|
101 | 101 | "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json", |
102 | 102 | "num_prompts": 200 |
103 | 103 | } |
| 104 | + }, |
| 105 | + { |
| 106 | + "test_name": "serving_gpt_oss_20b_tp1_random_in5k_out8k", |
| 107 | + "qps_list": [1, 4, 16, "inf"], |
| 108 | + "server_parameters": { |
| 109 | + "model": "openai/gpt-oss-20b", |
| 110 | + "tensor_parallel_size": 1, |
| 111 | + "load_format": "dummy", |
| 112 | + "context_length": 16384 |
| 113 | + }, |
| 114 | + "client_parameters": { |
| 115 | + "model": "openai/gpt-oss-20b", |
| 116 | + "backend": "vllm", |
| 117 | + "dataset_name": "random", |
| 118 | + "num_prompts": 200, |
| 119 | + "random_input_len": 5250, |
| 120 | + "random_output_len": 8250 |
| 121 | + } |
| 122 | + }, |
| 123 | + { |
| 124 | + "test_name": "serving_gpt_oss_120b_tp4_random_in5k_out8k", |
| 125 | + "qps_list": [1, 4, 16, "inf"], |
| 126 | + "server_parameters": { |
| 127 | + "model": "openai/gpt-oss-120b", |
| 128 | + "tensor_parallel_size": 4, |
| 129 | + "load_format": "dummy", |
| 130 | + "context_length": 16384 |
| 131 | + }, |
| 132 | + "client_parameters": { |
| 133 | + "model": "openai/gpt-oss-120b", |
| 134 | + "backend": "vllm", |
| 135 | + "dataset_name": "random", |
| 136 | + "num_prompts": 200, |
| 137 | + "random_input_len": 5250, |
| 138 | + "random_output_len": 8250 |
| 139 | + } |
| 140 | + }, |
| 141 | + { |
| 142 | + "test_name": "serving_deepseek_v3_tp8_random_in5k_out8k", |
| 143 | + "qps_list": [1, 4, 16, "inf"], |
| 144 | + "server_parameters": { |
| 145 | + "model": "deepseek-ai/DeepSeek-V3.1", |
| 146 | + "tensor_parallel_size": 8, |
| 147 | + "load_format": "dummy", |
| 148 | + "context_length": 16384 |
| 149 | + }, |
| 150 | + "client_parameters": { |
| 151 | + "model": "deepseek-ai/DeepSeek-V3.1", |
| 152 | + "backend": "vllm", |
| 153 | + "dataset_name": "random", |
| 154 | + "num_prompts": 200, |
| 155 | + "random_input_len": 5250, |
| 156 | + "random_output_len": 8250 |
| 157 | + } |
| 158 | + }, |
| 159 | + { |
| 160 | + "test_name": "serving_deepseek_r1_tp8_random_in5k_out8k", |
| 161 | + "qps_list": [1, 4, 16, "inf"], |
| 162 | + "server_parameters": { |
| 163 | + "model": "deepseek-ai/DeepSeek-R1", |
| 164 | + "tensor_parallel_size": 8, |
| 165 | + "load_format": "dummy", |
| 166 | + "context_length": 16384 |
| 167 | + }, |
| 168 | + "client_parameters": { |
| 169 | + "model": "deepseek-ai/DeepSeek-R1", |
| 170 | + "backend": "vllm", |
| 171 | + "dataset_name": "random", |
| 172 | + "num_prompts": 200, |
| 173 | + "random_input_len": 5250, |
| 174 | + "random_output_len": 8250 |
| 175 | + } |
104 | 176 | } |
105 | 177 | ] |
0 commit comments