33 lines
1.2 KiB
TOML
33 lines
1.2 KiB
TOML
# <data root>/config/apps/judge.toml on nh3-svos — app-judge 0.0.1 (U19), from apps/judge/examples/judge.toml at v0.0.16; + explain_seconds for 0.0.2 (v0.0.24).
|
|
# Prime 2026-10-18: drafting on gen-large (local, free); tasks may run on the six local chat models (the hermes runtime's
|
|
# `models` in instance.toml). The gateway key (vault nh3-svos/svos-judge-gateway-key, LiteLLM alias svos-app-judge) is
|
|
# scoped to exactly those six, so the /tasks model select offers exactly them.
|
|
|
|
[gateway]
|
|
base = "http://10.250.50.70:4000/v1"
|
|
credential_env = "SVOS_JUDGE_GATEWAY_KEY" # /etc/svos/judge.env (root 0600); a project-scoped key, never the shared one
|
|
|
|
[propose]
|
|
model = "gen-large" # hosted_vllm qwen3.8-flash on fv-ml1 :8022 — free; step up only on Prime's call
|
|
max_tokens = 4000
|
|
|
|
[admit]
|
|
scan_app = "app-terminal"
|
|
|
|
[[routes]]
|
|
role = "*"
|
|
runtime = "hermes"
|
|
model = "gen-large"
|
|
|
|
[limits]
|
|
models_seconds = 10
|
|
propose_seconds = 120
|
|
admit_seconds = 8
|
|
route_seconds = 5
|
|
catalog_seconds = 30
|
|
concurrency = 8
|
|
reply_bytes = 1048576
|
|
brief_bytes = 262144
|
|
health_seconds = 5
|
|
explain_seconds = 60 # U22 (judge 0.0.2, v0.0.24): one deadline per explain call, the model call included; 5..60, the page waits 65 s
|