Crate cortiq_engine 

Cortiq inference engine — sparse forward pass, attention, tokenization, sampling.

Re-exports

pub use nystrom::NystromState;

pub use pipeline::GenerateResult;

pub use pipeline::Pipeline;

pub use pipeline::TokenCallback;

pub use pipeline::TokenTrace;

pub use runtime::CortiqRuntime;

pub use sampler::SamplerConfig;

Modules

attention

Attention forward pass — GQA with RoPE, head masking, head-major KV.

fcd

FCD polish trainer — native Rust quality-polish for O(1)-converted models (docs/RUST_FCD.md). Removes the last Python dependency from the cortiq convert --o1 pipeline.

fcd_ops

FCD polish — hand-rolled forward/backward operators.

gptq_capture

Calibration-time Hessian capture for GPTQ / the holographic transfer.

gpu

Facade for GPU backends: a single call entry point for qtensor/pipeline/ linear_core. Job types and the threshold are canonical HERE; behind the facade dispatch goes to a platform backend:

inference

CPU forward-pass primitives: RMS-norm, SiLU, sparse SwiGLU FFN.

kv_cache

KV cache — per-layer, head-major storage.

linear_core

Linear-attention cores, selected by arch.linear_core.kind (descriptor-driven operators — Patent 15 claim 8).

loader

Weight loader: CMF tensor directory → Pipeline.

nystrom

Nyström (landmark) attention kernel — streaming per-GQA-group runtime for long-context attn_type: nystrom layers.

pipeline

Full inference pipeline: tokenize → embed → layers → lm_head → sample → decode.

pool

Persistent worker pool for row-parallel matvecs.

qtensor

QTensor — weight tensor with pluggable storage.

router

Recon-argmin skill routing (spec §9, P1 signal-consistency): the container’s selection descriptors define per-skill affine subspaces over φ(x); the winner is the skill that reconstructs φ best. No trained gate — routing is a property of the skills themselves.

runtime

Runtime state management — model, active task, metrics.

sampler

Token sampling — temperature, top-p, top-k, min-p, repetition penalty.

skillbake

Native DTG-MA skill bake (Patent 2) — no Python, no torch.

swarm

Dynamic per-token skill routing with hysteresis (spec §9 made runtime; VMF-2026 experiment №2).

tokenizer

Byte-level BPE tokenizer — HF tokenizer.json parity.