DeepSeek-R1-Distill-Qwen-32B-quantized.w4a16</th>
+      <td>1.30</td>
+      <td>5.9</td>
+      <td>13366</td>
+      <td>2.5</td>
+      <td>5733</td>
+      <td>2.4</td>
+      <td>5409</td>
+      <td>1.6</td>
+      <td>3525</td>
+      <td>1.2</td>
+      <td>2757</td>
+      <td>0.7</td>
+      <td>1663</td>
+      <td>0.3</td>
+      <td>676</td>
+      <td>0.1</td>
+      <td>214</td>
+    </tr>
+    <tr>
+      <th rowspan="3" valign="top">A100x2</th>
+      <th>deepseek-ai/DeepSeek-R1-Distill-Qwen-32B</th>
+      <td>---</td>
+      <td>12.9</td>
+      <td>13016</td>
+      <td>5.8</td>
+      <td>5848</td>
+      <td>6.3</td>
+      <td>6348</td>
+      <td>5.1</td>
+      <td>5146</td>
+      <td>2.0</td>
+      <td>1988</td>
+      <td>1.5</td>
+      <td>1463</td>
+      <td>0.9</td>
+      <td>869</td>
+      <td>0.2</td>
+      <td>192</td>
+    </tr>
+    <tr>
+      <th>neuralmagic/DeepSeek-R1-Distill-Qwen-32B-quantized.w8a8</th>
+      <td>1.52</td>
+      <td>21.4</td>
+      <td>21479</td>
+      <td>8.9</td>
+      <td>8948</td>
+      <td>10.6</td>
+      <td>10611</td>
+      <td>8.2</td>
+      <td>8197</td>
+      <td>3.0</td>
+      <td>3018</td>
+      <td>2.0</td>
+      <td>2054</td>
+      <td>1.2</td>
+      <td>1241</td>
+      <td>0.3</td>
+      <td>264</td>
+    </tr>
+    <tr>
+      <th>neuralmagic/DeepSeek-R1-Distill-Qwen-32B-quantized.w4a16</th>
+      <td>1.09</td>
+      <td>13.5</td>
+      <td>13568</td>
+      <td>6.5</td>
+      <td>6509</td>
+      <td>6.0</td>
+      <td>6075</td>
+      <td>4.7</td>
+      <td>4754</td>
+      <td>2.8</td>
+      <td>2790</td>
+      <td>1.6</td>
+      <td>1651</td>
+      <td>0.9</td>
+      <td>862</td>
+      <td>0.2</td>
+      <td>225</td>
+    </tr>
+    <tr>
+      <th rowspan="3" valign="top">H100x2</th>
+      <th>deepseek-ai/DeepSeek-R1-Distill-Qwen-32B</th>
+      <td>---</td>
+      <td>25.5</td>
+      <td>14392</td>
+      <td>12.5</td>
+      <td>7035</td>
+      <td>14.0</td>
+      <td>7877</td>
+      <td>11.3</td>
+      <td>6364</td>
+      <td>3.6</td>
+      <td>2041</td>
+      <td>2.7</td>
+      <td>1549</td>
+      <td>1.9</td>
+      <td>1057</td>
+      <td>0.4</td>
+      <td>200</td>
+    </tr>
+    <tr>
+      <th>neuralmagic/DeepSeek-R1-Distill-Qwen-32B-FP8-dynamic</th>
+      <td>1.46</td>
+      <td>46.7</td>
+      <td>25538</td>
+      <td>20.3</td>
+      <td>11082</td>
+      <td>23.3</td>
+      <td>12728</td>
+      <td>18.4</td>
+      <td>10049</td>
+      <td>5.3</td>
+      <td>2881</td>
+      <td>3.7</td>
+      <td>2097</td>
+      <td>2.6</td>
+      <td>1445</td>
+      <td>0.5</td>
+      <td>256</td>
+    </tr>
+    <tr>
+      <th>neuralmagic/DeepSeek-R1-Distill-Qwen-32B-quantized.w4a16</th>
+      <td>1.23</td>
+      <td>36.9</td>
+      <td>20172</td>
+      <td>17.4</td>
+      <td>9500</td>
+      <td>18.0</td>
+      <td>9822</td>
+      <td>14.2</td>
+      <td>7755</td>
+      <td>5.3</td>
+      <td>2900</td>
+      <td>3.3</td>
+      <td>1867</td>
+      <td>2.3</td>
+      <td>1265</td>
+      <td>0.4</td>
+      <td>241</td>
+    </tr>
+  </tbody>
+</table>
+
+**Use case profiles: prompt tokens / generation tokens
+
+**QPS: Queries per second.
+
+**QPD: Queries per dollar, based on on-demand cost at [Lambda Labs](https://lambdalabs.com/service/gpu-cloud) (observed on 2/18/2025).