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
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232

use super::*;

#[test]
fn geometry_constants_match_fluidaudio() {
  // FluidAudio `VadManager`: chunkSize 4096, contextSize 64, modelInputSize
  // 4160, stateSize 128 (`VadManager.swift:22-25`).
  assert_eq!(CHUNK_SAMPLES, 4096);
  assert_eq!(CONTEXT_SAMPLES, 64);
  assert_eq!(MODEL_INPUT_SAMPLES, 4160);
  assert_eq!(MODEL_INPUT_SAMPLES, CONTEXT_SAMPLES + CHUNK_SAMPLES);
  assert_eq!(STATE_SIZE, 128);
}

// ── prepare_chunk (FluidAudio processChunk padding) ────────────────────────

#[test]
fn prepare_chunk_exact_length_passes_through() {
  let chunk: Vec<f32> = (0..CHUNK_SAMPLES).map(|i| i as f32).collect();
  let out = prepare_chunk(&chunk).expect("exact length");
  assert_eq!(&out[..], &chunk[..]);
}

#[test]
fn prepare_chunk_pads_short_by_repeating_last_sample() {
  // Repeat-LAST, not zeros (`VadManager.swift:174-178`).
  let chunk = [0.5f32, -0.25, 0.75];
  let out = prepare_chunk(&chunk).expect("short chunk");
  assert_eq!(&out[..3], &chunk[..]);
  // Every padded slot is the last real sample, 0.75 — a zero-padding mutation
  // (`padded[chunk.len()..]` left at 0.0) turns this red.
  assert!(
    out[3..].iter().all(|&v| v == 0.75),
    "short chunk must pad by repeating the last sample"
  );
}

#[test]
fn prepare_chunk_empty_pads_with_zero() {
  // `chunk.last ?? 0.0`.
  let out = prepare_chunk(&[]).expect("empty chunk");
  assert!(out.iter().all(|&v| v == 0.0));
}

#[test]
fn prepare_chunk_rejects_over_long() {
  let chunk = vec![0.0f32; CHUNK_SAMPLES + 1];
  let err = prepare_chunk(&chunk).expect_err("over-long must reject, not truncate");
  assert_eq!(
    err,
    InferError::ChunkTooLong(ChunkTooLong::new(CHUNK_SAMPLES + 1, CHUNK_SAMPLES))
  );
}

// ── assemble_window (context then chunk) ───────────────────────────────────

#[test]
fn assemble_window_places_context_then_chunk() {
  let mut context = [0.0f32; CONTEXT_SAMPLES];
  for (i, c) in context.iter_mut().enumerate() {
    *c = i as f32;
  }
  let mut chunk = [0.0f32; CHUNK_SAMPLES];
  for (i, c) in chunk.iter_mut().enumerate() {
    *c = 1000.0 + i as f32;
  }
  let window = assemble_window(&context, &chunk);

  // Context occupies [0..64], chunk occupies [64..4160]. A swap of the copy
  // order (chunk first, context last) turns both asserts red.
  assert_eq!(&window[..CONTEXT_SAMPLES], &context[..]);
  assert_eq!(&window[CONTEXT_SAMPLES..], &chunk[..]);
  assert_eq!(window[CONTEXT_SAMPLES - 1], 63.0, "last context sample");
  assert_eq!(window[CONTEXT_SAMPLES], 1000.0, "first chunk sample");
}

// ── next_context (last 64 of the padded chunk) ─────────────────────────────

#[test]
fn next_context_is_the_last_64_samples() {
  let mut chunk = [0.0f32; CHUNK_SAMPLES];
  for (i, c) in chunk.iter_mut().enumerate() {
    *c = i as f32;
  }
  let context = next_context(&chunk);
  // Exactly chunk[4032..4096]. A one-sample offset skew (the exact stitching
  // bug the model-gated `misaligned_context_changes_the_probability` also
  // catches) turns this red.
  assert_eq!(
    context[0],
    (CHUNK_SAMPLES - CONTEXT_SAMPLES) as f32,
    "4032.0"
  );
  assert_eq!(
    context[CONTEXT_SAMPLES - 1],
    (CHUNK_SAMPLES - 1) as f32,
    "4095.0"
  );
  for (i, &v) in context.iter().enumerate() {
    assert_eq!(v, (CHUNK_SAMPLES - CONTEXT_SAMPLES + i) as f32);
  }
}

#[test]
fn stitching_carries_the_previous_chunks_tail_forward() {
  // The end-to-end stitching invariant, hermetically: the context that the
  // next window's [0..64] would carry is exactly the previous (padded)
  // chunk's last 64 samples.
  let mut chunk0 = [0.0f32; CHUNK_SAMPLES];
  for (i, c) in chunk0.iter_mut().enumerate() {
    *c = (i as f32) * 0.001;
  }
  let carried = next_context(&chunk0);
  let chunk1 = [7.0f32; CHUNK_SAMPLES];
  let window1 = assemble_window(&carried, &chunk1);
  assert_eq!(
    &window1[..CONTEXT_SAMPLES],
    &chunk0[CHUNK_SAMPLES - CONTEXT_SAMPLES..]
  );
  assert_eq!(&window1[CONTEXT_SAMPLES..], &chunk1[..]);
}

// ── finite input scan ──────────────────────────────────────────────────────

#[test]
fn check_finite_input_accepts_finite() {
  assert!(check_finite_input(&[0.0, 1.0, -1.0, 0.5]).is_ok());
}

#[test]
fn check_finite_input_rejects_nan_at_its_index() {
  let mut window = [0.0f32; 8];
  window[5] = f32::NAN;
  assert_eq!(
    check_finite_input(&window),
    Err(InferError::NonFiniteInput(5))
  );
}

#[test]
fn check_finite_input_rejects_infinity() {
  let mut window = [0.0f32; 8];
  window[2] = f32::INFINITY;
  assert_eq!(
    check_finite_input(&window),
    Err(InferError::NonFiniteInput(2))
  );
}

// ── output shape re-check ──────────────────────────────────────────────────

#[test]
fn check_output_shape_accepts_the_contract() {
  assert!(check_output_shape(&[1, 1, 1], names::VAD_OUTPUT, &[1, 1, 1]).is_ok());
  assert!(check_output_shape(&[1, STATE_SIZE], names::NEW_HIDDEN_STATE, &[1, STATE_SIZE]).is_ok());
}

#[test]
fn check_output_shape_rejects_a_divergent_shape() {
  let err =
    check_output_shape(&[1, 1], names::VAD_OUTPUT, &[1, 1, 1]).expect_err("wrong rank must fail");
  assert_eq!(
    err,
    InferError::OutputShape(OutputShape::new(
      names::VAD_OUTPUT,
      vec![1, 1],
      vec![1, 1, 1]
    ))
  );
}

// ── VadState ───────────────────────────────────────────────────────────────

#[test]
fn vad_state_initial_is_all_zero() {
  let s = VadState::initial();
  assert!(s.hidden().iter().all(|&v| v == 0.0));
  assert!(s.cell().iter().all(|&v| v == 0.0));
  assert!(s.context().iter().all(|&v| v == 0.0));
  assert_eq!(s, VadState::default());
}

#[test]
fn vad_state_from_parts_round_trips_through_accessors() {
  let hidden = [1.0f32; STATE_SIZE];
  let cell = [2.0f32; STATE_SIZE];
  let context = [3.0f32; CONTEXT_SAMPLES];
  let s = VadState::from_parts(hidden, cell, context);
  assert_eq!(s.hidden(), &hidden);
  assert_eq!(s.cell(), &cell);
  assert_eq!(s.context(), &context);
}

// ── VadModelOptions ────────────────────────────────────────────────────────

#[test]
fn vad_model_options_default_is_all() {
  assert_eq!(VadModelOptions::new().compute(), ComputeUnits::All);
  assert_eq!(VadModelOptions::default(), VadModelOptions::new());
  assert_eq!(DEFAULT_VAD_COMPUTE, ComputeUnits::All);
}

#[test]
fn vad_model_options_with_and_set_compute() {
  let opts = VadModelOptions::new().with_compute(ComputeUnits::CpuOnly);
  assert_eq!(opts.compute(), ComputeUnits::CpuOnly);
  let mut opts = VadModelOptions::new();
  opts.set_compute(ComputeUnits::CpuAndNeuralEngine);
  assert_eq!(opts.compute(), ComputeUnits::CpuAndNeuralEngine);
}

#[cfg(feature = "serde")]
#[test]
fn vad_model_options_serde_roundtrip_and_pinned_spelling() {
  let opts = VadModelOptions::new().with_compute(ComputeUnits::CpuAndGpu);
  let json = serde_json::to_string(&opts).unwrap();
  assert_eq!(json, "{\"compute\":\"cpu_and_gpu\"}");
  let back: VadModelOptions = serde_json::from_str(&json).unwrap();
  assert_eq!(back, opts);
}

#[cfg(feature = "serde")]
#[test]
fn vad_model_options_serde_missing_compute_defaults_to_all() {
  let opts: VadModelOptions = serde_json::from_str("{}").unwrap();
  assert_eq!(opts.compute(), DEFAULT_VAD_COMPUTE);
}

#[cfg(feature = "serde")]
#[test]
fn vad_model_options_serde_unknown_compute_spelling_is_rejected() {
  assert!(serde_json::from_str::<VadModelOptions>("{\"compute\":\"gpu\"}").is_err());
}