Scenario

The problem a run is given: a model, the cluster it may use, the traffic it serves, the coefficient sets that price it, and the engine release it runs. Everything in it is fixed for the run. It names catalog and registry entries rather than containing them.

kind: Scenario
name: deepseek-v3-on-h200
model: deepseek-v3
coefficients:
  - cost-model-primitives
  - cost-model-collectives
  - cost-model-host-overheads
  - cost-model-attention
  - cost-model-recurrent
engine_version: "0.29.0"
workload:
  shape: summarization
cluster:
  hardware: h200
  fabric: ib-400g
  nodes: 4
  gpus_per_node: 8
  storage:
    - cpu_dram

Keys

Key	Type	Required	Meaning
kind	string	yes	Always Scenario.
name	string	yes	The scenario's identity.
model	string	yes	A model directory under models/ in the catalog.
coefficients	list of strings	yes	Coefficient sets in the registry, named by file, at least one. How a cost model combines them is under Selecting an entry.
engine_version	string	yes	The exact release string of the engine the run describes, such as 0.29.0. Selects the rules pack.
workload	mapping	no	The traffic, as a shape, a trace or a replay. See Workload. Omitted for a question about capacity alone.
routing	mapping	no	Pins expert-parallel routing imbalance for the run, by a routing-profile set in the registry or an inline value. See Scenario pin. Omitted, the workload and the registry's default sets decide, else evenly loaded.
acceptance	mapping	no	Pins speculative acceptance for the run, by a profile set in the registry or an inline value. See Scenario pin. Omitted, the workload and the registry decide.
cluster	mapping	yes	The hardware available, below.

cluster

Key	Type	Required	Meaning
hardware	string	yes	A chip under hardware/ in the catalog. Every node carries it.
fabric	string	if nodes > 1	A fabric under networks/ in the catalog. Cross-node cost comes from it.
nodes	integer	yes	Node count, at least 1.
gpus_per_node	integer	yes	GPUs per node, at least 1.
gpus_per_rack	integer	no	GPUs per rack, where a rack boundary is a different link from a node boundary. A multiple of gpus_per_node.
storage	list of strings	no	Storage classes from the catalog's devices/storage.yaml that the cluster provides. If any are listed, a deployment's offload tiers must use them.

Checks

A scenario is invalid if

• kind is not Scenario, or name, model or engine_version is empty;
• coefficients is empty;
• cluster.hardware is empty, or nodes or gpus_per_node is less than 1;
• gpus_per_rack is set and is not a multiple of gpus_per_node;
• the cluster has more than one node and names no fabric;
• a storage entry is empty or appears twice;
• workload is present and is not a valid binding.
• acceptance is present and is not a valid pin.

Whether the names resolve is checked by ValidateAgainstCatalog, not by Validate; see Validation.

Source: spec/scenario.