main
json 288 lines 7.71 KB
Raw
1 {
2 "defaults": {
3 "qps_list": [
4 "inf"
5 ],
6 "max_concurrency_list": [12, 16, 24, 32, 64, 128, 200],
7 "server_environment_variables": {
8 "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1,
9 "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120,
10 "VLLM_CPU_SGL_KERNEL": 1,
11 "VLLM_CPU_KVCACHE_SPACE": 40
12 },
13 "server_parameters": {
14 "model": "meta-llama/Llama-3.1-8B-Instruct",
15 "tensor_parallel_size": 1,
16 "dtype": "bfloat16",
17 "distributed_executor_backend": "mp",
18 "block_size": 128,
19 "trust_remote_code": "",
20 "disable_log_stats": "",
21 "max_num_batched_tokens": 2048,
22 "max_num_seqs": 256
23 },
24 "client_parameters": {
25 "model": "meta-llama/Llama-3.1-8B-Instruct",
26 "backend": "vllm",
27 "ignore-eos": "",
28 "temperature": 0,
29 "num_prompts": 200
30 }
31 },
32 "tests": [
33 {
34 "test_name": "serving_llama8B_tp1_random_128_128",
35 "server_parameters": {
36 },
37 "client_parameters": {
38 "dataset_name": "random",
39 "random-input-len": 128,
40 "random-output-len": 128
41 }
42 },
43 {
44 "test_name": "serving_llama8B_int4_tp1_random_128_128",
45 "server_parameters": {
46 "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4"
47 },
48 "client_parameters": {
49 "model": "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
50 "dataset_name": "random",
51 "random-input-len": 128,
52 "random-output-len": 128
53 }
54 },
55 {
56 "test_name": "serving_llama8B_int8_tp1_random_128_128",
57 "server_parameters": {
58 "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8"
59 },
60 "client_parameters": {
61 "model": "RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",
62 "dataset_name": "random",
63 "random-input-len": 128,
64 "random-output-len": 128
65 }
66 },
67 {
68 "test_name": "serving_llama1B_tp1_random_128_128",
69 "server_parameters": {
70 "model": "meta-llama/Llama-3.2-1B"
71 },
72 "client_parameters": {
73 "model": "meta-llama/Llama-3.2-1B",
74 "dataset_name": "random",
75 "random-input-len": 128,
76 "random-output-len": 128
77 }
78 },
79 {
80 "test_name": "serving_llama3B_tp1_random_128_128",
81 "server_parameters": {
82 "model": "meta-llama/Llama-3.2-3B-Instruct"
83 },
84 "client_parameters": {
85 "model": "meta-llama/Llama-3.2-3B-Instruct",
86 "dataset_name": "random",
87 "random-input-len": 128,
88 "random-output-len": 128
89 }
90 },
91 {
92 "test_name": "serving_llama70B_tp1_random_128_128",
93 "server_parameters": {
94 "model": "meta-llama/Llama-3.3-70B-Instruct"
95 },
96 "client_parameters": {
97 "model": "meta-llama/Llama-3.3-70B-Instruct",
98 "dataset_name": "random",
99 "random-input-len": 128,
100 "random-output-len": 128
101 }
102 },
103 {
104 "test_name": "serving_granite2B_tp1_random_128_128",
105 "server_parameters": {
106 "model": "ibm-granite/granite-3.2-2b-instruct"
107 },
108 "client_parameters": {
109 "model": "ibm-granite/granite-3.2-2b-instruct",
110 "dataset_name": "random",
111 "random-input-len": 128,
112 "random-output-len": 128
113 }
114 },
115 {
116 "test_name": "serving_qwen1.7B_tp1_random_128_128",
117 "server_parameters": {
118 "model": "Qwen/Qwen3-1.7B"
119 },
120 "client_parameters": {
121 "model": "Qwen/Qwen3-1.7B",
122 "dataset_name": "random",
123 "random-input-len": 128,
124 "random-output-len": 128
125 }
126 },
127 {
128 "test_name": "serving_qwen4B_tp1_random_128_128",
129 "server_parameters": {
130 "model": "Qwen/Qwen3-4B"
131 },
132 "client_parameters": {
133 "model": "Qwen/Qwen3-4B",
134 "dataset_name": "random",
135 "random-input-len": 128,
136 "random-output-len": 128
137 }
138 },
139 {
140 "test_name": "serving_qwen8B_tp1_random_128_128",
141 "server_parameters": {
142 "model": "Qwen/Qwen3-8B"
143 },
144 "client_parameters": {
145 "model": "Qwen/Qwen3-8B",
146 "dataset_name": "random",
147 "random-input-len": 128,
148 "random-output-len": 128
149 }
150 },
151 {
152 "test_name": "serving_qwen14B_tp1_random_128_128",
153 "server_parameters": {
154 "model": "Qwen/Qwen3-14B"
155 },
156 "client_parameters": {
157 "model": "Qwen/Qwen3-14B",
158 "dataset_name": "random",
159 "random-input-len": 128,
160 "random-output-len": 128
161 }
162 },
163 {
164 "test_name": "serving_qwen30B_tp1_random_128_128",
165 "server_parameters": {
166 "model": "Qwen/Qwen3-30B-A3B"
167 },
168 "client_parameters": {
169 "model": "Qwen/Qwen3-30B-A3B",
170 "dataset_name": "random",
171 "random-input-len": 128,
172 "random-output-len": 128
173 }
174 },
175 {
176 "test_name": "serving_glm9B_tp1_random_128_128",
177 "server_parameters": {
178 "model": "zai-org/glm-4-9b-hf"
179 },
180 "client_parameters": {
181 "model": "zai-org/glm-4-9b-hf",
182 "dataset_name": "random",
183 "random-input-len": 128,
184 "random-output-len": 128
185 }
186 },
187 {
188 "test_name": "serving_gemma7B_tp1_random_128_128",
189 "server_parameters": {
190 "model": "google/gemma-7b"
191 },
192 "client_parameters": {
193 "model": "google/gemma-7b",
194 "dataset_name": "random",
195 "random-input-len": 128,
196 "random-output-len": 128
197 }
198 },
199 {
200 "test_name": "serving_gemma3-4b_tp1_random_128_128",
201 "server_environment_variables": {
202 "VLLM_CPU_SGL_KERNEL": 0
203 },
204 "server_parameters": {
205 "model": "google/gemma-3-4b-it"
206 },
207 "client_parameters": {
208 "model": "google/gemma-3-4b-it",
209 "dataset_name": "random",
210 "random-input-len": 128,
211 "random-output-len": 128
212 }
213 },
214 {
215 "test_name": "serving_gemma3-12b_tp1_random_128_128",
216 "server_environment_variables": {
217 "VLLM_CPU_SGL_KERNEL": 0
218 },
219 "server_parameters": {
220 "model": "google/gemma-3-12b-it"
221 },
222 "client_parameters": {
223 "model": "google/gemma-3-12b-it",
224 "dataset_name": "random",
225 "random-input-len": 128,
226 "random-output-len": 128
227 }
228 },
229 {
230 "test_name": "serving_gemma4-4b_tp1_random_128_128",
231 "server_environment_variables": {
232 "VLLM_CPU_SGL_KERNEL": 0
233 },
234 "server_parameters": {
235 "model": "google/gemma-4-E4B-it"
236 },
237 "client_parameters": {
238 "model": "google/gemma-4-E4B-it",
239 "dataset_name": "random",
240 "random-input-len": 128,
241 "random-output-len": 128
242 }
243 },
244 {
245 "test_name": "serving_gemma4-2b_tp1_random_128_128",
246 "server_environment_variables": {
247 "VLLM_CPU_SGL_KERNEL": 0
248 },
249 "server_parameters": {
250 "model": "google/gemma-4-E2B-it"
251 },
252 "client_parameters": {
253 "model": "google/gemma-4-E2B-it",
254 "dataset_name": "random",
255 "random-input-len": 128,
256 "random-output-len": 128
257 }
258 },
259 {
260 "test_name": "serving_gemma4-26b_tp1_random_128_128",
261 "server_environment_variables": {
262 "VLLM_CPU_SGL_KERNEL": 0,
263 "VLLM_CPU_ATTN_SPLIT_KV": 0
264 },
265 "server_parameters": {
266 "model": "google/gemma-4-26B-A4B-it"
267 },
268 "client_parameters": {
269 "model": "google/gemma-4-26B-A4B-it",
270 "dataset_name": "random",
271 "random-input-len": 128,
272 "random-output-len": 128
273 }
274 },
275 {
276 "test_name": "serving_phi4_tp1_random_128_128",
277 "server_parameters": {
278 "model": "microsoft/Phi-4-reasoning"
279 },
280 "client_parameters": {
281 "model": "microsoft/Phi-4-reasoning",
282 "dataset_name": "random",
283 "random-input-len": 128,
284 "random-output-len": 128
285 }
286 }
287 ]
288 }