Expand description
cortiq-core — CMF v2 container: types, tensor directory, masks, quant.
Format specification: docs/CMF_V2_SPEC.md.
Re-exports§
pub use format::AppendReport;pub use format::CMF_MAGIC;pub use format::CMF_VERSION;pub use format::CmfError;pub use format::CmfHeader;pub use format::CmfModel;pub use format::PendingAppend;pub use format::RoutingCalibration;pub use format::SelectionDescriptor;pub use format::SkillRecord;pub use format::SparseIndexEntry;pub use format::TensorEntry;pub use format::TensorSpec;pub use format::TensorSpecRef;pub use format::build_sparse_index;pub use format::dspark_sidecar_path;pub use format::mtp_sidecar_path;pub use hash::hash64;pub use knowledge::ExecHashes;pub use knowledge::ExpertAppend;pub use knowledge::ExpertTensorSpec;pub use knowledge::GenomeInfo;pub use knowledge::GenomeParent;pub use knowledge::LineageEvent;pub use knowledge::LookupInfo;pub use knowledge::PhiSpec;pub use knowledge::RouterPolicy;pub use knowledge::Segment;pub use knowledge::SkillBound;pub use knowledge::SkillOverride;pub use knowledge::StateEffect;pub use knowledge::expert_append_base;pub use knowledge::expert_append_layout;pub use knowledge::expert_append_state_effect;pub use knowledge::ffn_replace_state_effect;pub use knowledge::genome_moe_experts;pub use knowledge::key_hash;pub use knowledge::lookup_state_effect;pub use knowledge::lookup_tensor_name;pub use knowledge::lookup_tensors;pub use knowledge::moe_layers;pub use knowledge::normalize_key;pub use knowledge::normalized_key_hash;pub use knowledge::trunk_expert_rank;pub use knowledge::trunk_hash;pub use knowledge::trunk_hash_exec;pub use knowledge::validate_expert_append_values;pub use knowledge::validate_lookup_values;pub use mask::MaskCatalog;pub use mask::MaskDiff;pub use mask::MaskPriority;pub use mask::Quality;pub use mask::TaskMask;pub use types::AnchorCoreConfig;pub use types::BlockDiffusionConfig;pub use types::ExecutionMode;pub use types::FarFieldConfig;pub use types::G3nConfig;pub use types::LayerStats;pub use types::LayerType;pub use types::LinearCoreConfig;pub use types::MlaConfig;pub use types::ModelArch;pub use types::MoeConfig;pub use types::MtpConfig;pub use types::NormStyle;pub use types::PerformanceMetrics;pub use types::PrismHadamardConfig;pub use types::QuantType;pub use types::Qwen4ExpConfig;pub use types::SimdType;pub use types::TensorDtype;
Modules§
- blr
- Binary low-rank weights (
blr, dtype 19) — docs/BLR_NVG.md §2. - format
- CMF v2 binary container — envelope, tensor directory, mmap access.
- hadamard
- Signed, normalized Sylvester Walsh-Hadamard transforms used by the Prism/Bonsai ternary extension.
- hash
- Per-tensor 64-bit hash — corruption detection for mmap’d weights and backbone-tensor dedup between skill files.
- knowledge
- Format v2 “knowledge without forgetting” (spec §9.2–§9.5): a FROZEN genome, skill records v2 that own their tensors over it, and the backbone-gated router policy.
- mask
- Task mask management — per-task neuron/head/layer masks.
- mxfp4
mxfp4(dtype 21): OCP Microscaling FP4 — E2M1 elements with one E8M0 scale per 32 consecutive elements of a row.- quant
- Canonical quantization layouts and scalar dequantization.
- types
- Core types for CMF format.