Update README.md

README.md

@@ -261,3 +261,246 @@ lm_eval \
     </tr>
   </tbody>
 </table>
+
+## Inference Performance
+
+
+This model achieves up to 1.1x speedup in single-stream 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-1.5B-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 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-1.5B</th>
+      <td>---</td>
+      <td>0.8</td>
+      <td>5667</td>
+      <td>1.6</td>
+      <td>2776</td>
+      <td>0.8</td>
+      <td>5515</td>
+      <td>0.8</td>
+      <td>5466</td>
+      <td>6.4</td>
+      <td>705</td>
+      <td>6.5</td>
+      <td>697</td>
+      <td>3.5</td>
+      <td>1295</td>
+      <td>18.3</td>
+      <td>246</td>
+    </tr>
+    <tr>
+      <th>neuralmagic/DeepSeek-R1-Distill-Qwen-1.5B-quantized.w8a8</th>
+      <td>1.14</td>
+      <td>0.7</td>
+      <td>6635</td>
+      <td>1.3</td>
+      <td>3340</td>
+      <td>0.7</td>
+      <td>6396</td>
+      <td>0.7</td>
+      <td>6343</td>
+      <td>5.3</td>
+      <td>845</td>
+      <td>5.4</td>
+      <td>832</td>
+      <td>2.9</td>
+      <td>1547</td>
+      <td>21.3</td>
+      <td>211</td>
+    </tr>
+    <tr>
+      <th>neuralmagic/DeepSeek-R1-Distill-Qwen-1.5B-quantized.w4a16</th>
+      <td>1.38</td>
+      <td>0.5</td>
+      <td>8293</td>
+      <td>1.1</td>
+      <td>4184</td>
+      <td>0.6</td>
+      <td>7976</td>
+      <td>0.6</td>
+      <td>7504</td>
+      <td>4.3</td>
+      <td>1051</td>
+      <td>4.4</td>
+      <td>1033</td>
+      <td>2.5</td>
+      <td>1819</td>
+      <td>21.1</td>
+      <td>213</td>
+    </tr>
+    <tr>
+      <th rowspan="3" valign="top">A100x1</th>
+      <th>deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B</th>
+      <td>---</td>
+      <td>0.6</td>
+      <td>3359</td>
+      <td>1.2</td>
+      <td>1654</td>
+      <td>0.6</td>
+      <td>3286</td>
+      <td>0.6</td>
+      <td>3241</td>
+      <td>4.7</td>
+      <td>424</td>
+      <td>4.9</td>
+      <td>411</td>
+      <td>2.6</td>
+      <td>778</td>
+      <td>21.1</td>
+      <td>95</td>
+    </tr>
+    <tr>
+      <th>neuralmagic/DeepSeek-R1-Distill-Qwen-1.5B-quantized.w8a8</th>
+      <td>1.05</td>
+      <td>0.6</td>
+      <td>3531</td>
+      <td>1.1</td>
+      <td>1807</td>
+      <td>0.6</td>
+      <td>3427</td>
+      <td>0.6</td>
+      <td>3480</td>
+      <td>4.5</td>
+      <td>448</td>
+      <td>4.5</td>
+      <td>447</td>
+      <td>2.4</td>
+      <td>842</td>
+      <td>23.5</td>
+      <td>86</td>
+    </tr>
+    <tr>
+      <th>neuralmagic/DeepSeek-R1-Distill-Qwen-1.5B-quantized.w4a16</th>
+      <td>1.03</td>
+      <td>0.6</td>
+      <td>3469</td>
+      <td>1.1</td>
+      <td>1751</td>
+      <td>0.6</td>
+      <td>3403</td>
+      <td>0.6</td>
+      <td>3407</td>
+      <td>4.5</td>
+      <td>447</td>
+      <td>4.6</td>
+      <td>435</td>
+      <td>2.5</td>
+      <td>815</td>
+      <td>23.3</td>
+      <td>86</td>
+    </tr>
+    <tr>
+      <th rowspan="3" valign="top">H100x1</th>
+      <th>deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B</th>
+      <td>---</td>
+      <td>0.4</td>
+      <td>2604</td>
+      <td>0.8</td>
+      <td>1299</td>
+      <td>0.4</td>
+      <td>2543</td>
+      <td>0.4</td>
+      <td>2551</td>
+      <td>3.3</td>
+      <td>330</td>
+      <td>3.4</td>
+      <td>326</td>
+      <td>1.8</td>
+      <td>612</td>
+      <td>14.0</td>
+      <td>78</td>
+    </tr>
+    <tr>
+      <th>neuralmagic/DeepSeek-R1-Distill-Qwen-1.5B-FP8-dynamic</th>
+      <td>1.04</td>
+      <td>0.4</td>
+      <td>2694</td>
+      <td>0.8</td>
+      <td>1364</td>
+      <td>0.4</td>
+      <td>2670</td>
+      <td>0.4</td>
+      <td>2639</td>
+      <td>3.2</td>
+      <td>347</td>
+      <td>3.2</td>
+      <td>341</td>
+      <td>1.6</td>
+      <td>673</td>
+      <td>14.1</td>
+      <td>78</td>
+    </tr>
+    <tr>
+      <th>neuralmagic/DeepSeek-R1-Distill-Qwen-1.5B-quantized.w4a16</th>
+      <td>0.84</td>
+      <td>0.5</td>
+      <td>2111</td>
+      <td>1.0</td>
+      <td>1065</td>
+      <td>0.5</td>
+      <td>2068</td>
+      <td>0.5</td>
+      <td>2119</td>
+      <td>4.1</td>
+      <td>270</td>
+      <td>4.1</td>
+      <td>265</td>
+      <td>2.1</td>
+      <td>530</td>
+      <td>15.1</td>
+      <td>73</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).