Update README.md

README.md

@@ -260,3 +260,478 @@ lm_eval \
     </tr>
   </tbody>
 </table>
+
+## Inference Performance
+
+
+This model achieves up to 1.4x speedup in single-stream deployment and up to 1.3x 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-Llama-8B-FP8-dynamic --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-Llama-8B</th>
+      <td>---</td>
+      <td>3.0</td>
+      <td>1511</td>
+      <td>6.0</td>
+      <td>755</td>
+      <td>3.0</td>
+      <td>1483</td>
+      <td>3.1</td>
+      <td>1462</td>
+      <td>23.6</td>
+      <td>191</td>
+      <td>24.0</td>
+      <td>188</td>
+      <td>12.7</td>
+      <td>353</td>
+      <td>41.1</td>
+      <td>110</td>
+    </tr>
+    <tr>
+      <th>neuralmagic/DeepSeek-R1-Distill-Llama-8B-quantized.w8a8</th>
+      <td>1.53</td>
+      <td>1.9</td>
+      <td>2356</td>
+      <td>3.8</td>
+      <td>1175</td>
+      <td>2.0</td>
+      <td>2291</td>
+      <td>2.0</td>
+      <td>2207</td>
+      <td>15.2</td>
+      <td>297</td>
+      <td>15.5</td>
+      <td>290</td>
+      <td>8.5</td>
+      <td>531</td>
+      <td>28.6</td>
+      <td>157</td>
+    </tr>
+    <tr>
+      <th>neuralmagic/DeepSeek-R1-Distill-Llama-8B-quantized.w4a16</th>
+      <td>2.35</td>
+      <td>1.2</td>
+      <td>3870</td>
+      <td>2.3</td>
+      <td>1918</td>
+      <td>1.3</td>
+      <td>3492</td>
+      <td>1.3</td>
+      <td>3335</td>
+      <td>9.1</td>
+      <td>492</td>
+      <td>9.5</td>
+      <td>472</td>
+      <td>5.8</td>
+      <td>771</td>
+      <td>22.7</td>
+      <td>198</td>
+    </tr>
+    <tr>
+      <th rowspan="3" valign="top">A100x1</th>
+      <th>deepseek-ai/DeepSeek-R1-Distill-Llama-8B</th>
+      <td>---</td>
+      <td>1.5</td>
+      <td>1308</td>
+      <td>3.1</td>
+      <td>657</td>
+      <td>1.6</td>
+      <td>1274</td>
+      <td>1.6</td>
+      <td>1263</td>
+      <td>12.1</td>
+      <td>166</td>
+      <td>12.4</td>
+      <td>162</td>
+      <td>6.5</td>
+      <td>308</td>
+      <td>25.6</td>
+      <td>78</td>
+    </tr>
+    <tr>
+      <th>neuralmagic/DeepSeek-R1-Distill-Llama-8B-quantized.w8a8</th>
+      <td>1.30</td>
+      <td>1.1</td>
+      <td>1763</td>
+      <td>2.3</td>
+      <td>882</td>
+      <td>1.2</td>
+      <td>1716</td>
+      <td>1.2</td>
+      <td>1698</td>
+      <td>9.0</td>
+      <td>223</td>
+      <td>9.2</td>
+      <td>218</td>
+      <td>4.9</td>
+      <td>409</td>
+      <td>25.7</td>
+      <td>78</td>
+    </tr>
+    <tr>
+      <th>neuralmagic/DeepSeek-R1-Distill-Llama-8B-quantized.w4a16</th>
+      <td>1.76</td>
+      <td>0.8</td>
+      <td>2501</td>
+      <td>1.6</td>
+      <td>1236</td>
+      <td>0.9</td>
+      <td>2350</td>
+      <td>0.9</td>
+      <td>2287</td>
+      <td>6.4</td>
+      <td>316</td>
+      <td>6.6</td>
+      <td>306</td>
+      <td>3.7</td>
+      <td>544</td>
+      <td>24.7</td>
+      <td>82</td>
+    </tr>
+    <tr>
+      <th rowspan="3" valign="top">H100x1</th>
+      <th>deepseek-ai/DeepSeek-R1-Distill-Llama-8B</th>
+      <td>---</td>
+      <td>1.0</td>
+      <td>1146</td>
+      <td>1.9</td>
+      <td>574</td>
+      <td>1.0</td>
+      <td>1128</td>
+      <td>1.0</td>
+      <td>1111</td>
+      <td>7.6</td>
+      <td>144</td>
+      <td>7.7</td>
+      <td>142</td>
+      <td>4.1</td>
+      <td>266</td>
+      <td>16.3</td>
+      <td>67</td>
+    </tr>
+    <tr>
+      <th>neuralmagic/DeepSeek-R1-Distill-Llama-8B-FP8-dynamic</th>
+      <td>1.25</td>
+      <td>0.7</td>
+      <td>1567</td>
+      <td>1.4</td>
+      <td>758</td>
+      <td>0.7</td>
+      <td>1484</td>
+      <td>0.7</td>
+      <td>1462</td>
+      <td>5.7</td>
+      <td>191</td>
+      <td>5.8</td>
+      <td>189</td>
+      <td>3.2</td>
+      <td>347</td>
+      <td>22.5</td>
+      <td>49</td>
+    </tr>
+    <tr>
+      <th>neuralmagic/DeepSeek-R1-Distill-Llama-8B-quantized.w4a16</th>
+      <td>1.30</td>
+      <td>0.7</td>
+      <td>1527</td>
+      <td>1.4</td>
+      <td>768</td>
+      <td>0.7</td>
+      <td>1495</td>
+      <td>0.7</td>
+      <td>1463</td>
+      <td>5.6</td>
+      <td>194</td>
+      <td>5.7</td>
+      <td>190</td>
+      <td>3.1</td>
+      <td>350</td>
+      <td>14.7</td>
+      <td>74</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-Llama-8B</th>
+      <td>---</td>
+      <td>12.6</td>
+      <td>56742</td>
+      <td>5.7</td>
+      <td>25687</td>
+      <td>6.5</td>
+      <td>29349</td>
+      <td>5.2</td>
+      <td>23259</td>
+      <td>1.6</td>
+      <td>7250</td>
+      <td>1.2</td>
+      <td>5181</td>
+      <td>0.8</td>
+      <td>3445</td>
+      <td>0.1</td>
+      <td>616</td>
+    </tr>
+    <tr>
+      <th>neuralmagic/DeepSeek-R1-Distill-Llama-8B-quantized.w8a8</th>
+      <td>1.34</td>
+      <td>17.4</td>
+      <td>78101</td>
+      <td>7.6</td>
+      <td>34351</td>
+      <td>8.8</td>
+      <td>39790</td>
+      <td>7.0</td>
+      <td>31532</td>
+      <td>2.3</td>
+      <td>10405</td>
+      <td>1.5</td>
+      <td>6960</td>
+      <td>1.0</td>
+      <td>4355</td>
+      <td>0.2</td>
+      <td>785</td>
+    </tr>
+    <tr>
+      <th>neuralmagic/DeepSeek-R1-Distill-Llama-8B-quantized.w4a16</th>
+      <td>0.91</td>
+      <td>10.9</td>
+      <td>48964</td>
+      <td>5.1</td>
+      <td>22989</td>
+      <td>4.8</td>
+      <td>21791</td>
+      <td>3.8</td>
+      <td>17039</td>
+      <td>2.2</td>
+      <td>9726</td>
+      <td>1.2</td>
+      <td>5434</td>
+      <td>0.6</td>
+      <td>2544</td>
+      <td>0.1</td>
+      <td>578</td>
+    </tr>
+    <tr>
+      <th rowspan="3" valign="top">A100x1</th>
+      <th>deepseek-ai/DeepSeek-R1-Distill-Llama-8B</th>
+      <td>---</td>
+      <td>24.5</td>
+      <td>49296</td>
+      <td>11.3</td>
+      <td>22657</td>
+      <td>13.0</td>
+      <td>26047</td>
+      <td>10.5</td>
+      <td>21020</td>
+      <td>3.5</td>
+      <td>7029</td>
+      <td>2.5</td>
+      <td>4995</td>
+      <td>1.7</td>
+      <td>3503</td>
+      <td>0.3</td>
+      <td>659</td>
+    </tr>
+    <tr>
+      <th>neuralmagic/DeepSeek-R1-Distill-Llama-8B-quantized.w8a8</th>
+      <td>1.27</td>
+      <td>30.8</td>
+      <td>62042</td>
+      <td>14.1</td>
+      <td>28419</td>
+      <td>17.2</td>
+      <td>34554</td>
+      <td>13.8</td>
+      <td>27719</td>
+      <td>4.6</td>
+      <td>9299</td>
+      <td>3.1</td>
+      <td>6215</td>
+      <td>2.2</td>
+      <td>4331</td>
+      <td>0.4</td>
+      <td>807</td>
+    </tr>
+    <tr>
+      <th>neuralmagic/DeepSeek-R1-Distill-Llama-8B-quantized.w4a16</th>
+      <td>0.97</td>
+      <td>22.7</td>
+      <td>45708</td>
+      <td>10.5</td>
+      <td>21216</td>
+      <td>11.1</td>
+      <td>22353</td>
+      <td>8.9</td>
+      <td>17939</td>
+      <td>3.9</td>
+      <td>7758</td>
+      <td>2.6</td>
+      <td>5241</td>
+      <td>1.6</td>
+      <td>3196</td>
+      <td>0.4</td>
+      <td>718</td>
+    </tr>
+    <tr>
+      <th rowspan="3" valign="top">H100x1</th>
+      <th>deepseek-ai/DeepSeek-R1-Distill-Llama-8B</th>
+      <td>---</td>
+      <td>49.0</td>
+      <td>53593</td>
+      <td>22.6</td>
+      <td>24750</td>
+      <td>28.3</td>
+      <td>30971</td>
+      <td>22.9</td>
+      <td>25035</td>
+      <td>7.2</td>
+      <td>7912</td>
+      <td>5.1</td>
+      <td>5561</td>
+      <td>3.6</td>
+      <td>3939</td>
+      <td>0.6</td>
+      <td>703</td>
+    </tr>
+    <tr>
+      <th>neuralmagic/DeepSeek-R1-Distill-Llama-8B-FP8-dynamic</th>
+      <td>1.14</td>
+      <td>57.1</td>
+      <td>62517</td>
+      <td>26.0</td>
+      <td>28440</td>
+      <td>34.5</td>
+      <td>37781</td>
+      <td>28.7</td>
+      <td>31360</td>
+      <td>7.2</td>
+      <td>7877</td>
+      <td>5.4</td>
+      <td>5923</td>
+      <td>4.3</td>
+      <td>4697</td>
+      <td>0.7</td>
+      <td>782</td>
+    </tr>
+    <tr>
+      <th>neuralmagic/DeepSeek-R1-Distill-Llama-8B-quantized.w4a16</th>
+      <td>1.01</td>
+      <td>49.8</td>
+      <td>54452</td>
+      <td>22.9</td>
+      <td>25035</td>
+      <td>28.5</td>
+      <td>31162</td>
+      <td>23.0</td>
+      <td>25200</td>
+      <td>6.8</td>
+      <td>7493</td>
+      <td>5.0</td>
+      <td>5431</td>
+      <td>3.7</td>
+      <td>4079</td>
+      <td>0.7</td>
+      <td>787</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).