coremlit 0.1.0

Safe, synchronous CoreML runtime for macOS (CPU/GPU/Neural Engine) with opt-in on-device multimodal pipelines: speech (Whisper STT, forced alignment, speaker diarization, Silero VAD), AudioSet sound-event tagging, and audio/text/image embeddings (CLAP, granite, SigLIP)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53

//! Embedding producers.
//!
//! A namespace of feature-gated submodules mirroring [`crate::audio`], each a
//! self-contained CoreML embedding pipeline built over the always-compiled
//! runtime core (`Model` / `MultiArray` / `Features`). Enabling a module's
//! feature is the only way it compiles, and `default = []` pulls none of them.
//!
//! - `clap` — CLAP-HTSAT dual-tower audio+text embeddings, a former standalone
//!   kit crate collapsed here per the mono-crate restructure (feature `clap`;
//!   the textclap parity oracle lives in the sibling `coremlit-parity`
//!   package's `clap-oracle` feature).
//! - `granite` — general text sentence-embeddings on CoreML (the embedkit
//!   phase), first model `granite-embedding-97m-multilingual-r2` (feature
//!   `granite`; committed transformers-fp32 goldens as the parity oracle — NO
//!   ort). The T2/T3 core is landed; long-input windowing (T4) is a later phase.
//!   `gemma` was DROPPED per the embedkit design's Amendment 3.
//! - `siglip` — SigLIP 2 (`siglip2-base-patch16-naflex`) dual-tower image+text
//!   embeddings on CoreML, a shared 768-dim joint space with cross-modal `rank`
//!   (feature `siglip`; committed transformers-fp32 goldens as the parity
//!   oracle — NO ort). NaFlex resizes natively to a fixed patch budget, so no
//!   windowing is used. The hermetic preprocessing + embedding core is landed;
//!   the model-gated parity gates await the staged conversion.
//!
//! - `face` — face embedding for the video identity path (feature `face`): the
//!   ArcFace 5-point similarity alignment as an explicit, golden-tested step,
//!   plus a manifest-driven CoreML embedder over a CALLER-supplied artifact.
//!   Its one staged artifact — InsightFace's `w600k_r50`, research-only at both
//!   the weights and the corpus layer — is a manifest module of its own behind
//!   `commercial-face-arcface`, a feature never in `default`. The alignment and
//!   the embedder are encumbered by nothing and stay ungated; see the module's
//!   own doc for why only the artifact is behind the gate.
//!
//! See the crate README's layering map for module authority. The `video/`
//! sibling namespace is likewise reserved, but is not created until a video kit
//! exists (README note).

#[cfg(feature = "clap")]
pub mod clap;

#[cfg(feature = "granite")]
pub mod granite;

#[cfg(feature = "siglip")]
pub mod siglip;

#[cfg(feature = "face")]
pub mod face;

/// The structural post-processor check the text doors share. Crate-private —
/// its one public type is re-exported by each consuming door's `error` module,
/// where a caller matching on that door's `Error` already looks.
#[cfg(any(feature = "clap", feature = "siglip"))]
pub(crate) mod tokenizer_guard;