Crate irithyll_core 

Core types and inference engine for irithyll streaming ML.

irithyll-core is the #![no_std] foundation shared by the full irithyll crate and embedded targets. It provides loss functions, the streaming Observation trait, a compact 12-byte packed node format, and branch-free ensemble traversal for deploying trained models on bare metal (Cortex-M0+, 32 KB flash).

The crate has a hard dependency boundary: only libm is mandatory. All dynamic-allocation paths (histogram binning, tree construction, drift detectors, neural architectures) gate on the alloc feature.

Feature Flags

Feature	Description
alloc	Enables dynamic-allocation types: trees, ensembles, drift, neural
std	Enables alloc plus standard I/O (required for full irithyll crate)
serde	Derives Serialize/Deserialize on config and snapshot types
kmeans-binning	K-means histogram binning strategy (requires alloc)
parallel	Rayon-parallel tree training (requires alloc)
simd	AVX2 histogram accumulation acceleration (requires std)
simd-avx2	Explicit AVX2 SIMD intrinsics (requires std)
embedded-bench	Cortex-M semihosting bench helpers
iai-bench	iai-callgrind regression bench source

Embedded Inference

Train a model with the full irithyll crate, then export:

use irithyll_core::{EnsembleView, FormatError};

// packed_bytes is a &[u8] from irithyll::export_embedded()
let packed_bytes: &[u8] = &[];
let view = EnsembleView::from_bytes(packed_bytes)?;
let prediction = view.predict(&[1.0f32, 2.0, 3.0]);

The packed format stores each node in 12 bytes (5 nodes per cache line). Traversal is branch-free: child selection uses cmov/csel-equivalent conditionals that avoid pipeline stalls on Cortex-M.

Re-exports

pub use error::FormatError;

pub use packed::EnsembleHeader;

pub use packed::PackedNode;

pub use packed::TreeEntry;

pub use packed_i16::PackedNodeI16;

pub use packed_i16::QuantizedEnsembleHeader;

pub use view::EnsembleView;

pub use view_i16::QuantizedEnsembleView;

pub use loss::Loss;

pub use loss::LossType;

pub use sample::Sample;alloc

pub use sample::Observation;

pub use sample::SampleRef;

pub use drift::DriftSignal;

pub use drift::DriftDetector;alloc

pub use drift::DriftDetectorState;alloc

pub use error::ConfigError;alloc

pub use error::IrithyllError;alloc

pub use error::Result;alloc

Modules

attentionalloc

Unified streaming linear attention engine.

continualalloc

Continual learning strategies for streaming neural models.

drift

Concept drift detection algorithms.

ensemblealloc

SGBT ensemble orchestrator – the core boosting loop.

error

Error types for irithyll-core.

featurealloc

Feature type declarations for streaming tree construction.

histogramalloc

Histogram-based feature binning for streaming tree construction.

learneralloc

Unified streaming learner trait for polymorphic model composition.

loss

Loss functions for gradient boosting.

lstmalloc

sLSTM (stabilized LSTM) cell with exponential gating and log-domain stabilization.

math

Platform-agnostic f64 math operations.

mgradealloc

mGRADE (Minimal Recurrent Gating with Delay Convolutions) core cells.

packed

12-byte packed node format and ensemble binary layout.

packed_i16

8-byte quantized packed node format for integer-only inference.

quantize

f64 → f32 quantization utilities for packed export.

reservoiralloc

Reservoir computing primitives for streaming temporal models.

rng

Deterministic xorshift64 PRNG for reproducible initialization.

sample

Core observation trait and zero-copy sample types.

simd

SIMD-accelerated math primitives for neural forward passes.

snnalloc

Spiking Neural Networks with online e-prop learning.

ssmalloc

State Space Models for streaming temporal feature extraction.

streaming_primitives

Shared streaming primitives for irithyll-core model building blocks.

traverse

Branch-free tree traversal for packed nodes.

traverse_i16

Branch-free tree traversal for quantized i16 packed nodes.

treealloc

Streaming decision trees with Hoeffding-bound split decisions.

turbo_quantalloc

TurboQuant multi-mode weight quantization with randomized Hadamard rotation.

view

Zero-copy, zero-alloc inference view over a packed ensemble binary.

view_i16

Zero-copy, zero-alloc inference view over a quantized (int16) ensemble binary.