Update README.md
Browse files
README.md
CHANGED
|
@@ -65,6 +65,39 @@ export_safetensors(model, output_dir)
|
|
| 65 |
tokenizer.save_pretrained(output_dir)
|
| 66 |
```
|
| 67 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 68 |
# Deployment
|
| 69 |
|
| 70 |
This model can be deployed efficiently using the [vLLM](https://docs.vllm.ai/en/latest/) backends.
|
|
|
|
| 65 |
tokenizer.save_pretrained(output_dir)
|
| 66 |
```
|
| 67 |
|
| 68 |
+
### Accuracy
|
| 69 |
+
|
| 70 |
+
<table>
|
| 71 |
+
<tr>
|
| 72 |
+
<td><strong>Benchmark</strong>
|
| 73 |
+
</td>
|
| 74 |
+
<td><strong>DeepSeek-R1-0528</strong>
|
| 75 |
+
</td>
|
| 76 |
+
<td><strong>DeepSeek-R1-0528-ptpc(this model)</strong>
|
| 77 |
+
</td>
|
| 78 |
+
</tr>
|
| 79 |
+
<tr>
|
| 80 |
+
<td>GPQA-diamond
|
| 81 |
+
</td>
|
| 82 |
+
<td>80.72
|
| 83 |
+
</td>
|
| 84 |
+
<td>80.05
|
| 85 |
+
</td>
|
| 86 |
+
</tr>
|
| 87 |
+
</table>
|
| 88 |
+
|
| 89 |
+
### Reproduction
|
| 90 |
+
|
| 91 |
+
vllm version: 0.10.1.1+rocm710
|
| 92 |
+
|
| 93 |
+
aiter version: 0.1.6.post2.dev55+g59bd8ff2c
|
| 94 |
+
|
| 95 |
+
lighteval version: 0.12.2
|
| 96 |
+
```
|
| 97 |
+
MODEL_ARGS="model_name=/amd/DeepSeek-R1-0528-ptpc,dtype=auto,tensor_parallel_size=8,max_model_length=71536,max_num_batched_tokens=65536,gpu_memory_utilization=0.9,generation_parameters={max_new_tokens:65536,temperature:0.6,top_p:0.95,seed:100}"
|
| 98 |
+
lighteval vllm $MODEL_ARGS "lighteval|gpqa:diamond|0"
|
| 99 |
+
```
|
| 100 |
+
|
| 101 |
# Deployment
|
| 102 |
|
| 103 |
This model can be deployed efficiently using the [vLLM](https://docs.vllm.ai/en/latest/) backends.
|