mirror of
https://github.com/JustVugg/colibri.git
synced 2026-10-02 02:54:37 +08:00
74 lines
2.4 KiB
JSON
74 lines
2.4 KiB
JSON
{
|
|
"schema": "colibri.serving-baseline/v1",
|
|
"experiment": "REPLACE-experiment-id",
|
|
"comparison": "deployment",
|
|
"residency": "fully_resident",
|
|
"workload": "workload.jsonl",
|
|
"hardware": {
|
|
"host": "REPLACE-host",
|
|
"cpu": "REPLACE-cpu",
|
|
"ram": "REPLACE-ram",
|
|
"gpu": "REPLACE-gpu",
|
|
"storage": "REPLACE-storage",
|
|
"os": "REPLACE-os",
|
|
"driver": "REPLACE-driver"
|
|
},
|
|
"model": {
|
|
"source_revision": "REPLACE-source_revision",
|
|
"tokenizer_sha256": "REPLACE-tokenizer_sha256",
|
|
"chat_template_sha256": "REPLACE-chat_template_sha256"
|
|
},
|
|
"cache_policy": "REPLACE-reset-before-round; warm-with-recorded-workload",
|
|
"reasoning_policy": "REPLACE-explicit-thinking-setting",
|
|
"speculation_policy": "REPLACE-disabled-for-initial-baseline",
|
|
"quality_protocol": "REPLACE-quality-dataset-revision-metric-and-acceptance-threshold",
|
|
"matrix": {
|
|
"concurrency": [
|
|
1,
|
|
4,
|
|
8,
|
|
16
|
|
],
|
|
"rounds": 3,
|
|
"repeats": 8,
|
|
"warmup_requests": 2,
|
|
"max_tokens": 128,
|
|
"temperature": 0,
|
|
"timeout": 120,
|
|
"slo_first_output": 5,
|
|
"slo_duration": 120
|
|
},
|
|
"engines": {
|
|
"colibri": {
|
|
"base_url": "http://127.0.0.1:8000/v1",
|
|
"served_model": "REPLACE-model-alias",
|
|
"revision": "REPLACE-commit-or-image-digest",
|
|
"launch_command": "REPLACE-exact-command-with-secret-environment-references",
|
|
"artifact_sha256": "REPLACE-weight-manifest-sha256",
|
|
"weight_format": "REPLACE-format",
|
|
"quantization": "REPLACE-weight-and-KV-precision",
|
|
"api_key_env": "OPENAI_API_KEY"
|
|
},
|
|
"sglang": {
|
|
"base_url": "http://127.0.0.1:8000/v1",
|
|
"served_model": "REPLACE-model-alias",
|
|
"revision": "REPLACE-commit-or-image-digest",
|
|
"launch_command": "REPLACE-exact-command-with-secret-environment-references",
|
|
"artifact_sha256": "REPLACE-weight-manifest-sha256",
|
|
"weight_format": "REPLACE-format",
|
|
"quantization": "REPLACE-weight-and-KV-precision",
|
|
"api_key_env": "OPENAI_API_KEY"
|
|
},
|
|
"vllm": {
|
|
"base_url": "http://127.0.0.1:8000/v1",
|
|
"served_model": "REPLACE-model-alias",
|
|
"revision": "REPLACE-commit-or-image-digest",
|
|
"launch_command": "REPLACE-exact-command-with-secret-environment-references",
|
|
"artifact_sha256": "REPLACE-weight-manifest-sha256",
|
|
"weight_format": "REPLACE-format",
|
|
"quantization": "REPLACE-weight-and-KV-precision",
|
|
"api_key_env": "OPENAI_API_KEY"
|
|
}
|
|
}
|
|
}
|