Expand description
KV-cache compression contracts and CPU reference paths.
The kv feature is experimental and default-off. It provides typed
contracts, role-aware policy decisions, and CPU reference encode/decode
helpers for KV-cache pages. It is not a production serving backend.
Re-exports§
pub use attention_ref::compare_attention_fixture;pub use attention_ref::reference_attention_logits;pub use attention_ref::reference_value_aggregation;pub use block::KvBlockEncodingV1;pub use block::KvEncodedBlockV1;pub use calibration::calibrate_kv_tensor;pub use calibration::KvCalibrationSummaryV1;pub use codec::decode_kv_pages;pub use codec::encode_kv_tensor;pub use codec::KvDecodedTensorV1;pub use codec::KvEncodedTensorV1;pub use layout::KvCacheLayoutV1;pub use layout::KvLayoutOrder;pub use layout::KvPageGeometryV1;pub use page::KvEncodedPageV1;pub use policy::decide_kv_compression;pub use policy::KvCompressionDecisionV1;pub use policy::KvCompressionPolicyV1;pub use policy::KvCompressionStrategyV1;pub use policy::KvDecisionActionV1;pub use policy::KvDecisionReasonV1;pub use profile::KvAxisPolicyV1;pub use profile::KvCompressionProfileV1;pub use profile::KvFallbackModeV1;pub use profile::KvFallbackPolicyV1;pub use profile::KvProtectedPolicyV1;pub use profile::KvQualityBudgetV1;pub use quality::KvAttentionQualityReportV1;pub use quality::KvLayerHeadQualityV1;pub use receipt::kv_tensor_digest;pub use receipt::KvCompressionReceiptV1;pub use receipt::KvDecodeReceiptV1;pub use receipt::KvEvalReceiptV1;pub use receipt::KvOperationKindV1;pub use shape::KvAttentionKind;pub use shape::KvDType;pub use shape::KvRole;pub use shape::KvRopeState;pub use shape::KvTensorShapeV1;