| 1 | { |
| 2 | "defaults": { |
| 3 | "qps_list": [ |
| 4 | "inf" |
| 5 | ], |
| 6 | "max_concurrency_list": [ |
| 7 | 32, |
| 8 | 64, |
| 9 | 128 |
| 10 | ], |
| 11 | "server_environment_variables": { |
| 12 | "VLLM_ALLOW_LONG_MAX_MODEL_LEN": 1, |
| 13 | "VLLM_ENGINE_ITERATION_TIMEOUT_S": 120, |
| 14 | "VLLM_CPU_SGL_KERNEL": 1, |
| 15 | "VLLM_CPU_KVCACHE_SPACE": 40 |
| 16 | }, |
| 17 | "server_parameters": { |
| 18 | "dtype": "bfloat16", |
| 19 | "model": "jinaai/jina-embeddings-v3", |
| 20 | "trust_remote_code": "" |
| 21 | }, |
| 22 | "client_parameters": { |
| 23 | "model": "jinaai/jina-embeddings-v3", |
| 24 | "backend": "openai-embeddings", |
| 25 | "endpoint": "/v1/embeddings", |
| 26 | "dataset_name": "sharegpt", |
| 27 | "dataset_path": "ShareGPT_V3_unfiltered_cleaned_split.json", |
| 28 | "num_prompts": 200 |
| 29 | } |
| 30 | }, |
| 31 | "tests": [ |
| 32 | { |
| 33 | "test_name": "serving_jina_embed_v3_tp1_sharegpt", |
| 34 | "server_parameters": { |
| 35 | "tensor_parallel_size": 1 |
| 36 | }, |
| 37 | "client_parameters": {} |
| 38 | } |
| 39 | ] |
| 40 | } |