Update README.md

README.md

@@ -280,3 +280,479 @@ lm_eval \
     </tr>
   </tbody>
 </table>
+
+
+## Inference Performance
+
+
+This model achieves up to 1.6x speedup in both single-stream and 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 0.6.7.2, and [GuideLLM](https://github.com/neuralmagic/guidellm).
+
+<details>
+<summary>Benchmarking Command</summary>
+
+```
+guidellm --model neuralmagic/DeepSeek-R1-Distill-Qwen-7B-quantized.w8a8 --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 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>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">A6000x1</th>
+      <th>deepseek-ai/DeepSeek-R1-Distill-Qwen-7B</th>
+      <td>---</td>
+      <td>2.9</td>
+      <td>1576</td>
+      <td>5.7</td>
+      <td>788</td>
+      <td>2.9</td>
+      <td>1535</td>
+      <td>3.0</td>
+      <td>1496</td>
+      <td>22.6</td>
+      <td>199</td>
+      <td>23.2</td>
+      <td>194</td>
+      <td>12.1</td>
+      <td>370</td>
+      <td>38.5</td>
+      <td>117</td>
+    </tr>
+    <tr>
+      <th>neuralmagic/DeepSeek-R1-Distill-Qwen-7B-quantized.w8a8</th>
+      <td>1.56</td>
+      <td>1.8</td>
+      <td>2495</td>
+      <td>3.7</td>
+      <td>1223</td>
+      <td>1.9</td>
+      <td>2384</td>
+      <td>1.9</td>
+      <td>2393</td>
+      <td>14.3</td>
+      <td>315</td>
+      <td>14.8</td>
+      <td>304</td>
+      <td>7.9</td>
+      <td>572</td>
+      <td>25.3</td>
+      <td>178</td>
+    </tr>
+    <tr>
+      <th>neuralmagic/DeepSeek-R1-Distill-Qwen-7B-quantized.w4a16</th>
+      <td>2.41</td>
+      <td>1.1</td>
+      <td>4086</td>
+      <td>2.3</td>
+      <td>1998</td>
+      <td>1.2</td>
+      <td>3783</td>
+      <td>1.3</td>
+      <td>3527</td>
+      <td>8.6</td>
+      <td>526</td>
+      <td>8.8</td>
+      <td>512</td>
+      <td>5.2</td>
+      <td>860</td>
+      <td>22.7</td>
+      <td>198</td>
+    </tr>
+    <tr>
+      <th rowspan="3" valign="top">A100x1</th>
+      <th>deepseek-ai/DeepSeek-R1-Distill-Qwen-7B</th>
+      <td>---</td>
+      <td>1.4</td>
+      <td>1389</td>
+      <td>2.9</td>
+      <td>691</td>
+      <td>1.5</td>
+      <td>1358</td>
+      <td>1.5</td>
+      <td>1329</td>
+      <td>11.5</td>
+      <td>175</td>
+      <td>11.6</td>
+      <td>174</td>
+      <td>6.2</td>
+      <td>326</td>
+      <td>21.5</td>
+      <td>93</td>
+    </tr>
+    <tr>
+      <th>neuralmagic/DeepSeek-R1-Distill-Qwen-7B-quantized.w8a8</th>
+      <td>1.28</td>
+      <td>1.1</td>
+      <td>1850</td>
+      <td>2.2</td>
+      <td>905</td>
+      <td>1.1</td>
+      <td>1807</td>
+      <td>1.1</td>
+      <td>1750</td>
+      <td>8.6</td>
+      <td>233</td>
+      <td>8.7</td>
+      <td>230</td>
+      <td>4.7</td>
+      <td>431</td>
+      <td>23.1</td>
+      <td>87</td>
+    </tr>
+    <tr>
+      <th>neuralmagic/DeepSeek-R1-Distill-Qwen-7B-quantized.w4a16</th>
+      <td>1.72</td>
+      <td>0.8</td>
+      <td>2575</td>
+      <td>1.5</td>
+      <td>1298</td>
+      <td>0.8</td>
+      <td>2461</td>
+      <td>0.8</td>
+      <td>2382</td>
+      <td>6.1</td>
+      <td>331</td>
+      <td>6.2</td>
+      <td>323</td>
+      <td>3.6</td>
+      <td>566</td>
+      <td>22.7</td>
+      <td>89</td>
+    </tr>
+    <tr>
+      <th rowspan="3" valign="top">H100x1</th>
+      <th>deepseek-ai/DeepSeek-R1-Distill-Qwen-7B</th>
+      <td>---</td>
+      <td>0.9</td>
+      <td>1161</td>
+      <td>1.9</td>
+      <td>579</td>
+      <td>1.0</td>
+      <td>1138</td>
+      <td>1.0</td>
+      <td>1121</td>
+      <td>7.5</td>
+      <td>146</td>
+      <td>7.6</td>
+      <td>145</td>
+      <td>3.9</td>
+      <td>279</td>
+      <td>15.4</td>
+      <td>71</td>
+    </tr>
+    <tr>
+      <th>neuralmagic/DeepSeek-R1-Distill-Qwen-7B-FP8-dynamic</th>
+      <td>1.34</td>
+      <td>0.7</td>
+      <td>1585</td>
+      <td>1.4</td>
+      <td>786</td>
+      <td>0.7</td>
+      <td>1577</td>
+      <td>0.7</td>
+      <td>1524</td>
+      <td>5.3</td>
+      <td>207</td>
+      <td>5.5</td>
+      <td>197</td>
+      <td>2.9</td>
+      <td>382</td>
+      <td>14.3</td>
+      <td>77</td>
+    </tr>
+    <tr>
+      <th>neuralmagic/DeepSeek-R1-Distill-Qwen-7B-quantized.w4a16</th>
+      <td>1.33</td>
+      <td>0.7</td>
+      <td>1590</td>
+      <td>1.4</td>
+      <td>793</td>
+      <td>0.7</td>
+      <td>1549</td>
+      <td>0.7</td>
+      <td>1509</td>
+      <td>5.4</td>
+      <td>201</td>
+      <td>5.5</td>
+      <td>198</td>
+      <td>2.9</td>
+      <td>381</td>
+      <td>14.0</td>
+      <td>78</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">A6000x1</th>
+      <th>deepseek-ai/DeepSeek-R1-Distill-Qwen-7B</th>
+      <td>---</td>
+      <td>14.9</td>
+      <td>67138</td>
+      <td>7.1</td>
+      <td>32094</td>
+      <td>7.4</td>
+      <td>33096</td>
+      <td>5.9</td>
+      <td>26480</td>
+      <td>2.0</td>
+      <td>9004</td>
+      <td>1.5</td>
+      <td>6639</td>
+      <td>1.1</td>
+      <td>4938</td>
+      <td>0.3</td>
+      <td>1151</td>
+    </tr>
+    <tr>
+      <th>neuralmagic/DeepSeek-R1-Distill-Qwen-7B-quantized.w8a8</th>
+      <td>1.36</td>
+      <td>20.2</td>
+      <td>90956</td>
+      <td>8.8</td>
+      <td>39786</td>
+      <td>10.2</td>
+      <td>45963</td>
+      <td>8.1</td>
+      <td>36596</td>
+      <td>3.1</td>
+      <td>13968</td>
+      <td>2.1</td>
+      <td>9629</td>
+      <td>1.4</td>
+      <td>6374</td>
+      <td>0.3</td>
+      <td>1429</td>
+    </tr>
+    <tr>
+      <th>neuralmagic/DeepSeek-R1-Distill-Qwen-7B-quantized.w4a16</th>
+      <td>1.00</td>
+      <td>13.3</td>
+      <td>59681</td>
+      <td>6.1</td>
+      <td>27633</td>
+      <td>5.9</td>
+      <td>26689</td>
+      <td>4.7</td>
+      <td>20944</td>
+      <td>2.9</td>
+      <td>13108</td>
+      <td>1.9</td>
+      <td>8355</td>
+      <td>1.0</td>
+      <td>4362</td>
+      <td>0.3</td>
+      <td>1170</td>
+    </tr>
+    <tr>
+      <th rowspan="3" valign="top">A100x1</th>
+      <th>deepseek-ai/DeepSeek-R1-Distill-Qwen-7B</th>
+      <td>---</td>
+      <td>26.4</td>
+      <td>53073</td>
+      <td>13.0</td>
+      <td>26213</td>
+      <td>14.5</td>
+      <td>29110</td>
+      <td>11.4</td>
+      <td>22936</td>
+      <td>4.4</td>
+      <td>8749</td>
+      <td>3.3</td>
+      <td>6680</td>
+      <td>2.3</td>
+      <td>4634</td>
+      <td>0.5</td>
+      <td>1105</td>
+    </tr>
+    <tr>
+      <th>neuralmagic/DeepSeek-R1-Distill-Qwen-7B-quantized.w8a8</th>
+      <td>1.27</td>
+      <td>34.3</td>
+      <td>69009</td>
+      <td>14.8</td>
+      <td>29791</td>
+      <td>19.0</td>
+      <td>38214</td>
+      <td>15.7</td>
+      <td>31598</td>
+      <td>5.6</td>
+      <td>11186</td>
+      <td>4.2</td>
+      <td>8350</td>
+      <td>3.0</td>
+      <td>6020</td>
+      <td>0.7</td>
+      <td>1328</td>
+    </tr>
+    <tr>
+      <th>neuralmagic/DeepSeek-R1-Distill-Qwen-7B-quantized.w4a16</th>
+      <td>0.93</td>
+      <td>23.9</td>
+      <td>47993</td>
+      <td>12.0</td>
+      <td>24194</td>
+      <td>12.5</td>
+      <td>25239</td>
+      <td>10.0</td>
+      <td>20029</td>
+      <td>4.5</td>
+      <td>9055</td>
+      <td>3.3</td>
+      <td>6681</td>
+      <td>2.1</td>
+      <td>4156</td>
+      <td>0.5</td>
+      <td>1043</td>
+    </tr>
+    <tr>
+      <th rowspan="3" valign="top">H100x1</th>
+      <th>deepseek-ai/DeepSeek-R1-Distill-Qwen-7B</th>
+      <td>---</td>
+      <td>54.3</td>
+      <td>59410</td>
+      <td>26.0</td>
+      <td>28440</td>
+      <td>32.1</td>
+      <td>35154</td>
+      <td>26.7</td>
+      <td>29190</td>
+      <td>8.0</td>
+      <td>8700</td>
+      <td>6.6</td>
+      <td>7275</td>
+      <td>5.2</td>
+      <td>5669</td>
+      <td>1.2</td>
+      <td>1266</td>
+    </tr>
+    <tr>
+      <th>neuralmagic/DeepSeek-R1-Distill-Qwen-7B-FP8-dynamic</th>
+      <td>1.16</td>
+      <td>62.9</td>
+      <td>68818</td>
+      <td>30.3</td>
+      <td>33196</td>
+      <td>39.4</td>
+      <td>43132</td>
+      <td>31.1</td>
+      <td>34073</td>
+      <td>9.2</td>
+      <td>10058</td>
+      <td>7.1</td>
+      <td>7748</td>
+      <td>6.1</td>
+      <td>6714</td>
+      <td>1.3</td>
+      <td>1415</td>
+    </tr>
+    <tr>
+      <th>neuralmagic/DeepSeek-R1-Distill-Qwen-7B-quantized.w4a16</th>
+      <td>1.02</td>
+      <td>56.2</td>
+      <td>61483</td>
+      <td>26.7</td>
+      <td>29243</td>
+      <td>32.5</td>
+      <td>35592</td>
+      <td>26.9</td>
+      <td>29461</td>
+      <td>8.3</td>
+      <td>9072</td>
+      <td>6.4</td>
+      <td>7027</td>
+      <td>5.2</td>
+      <td>5731</td>
+      <td>1.2</td>
+      <td>1291</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).