Dumpster Dyva

Host population

Source breakdown

Service breakdown

Routing configurationneeds save

The fine-tunable hyperparameters for how a request finds a host — the diagram tracks them live as you adjust.

WorkersParallel hosts tried per request (fan-out). Hedge delay (s)How long the last-good host gets to respond before the rest of the pool is raced alongside it. Whoever answers first wins. 0 races everything at once. Keeps a dead-but-connected favourite from stalling the whole request. Warm connect wait (s)A phase inside the hedge lead: if the warm host hasn’t opened a TCP socket in this long, race the pool early (the warm host keeps running and can still win). Set it equal to the hedge delay to disable — then the warm host gets the full lead. Transport only; never cuts a connected host short. Minimum timeout (s)Floor on the per-host first-token wait. A small prompt uses exactly this; a big prompt scales up from here toward Maximum timeout. Also the base clock for the worker cap in the diagram. Maximum timeout (s)Ceiling on the per-host first-token wait — the longest we'll ever wait on a monster prompt. The budget scales from Minimum up to this with prompt size (at Input rate), and the SAME extra seconds are added to Fanout, so the worker/connection maximum is unchanged — a big prompt just waits longer on the same number of hosts. Set equal to Minimum timeout to disable scaling. Max run time (s)Hard ceiling on a single response's TOTAL run time, measured from when the worker starts. The timeouts above only cover the wait for the first token and the gap between tokens — a host that keeps trickling tokens (or keep-alives) can otherwise stream for hours. 0 disables. Default 1800 (30 min); can be set as high as 30 days. Fail over on max run timeWhat happens when Max run time is hit. Off (default): end the stream cleanly with what was produced — no failover, and the host is NOT penalised (a duration guard, not a quality verdict). On: treat it as a mid-stream cutoff and fail over to another host, carrying the partial answer forward — useful with a short cap (e.g. 300s) to re-race a slow host. Race watchdog (s)If a request's host search runs longer than this without finding a host, log its internal state (candidates tried, fan-out liveness, permits) every N seconds — so a wedged search surfaces in the log instead of hanging silently. Quiet for healthy requests, which resolve well under it. 0 = off. Default 30. Fanout (s)How often another wave of hosts is opened, independent of Timeout. Lower = widen the search faster; each attempt still gets its full Timeout, so a cold host isn't dropped to make room. Input rate (tok/s)Assumed prefill speed — how fast a capable host chews through the prompt before the first token. Turns prompt size into expected prefill time, so the race clocks stretch for a big prompt and it isn't dropped for taking the prefill we asked of it. Conservative on purpose (real hardware does several times this). Small prompts are unaffected.

Behaviorneeds save

Explore modeDiscover more of the unknown network: route hosts we haven't visited (unknown state) to the front of the line, ahead of known-good ones. Deliberately slower — a request may probe cold hosts first — but it maps out the pool. Off by default. Caution: the unvisited frontier is where the canary honeypots live, and poking a burst of fresh IPs is exactly what trips Cloudflare/WAF reputation — so this raises the chance of a soft IP ban on your egress. Use it for deliberate survey passes, not routine routing.

Thinking token keyFor OpenAI /v1 requests, which key(s) carry the model's reasoning. Every ticked key is emitted with the same text. OpenAI never standardized this field, so clones diverged — tick what your client reads. Emitting more than one can break a strict client that rejects unknown fields, so the default is the single OpenAI key. (The ollama /api/chat path always uses thinking and is unaffected.)

OpenAI — reasoning_content OpenRouter — reasoning custom

CloudskipperNever route to — or even connect to — hosts on the ticked clouds. The soft-ban tripwire can fire on the TCP connection itself, so a skipped cloud's hosts are dropped from every request before any connection is made. Clouds are identified by enrichment; the list below is the ones most represented in the current survey. Applies to every source and is always in effect, independent of Explore mode.

Presentationneeds save

Minimum model countHide models served by fewer than this many hosts from the /api/tags and /v1/models listings that third-party apps read. 0 = show all. (The dashboard always shows everything.) Base pathServe dyva under a secret URL path instead of the root. A path like /some-path moves the dashboard to /some-path/ and the API to /some-path/v1/models — deterrence by obscurity: a scanner hitting the default paths (/v1/models, /api/tags) finds nothing. Blank = served at the root. Takes effect on restart (like the listen port); the --base flag overrides this. GossiperShare high-confidence signals — starting with smoke-test results (real endpoint vs honeypot) — with peer dyva instances, scoped by source, so you skip re-probing a host a peer just characterized. The signal set expands conservatively over time. A peer's honeypot flag is trusted: the host is skipped without contact (re-verifying it would mean sending the trap a request — the exact thing the flag spares you). Privacy-bounded: a sync only reaches peers who already hold the same source URL. On by default; untick to opt out entirely. Gossiper endpointThe gossiper server to sync with. Defaults to the public reference instance; point it at your own, or clear it to run without one. Keep the trailing slash (a POST to the slashless form gets redirected and dropped). Model list overrideAdvertise a fixed list of model names from /api/tags and /v1/models instead of the discovered catalog, for tools that make you pick one specific model id. Every dyva model name is a routing pattern, so one entry like qwen3 covers every qwen3:* variant on every host. Tick Enabled and put one name per line in the box below (blank lines and #comments are ignored). Untick to go back to the real catalog — the list stays in the box so you can flip it back on. Names no host currently serves stay listed (count 0) so the menu can't shift under a tool that pinned one. Minimum model count doesn't apply to an explicit list.

Congenialityneeds save

Admin passwordProtect this Settings tab and its sources. When set, opening Settings requires this password — everything else (chat, models, the dashboard) stays public. Type a new value to set or change it; leave blank to keep the current one. Forgot it? Clear admin_pw in settings.json and restart. Localhost onlyRestrict inference endpoints to localhost only (a remote host that reaches this port can't chat or generate). Equivalent to the ./dyva -l flag; saved here it persists across restarts.

Additional sources

Each source fetches a JSON list of host rows from its url and maps fields onto host entries. Configured sources are tried before the built-ins, so they win on duplicate hosts.

mapping — each target field is one of:

{"field": "x"} — copy row["x"] from the source row
{"value": 0} — a constant

Map a server (or url) so the host can be keyed; models should resolve to a list of strings. CSV / transform-heavy sources aren't supported here.

Example

[
  {
    "name": "my-list",
    "url": "https://example.com/hosts.json",
    "mapping": {
      "server":  {"field": "url"},
      "models":  {"field": "models"},
      "service": {"value": "ollama"}
    }
  }
]

Extra JSON host sources, fetched before the built-ins (highest priority). A JSON array of {name, url, mapping}.

Server

Shut down dyvaStop the server process (like Ctrl+C in the terminal). You'll need host/terminal access to start it again. Admin-gated: requires the admin password when one is set.

Dyva

Models

Host

Host

Hosts

Host population

Map & compare syntax

Source breakdown

Service breakdown

Dyva

Models

Host

Host×

Hosts

Host population

Map & compare syntax

Source breakdown

Service breakdown

Host