mirror of
https://github.com/JustVugg/colibri.git
synced 2026-10-02 02:54:37 +08:00
JustVugg
0ed7ab527d
What it is, why the entropy is the part you cannot get from a generated answer, and the fact that it is server-only -- there is no CLI and no one-shot form, because a one-shot process would throw away the snapshot the mode exists for. Covers the HTTP endpoint field by field, the reply, the two protocol keys for anyone speaking the serve protocol directly, the three steps to score by hand, and the two mistakes that are easy to make: putting the option list in the prompt (48 tokens of 123 on a real case, and it biases the scoring), and summing log probabilities instead of averaging them, which penalises multi-token options -- measured, the sum picked a different answer than the same model's own greedy decoding. Measured costs included with the caveat that they come from a disk-streaming engine, so the ratios are the point and not the absolute seconds. Indexed under Serve models, next to the protocol reference it points at.
75 lines
1.8 KiB
JSON
75 lines
1.8 KiB
JSON
{
|
|
"$schema": "https://mintlify.com/docs.json",
|
|
"theme": "mint",
|
|
"name": "colibrì",
|
|
"description": "Run frontier mixture-of-experts models on consumer and heterogeneous hardware.",
|
|
"colors": {
|
|
"primary": "#00afaf",
|
|
"light": "#5fd7d7",
|
|
"dark": "#008b8b"
|
|
},
|
|
"navigation": {
|
|
"tabs": [
|
|
{
|
|
"tab": "Documentation",
|
|
"groups": [
|
|
{
|
|
"group": "Get started",
|
|
"pages": ["index", "quickstart"]
|
|
},
|
|
{
|
|
"group": "Serve models",
|
|
"pages": ["api-reference/overview", "api", "brio", "serve_protocol"]
|
|
},
|
|
{
|
|
"group": "Model engines",
|
|
"pages": ["deepseek-v4", "kimi_k3", "qwen38", "qwen36", "inkling"]
|
|
},
|
|
{
|
|
"group": "Configure and tune",
|
|
"pages": ["SETTINGS", "ENVIRONMENT", "tuning", "CACHE_ROUTE"]
|
|
},
|
|
{
|
|
"group": "Hardware backends",
|
|
"pages": ["cuda", "metal", "vulkan", "windows"]
|
|
},
|
|
{
|
|
"group": "Contribute",
|
|
"pages": ["benchmarks", "MAINTAINING-DOCS"]
|
|
}
|
|
]
|
|
},
|
|
{
|
|
"tab": "API Reference",
|
|
"openapi": {
|
|
"source": "./api-reference/openapi.json",
|
|
"directory": "api-reference/endpoints"
|
|
}
|
|
}
|
|
]
|
|
},
|
|
"navbar": {
|
|
"links": [
|
|
{
|
|
"label": "Website",
|
|
"href": "https://justvugg.github.io/colibri"
|
|
},
|
|
{
|
|
"label": "GitHub",
|
|
"href": "https://github.com/JustVugg/colibri",
|
|
"icon": "github"
|
|
}
|
|
],
|
|
"primary": {
|
|
"type": "button",
|
|
"label": "Quickstart",
|
|
"href": "/quickstart"
|
|
}
|
|
},
|
|
"footer": {
|
|
"socials": {
|
|
"github": "https://github.com/JustVugg/colibri"
|
|
}
|
|
}
|
|
}
|