add gemma-3-27b-it and qwen3_30B-A3B

Author: BoyuanFeng

vllm-benchmarks/benchmarks/serving-tests.json

@@ -117,5 +117,125 @@
             "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json",
             "num_prompts": 200
         }
+    },
+    {
+        "test_name": "serving_qwen3_30B-A3B_tp8_in200_out200",
+        "qps_list": ["inf"],
+        "server_parameters": {
+            "model": "Qwen/Qwen3-30B-A3B",
+            "tensor_parallel_size": 8,
+            "swap_space": 16,
+            "disable_log_stats": "",
+            "disable_log_requests": "",
+            "load_format": "dummy",
+            "max_model_len": 8192
+        },
+        "client_parameters": {
+            "model": "Qwen/Qwen3-30B-A3B",
+            "backend": "vllm",
+            "dataset_name": "random",
+            "random_input_len": 200,
+            "random_output_len": 200
+        }
+    },
+    {
+        "test_name": "serving_qwen3_30B-A3B_tp8_in1k_out2k",
+        "qps_list": ["inf"],
+        "server_parameters": {
+            "model": "Qwen/Qwen3-30B-A3B",
+            "tensor_parallel_size": 8,
+            "swap_space": 16,
+            "disable_log_stats": "",
+            "disable_log_requests": "",
+            "load_format": "dummy",
+            "max_model_len": 8192
+        },
+        "client_parameters": {
+            "model": "Qwen/Qwen3-30B-A3B",
+            "backend": "vllm",
+            "dataset_name": "random",
+            "random_input_len": 1024,
+            "random_output_len": 2048
+        }
+    },
+    {
+        "test_name": "serving_qwen3_30B-A3B_tp8_in5k_out1k",
+        "qps_list": ["inf"],
+        "server_parameters": {
+            "model": "Qwen/Qwen3-30B-A3B",
+            "tensor_parallel_size": 8,
+            "swap_space": 16,
+            "disable_log_stats": "",
+            "disable_log_requests": "",
+            "load_format": "dummy",
+            "max_model_len": 8192
+        },
+        "client_parameters": {
+            "model": "Qwen/Qwen3-30B-A3B",
+            "backend": "vllm",
+            "dataset_name": "random",
+            "random_input_len": 5120,
+            "random_output_len": 1024
+        }
+    },
+    {
+        "test_name": "serving_google/gemma_3_27b_it_tp8_in200_out200",
+        "qps_list": ["inf"],
+        "server_parameters": {
+            "model": "google/gemma-3-27b-it",
+            "tensor_parallel_size": 8,
+            "swap_space": 16,
+            "disable_log_stats": "",
+            "disable_log_requests": "",
+            "load_format": "dummy",
+            "max_model_len": 8192
+        },
+        "client_parameters": {
+            "model": "google/gemma-3-27b-it",
+            "backend": "vllm",
+            "dataset_name": "random",
+            "random_input_len": 200,
+            "random_output_len": 200
+        }
+    },
+    {
+        "test_name": "serving_google/gemma_3_27b_it_tp8_in1k_out2k",
+        "qps_list": ["inf"],
+        "server_parameters": {
+            "model": "google/gemma-3-27b-it",
+            "tensor_parallel_size": 8,
+            "swap_space": 16,
+            "disable_log_stats": "",
+            "disable_log_requests": "",
+            "load_format": "dummy",
+            "max_model_len": 8192
+        },
+        "client_parameters": {
+            "model": "google/gemma-3-27b-it",
+            "backend": "vllm",
+            "dataset_name": "random",
+            "random_input_len": 1024,
+            "random_output_len": 2048
+        }
+    },
+    {
+        "test_name": "serving_google/gemma_3_27b_it_tp8_in5k_out1k",
+        "qps_list": ["inf"],
+        "server_parameters": {
+            "model": "google/gemma-3-27b-it",
+            "tensor_parallel_size": 8,
+            "swap_space": 16,
+            "disable_log_stats": "",
+            "disable_log_requests": "",
+            "load_format": "dummy",
+            "max_model_len": 8192
+        },
+        "client_parameters": {
+            "model": "google/gemma-3-27b-it",
+            "backend": "vllm",
+            "dataset_name": "random",
+            "random_input_len": 5120,
+            "random_output_len": 1024
+        }
     }
 ]