main
json 142 lines 3.58 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_sharegpt",
35 "server_parameters": {
36 "tensor_parallel_size": 1
37 },
38 "client_parameters": {
39 "dataset_name": "sharegpt",
40 "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json"
41 }
42 },
43 {
44 "test_name": "serving_llama8B_tp2_sharegpt",
45 "server_parameters": {
46 "tensor_parallel_size": 2
47 },
48 "client_parameters": {
49 "dataset_name": "sharegpt",
50 "dataset_path": "./ShareGPT_V3_unfiltered_cleaned_split.json"
51 }
52 },
53 {
54 "test_name": "serving_llama8B_tp1_random_128_128",
55 "server_parameters": {
56 "tensor_parallel_size": 1
57 },
58 "client_parameters": {
59 "dataset_name": "random",
60 "random-input-len": 128,
61 "random-output-len": 128
62 }
63 },
64 {
65 "test_name": "serving_llama8B_tp2_random_128_128",
66 "server_parameters": {
67 "tensor_parallel_size": 2
68 },
69 "client_parameters": {
70 "dataset_name": "random",
71 "random-input-len": 128,
72 "random-output-len": 128
73 }
74 },
75 {
76 "test_name": "serving_llama8B_tp1_random_128_2048",
77 "server_parameters": {
78 "tensor_parallel_size": 1
79 },
80 "client_parameters": {
81 "dataset_name": "random",
82 "random-input-len": 128,
83 "random-output-len": 2048
84 }
85 },
86 {
87 "test_name": "serving_llama8B_tp2_random_128_2048",
88 "server_parameters": {
89 "tensor_parallel_size": 2
90 },
91 "client_parameters": {
92 "dataset_name": "random",
93 "random-input-len": 128,
94 "random-output-len": 2048
95 }
96 },
97 {
98 "test_name": "serving_llama8B_tp1_random_2048_128",
99 "server_parameters": {
100 "tensor_parallel_size": 1
101 },
102 "client_parameters": {
103 "dataset_name": "random",
104 "random-input-len": 2048,
105 "random-output-len": 128
106 }
107 },
108 {
109 "test_name": "serving_llama8B_tp2_random_2048_128",
110 "server_parameters": {
111 "tensor_parallel_size": 2
112 },
113 "client_parameters": {
114 "dataset_name": "random",
115 "random-input-len": 2048,
116 "random-output-len": 128
117 }
118 },
119 {
120 "test_name": "serving_llama8B_tp1_random_2048_2048",
121 "server_parameters": {
122 "tensor_parallel_size": 1
123 },
124 "client_parameters": {
125 "dataset_name": "random",
126 "random-input-len": 2048,
127 "random-output-len": 2048
128 }
129 },
130 {
131 "test_name": "serving_llama8B_tp2_random_2048_2048",
132 "server_parameters": {
133 "tensor_parallel_size": 2
134 },
135 "client_parameters": {
136 "dataset_name": "random",
137 "random-input-len": 2048,
138 "random-output-len": 2048
139 }
140 }
141 ]
142 }