Module cache

Expand description

Aggregate ownership and admission for backend-managed live caches. Portable cache/state policy and prompt-cache schemas.

Runtime ownership, admission, storage transitions, and backing-store scheduling live in eredu-runtime so this crate remains declarative.

Structs§

CacheBlockId: Stable identity for one immutable sealed cache block.
CacheRankIdentity: Optional rank identity included in a stable cache block identifier.
PromptCacheBlock: One cache block catalog entry in a prompt-cache manifest.
PromptCacheDescriptor: Caller-supplied identity and geometry for a reusable prefix cache.
PromptCacheManifest: Versioned metadata inspectable without loading backend arrays.
PromptCacheModelIdentity: Cache-relevant structure derived from a prepared model.
PromptCacheOptions: Explicit publication behavior for a reusable prefix cache.
PromptCacheStateSegment: One named contiguous state range in a portable prompt-cache identity.
PromptCacheStateTensor: One independently validated fixed-size state tensor catalog entry.
PromptCacheTopology: Rank-local topology recorded in a prompt-cache manifest.
StateComponentPolicy: Symbolic geometry and persistence behavior of one named state component.
StateTensorPolicy: Exact semantic role, symbolic shape, and dtype contract for a state tensor.

CachePolicyError: Invalid cache geometry or state-residency contract.
CacheRepresentation: Representation stored atomically in one cache block.
CacheTier: Logical location of a sealed cache block.
LayerCachePolicy: Exact state kind, attention policy, and tensor geometry for one decoder layer.
MutableStateResidency: Residency behaviors valid for mutable fixed-state tensors.
PoolingStateComponent: Semantic component of one append-only pooling stream.
PromptCacheError: Invalid reusable prompt-cache identity, schema, or catalog.
StateComponentRole: Stable semantic identity for one independently managed runtime component.
StateResidencyClass: Runtime ownership behavior for live model state.
StateTensorDimension: One dimension in a persisted fixed-state tensor.
StateTensorDtype: Accepted dtype family for one fixed-state tensor.
StateTensorOwner: Owner of one persisted non-attention state tensor.
StateTensorPresence: Condition under which a state tensor must be materialized.
StateTensorRole: Semantic role of one non-attention cache tensor.

derive_prompt_cache_architecture_fingerprint: Derives a stable cache architecture fingerprint from ordered semantic fields.
prompt_cache_token_fingerprint: Hashes exact prefix token IDs as little-endian u32 values.
validate_prompt_cache_model_identity: Verifies that a caller descriptor was derived from the prepared model.