Documentation
Settings and simulators
Settings holds the choices that apply across every tool: the simulator, the inference backend, the backend version, and the defaults a new view starts from.
Simulator
A simulator is the implementation that produces an estimate. It is not the
serving engine: vllm, sglang, and tensorrt-llm are backends that a
simulator models, and several simulators can model the same backend.
The simulator list comes from GET /simulators, not from this app. Each entry
has an id, a label, a version, and an optional revision. The entry marked
default is used when no explicit choice has been saved.
Switching simulators changes which models, GPU systems, backends, fabrics, and objectives are offered. The app reads those capabilities from the catalog and never assumes that two simulators support the same set. If a requested combination is unsupported, the API returns an explicit error; the app does not substitute another simulator or invent a value.
Backend and version
A backend version pin means something different in each simulator, so versions are stored per simulator. When you switch simulators, the app loads the version saved for that simulator rather than carrying the previous one across.
Some simulators publish a system_defaults entry per backend version. Leaving
the version empty lets the service resolve that default. Pinning a version is
how you reproduce a specific engine release.
Tools do not offer their own backend or version picker. They read the pair from Settings and report it read-only next to the values they do own, with a link back to this section. That keeps one effective pair across the app instead of per-page copies that could disagree — and it avoids saving an AISimulate version pin under another simulator, where the key would mean something else.
A shared link still records the backend pair that produced its numbers. Opening one applies that pair for the session only, exactly like a shared simulator, so the recipient sees the sender's request without having their saved settings rewritten.
The simulator badge
Every result carries a badge naming the simulator that produced it. The badge is read from the response, not inferred from your settings, because the service is the only place that decides which simulator ran. Treat it as provenance: two results with the same inputs and different badges are not expected to match.
Sharing a view
A shared link records the simulator, backend, and backend version used to build it. Opening a link never changes your saved settings:
- If the link's simulator matches yours, its inputs are applied.
- If it differs, the page shows a banner instead of switching. The banner names both simulators and offers a one-click Use … for this view action, which applies the link's provenance for the current session only.
- If the link's simulator is unavailable in this deployment, the existing unavailable notice is shown. There is no fallback.
What a link carries
A link records the view, not the page. Only values that differ from the page's starting state are written, so a link stays short and readable, and a field it leaves out keeps the value already in the recipient's form. Opening a partial link therefore never resets a control the sender did not touch.
Two cases where an omitted field would be ambiguous are carried explicitly:
- A disaggregated view writes each pool's parallelism separately — prefill and decode tensor parallel, pipeline parallel, worker count, and batch size — because the pools can differ in every one of them. An aggregated view writes none of them, so a link never describes controls the recipient cannot see.
- None in the recommend page's fabric list is a choice, not a blank. It is
written as
fabric=none, which asks for aggregate candidates only. Omittingfabricmeans something else: the link says nothing about the fabric, so the recipient keeps the fabric already selected in their form. On a fresh page that is the simulator's default.
Benchmark comparison
This is the one page whose shared link does not use a catalog system id. The
accelerator value is the label the benchmark dataset publishes (for example
H200, not h200_sxm), because the dataset is the source of that selection and
a measured row can exist for a GPU the catalog does not list. The page matches
the label to a catalog entry when it can and otherwise keeps the dataset label,
so the view opens on the same measured run the sender saw.
Related
- User scenarios: which tool to use for which ask
- REST API: catalog endpoints and response provenance