mstar.engine.resources.dsa.config

What a model declares for DeepSeek sparse attention (DSA): its config, spec and step.

Kept free of the manager so a submodule can declare a step without importing it.

Classes

DSAConfig(kv_cache, topk)
DSASpec(resource_key, nodes, config)
DSAStep([segments])

class mstar.engine.resources.dsa.config.DSAConfig(kv_cache: str, topk: int)

Bases: object

Parameters:

• kv_cache (str)

• topk (int)

kv_cache: str

topk: int

class mstar.engine.resources.dsa.config.DSASpec(resource_key: str, nodes: set[str], config: mstar.engine.resources.dsa.config.DSAConfig)

Bases: NodeResourceSpec

Parameters:

• resource_key (str)

• nodes (set[str])

• config (DSAConfig)

config: DSAConfig

depends_on()

Keys whose specs this one builds against; the engine resolves them into EngineResourceInfo.dependencies.

Return type:

set[str]

property resource_class: type[Resource]

What builds this spec. Imported inside the property, so declaring a resource stays free of the manager and its kernels.

The builder, not necessarily the class built: an attention spec names AttentionManager, whose build picks a backend subclass.

class mstar.engine.resources.dsa.config.DSAStep(segments: tuple[mstar.engine.resources.step.Segment, ...] | None = None)

Bases: ResourceStep

Parameters:

segments (tuple[Segment, ...] | None)