svod-model 0.1.0-alpha.3

Pretrained models inference abstraction.
Documentation 
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
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294

use svod_arch::ctc::{CtcDecoder, GreedyDecoder};
use svod_dtype::DType;
use svod_tensor::{Tensor, Variable};

use crate::gigaam::{ConvNormType, GigaAm, GigaAmConfig, GigaAmEncoderJit, SubsamplingMode};

pub(super) fn test_config() -> GigaAmConfig {
    GigaAmConfig {
        max_batch_size: 8,
        n_mels: 64,
        d_model: 32,
        n_heads: 4,
        n_layers: 2,
        d_ff: 128,
        conv_kernel: 5,
        subsampling_factor: 4,
        subsampling_mode: SubsamplingMode::Conv1d,
        subs_kernel_size: 5,
        conv_norm_type: ConvNormType::LayerNorm,
        vocab_size: 34,
        sample_rate: 16000,
        n_fft: 320,
        hop_length: 160,
        win_length: 320,
        mel_center: false,
        max_mel_frames: 512,
        max_encoder_frames: 128,
        decoder: CtcDecoder::Greedy(GreedyDecoder::new(Vec::new())),
        transducer: None,
    }
}

fn model_with_random_weights() -> GigaAm {
    GigaAm::with_random_weights(test_config())
}

fn read_prefix_f32(t: &Tensor, len: usize) -> Vec<f32> {
    let buf = t.buffer().unwrap();
    buf.as_array::<f32>().unwrap().as_slice().unwrap()[..len].to_vec()
}

// ---------------------------------------------------------------------------
// Cheap default tests (no realize): config + structural invariants.
// ---------------------------------------------------------------------------

#[test]
fn test_output_length_matches_forward() {
    let model = GigaAm::with_random_weights(test_config());

    let x = Tensor::full(&[1, 100, 64], 0.0f32, DType::Float32).unwrap();
    let out = model.encoder.subsampling.forward(&x).unwrap();
    let actual_t = out.shape().unwrap()[1].as_const().unwrap();
    assert_eq!(model.encoder.subsampling_output_length(100), actual_t);

    let x2 = Tensor::full(&[1, 50, 64], 0.0f32, DType::Float32).unwrap();
    let out2 = model.encoder.subsampling.forward(&x2).unwrap();
    let actual_t2 = out2.shape().unwrap()[1].as_const().unwrap();
    assert_eq!(model.encoder.subsampling_output_length(50), actual_t2);
}

#[test]
fn test_rope_cache_uses_encoder_bound() {
    let model = model_with_random_weights();
    let cfg = test_config();

    assert_eq!(model.encoder.cos_cache.shape().unwrap()[0].as_const().unwrap(), cfg.max_encoder_frames);
    assert_eq!(model.encoder.sin_cache.shape().unwrap()[0].as_const().unwrap(), cfg.max_encoder_frames);
    assert_ne!(cfg.max_encoder_frames, cfg.max_mel_frames);
}

#[test]
fn test_rope_cache_uses_pos_emb_max_len_as_base() {
    let model = model_with_random_weights();
    let cfg = test_config();
    let d_k = cfg.d_model / cfg.n_heads;
    let half_d = d_k / 2;
    let pos = 1usize;
    let freq_idx = 1usize;
    let angle = pos as f32 / (cfg.max_encoder_frames as f32).powf(2.0 * freq_idx as f32 / d_k as f32);
    let flat_idx = pos * half_d + freq_idx;

    let cos = model.encoder.cos_cache.as_vec::<f32>().unwrap();
    let sin = model.encoder.sin_cache.as_vec::<f32>().unwrap();

    assert!((cos[flat_idx] - angle.cos()).abs() < 1e-6);
    assert!((sin[flat_idx] - angle.sin()).abs() < 1e-6);
}

#[test]
fn test_subsampled_max_mel_fits_encoder_bound() {
    let model = model_with_random_weights();
    let cfg = test_config();
    let t_sub = model.encoder.subsampling_output_length(cfg.max_mel_frames);
    assert!(
        t_sub <= cfg.max_encoder_frames,
        "subsampled max_mel ({t_sub}) > max_encoder_frames ({})",
        cfg.max_encoder_frames
    );
}

#[test]
#[should_panic(expected = "with_b_bound(0) creates empty range")]
fn test_with_b_bound_panics_on_empty_range() {
    let model = model_with_random_weights();
    let _jit = GigaAmEncoderJit::new(model).with_b_bound(0);
}

// ---------------------------------------------------------------------------
// Heavy tests (realize / prepare): unique signal, gated behind --ignored.
// Run on demand when actively touching the GigaAM encoder, JIT plumbing, or
// the schedule pipeline. Each costs ~30s of prepare with random weights.
// ---------------------------------------------------------------------------

#[test]
#[ignore = "heavy: realize-based bounds-error JIT smoke"]
fn test_batched_jit_rejects_t_above_max_mel_frames() {
    let model = model_with_random_weights();
    let cfg = test_config();
    let mut jit = GigaAmEncoderJit::new(model);
    jit.prepare(crate::jit::InputSpec::f32(&[1, cfg.n_mels, cfg.max_mel_frames]), crate::jit::InputSpec::i32(&[1]))
        .unwrap();
    let err = jit.execute_with_vars(&[("b", 1), ("t", cfg.max_mel_frames as i64 + 1)]).unwrap_err();
    assert_runtime_bounds_err(err);
}

#[test]
#[ignore = "heavy: full encoder forward at max_mel_frames"]
fn test_encode_batch_near_max_mel_runs() {
    let model = model_with_random_weights();
    let cfg = test_config();
    let t = cfg.max_mel_frames;

    let x = Tensor::full(&[1, cfg.n_mels, t], 0.1f32, DType::Float32).unwrap();
    let lengths = Tensor::from_slice([t as i32]);
    let b_var = Variable::new("B", 1, cfg.max_batch_size as i64);
    let t_var = Variable::new("T", 1, cfg.max_mel_frames as i64);
    let b1 = b_var.bind(1).unwrap();
    let t_bound = t_var.bind(t as i64).unwrap();

    let mut out = model.encoder.forward_batch(&x, &lengths, &b1, &t_bound).unwrap();
    out.realize().unwrap();
    assert!(out.buffer().unwrap().size() > 0);
}

#[test]
#[ignore = "heavy: batched-vs-single numerical consistency (the suite's only correctness assert)"]
fn test_single_vs_batch_consistency() {
    let model = model_with_random_weights();
    let d = test_config().d_model;
    let n_mels = test_config().n_mels;
    let t = 10;
    let t_sub = model.encoder.subsampling_output_length(t);

    let x1 = Tensor::full(&[1, n_mels, t], 0.5f32, DType::Float32).unwrap();
    let x2 = Tensor::full(&[1, n_mels, t], 0.3f32, DType::Float32).unwrap();
    let lengths_single = Tensor::from_slice([t as i32]);

    let b_var = Variable::new("B", 1, test_config().max_batch_size as i64);
    let t_var = Variable::new("T", 1, test_config().max_mel_frames as i64);
    let b1 = b_var.bind(1).unwrap();
    let t1 = t_var.bind(t as i64).unwrap();

    let mut out1 = model.encoder.forward_batch(&x1, &lengths_single, &b1, &t1).unwrap();
    out1.realize().unwrap();
    let data1 = read_prefix_f32(&out1, d * t_sub);

    let mut out2 = model.encoder.forward_batch(&x2, &lengths_single, &b1, &t1).unwrap();
    out2.realize().unwrap();
    let data2 = read_prefix_f32(&out2, d * t_sub);

    let batch = {
        let mut x1r = x1.clone();
        x1r.realize().unwrap();
        let d1 = x1r.as_vec::<f32>().unwrap();
        let mut x2r = x2.clone();
        x2r.realize().unwrap();
        let d2 = x2r.as_vec::<f32>().unwrap();
        let mut batch_data = vec![0.0f32; 2 * n_mels * t];
        batch_data[..n_mels * t].copy_from_slice(&d1);
        batch_data[n_mels * t..].copy_from_slice(&d2);
        ndarray::Array3::from_shape_vec((2, n_mels, t), batch_data).unwrap()
    };
    let batch_tensor = Tensor::from_ndarray(&batch);
    let batch_lengths = Tensor::from_slice([t as i32, t as i32]);

    let b2 = b_var.bind(2).unwrap();
    let mut batch_out = model.encoder.forward_batch(&batch_tensor, &batch_lengths, &b2, &t1).unwrap();
    batch_out.realize().unwrap();
    let batch_data = read_prefix_f32(&batch_out, 2 * d * t_sub);

    assert_eq!(data1.len() * 2, batch_data.len());

    for (i, (&b, &s)) in batch_data[..data1.len()].iter().zip(data1.iter()).enumerate() {
        assert!((b - s).abs() < 1e-4, "batch[0] mismatch at {}: batch={} single={}", i, b, s);
    }
    for (i, (&b, &s)) in batch_data[data1.len()..].iter().zip(data2.iter()).enumerate() {
        assert!((b - s).abs() < 1e-4, "batch[1] mismatch at {}: batch={} single={}", i, b, s);
    }
}

#[test]
#[ignore = "heavy: NaN/Inf detector across encoder forward"]
fn test_encode_batch_full_lengths_finite() {
    let model = model_with_random_weights();
    let cfg = test_config();
    let t = 256usize;

    let x = Tensor::full(&[2, cfg.n_mels, t], 0.1f32, DType::Float32).unwrap();
    let lengths = Tensor::from_slice([t as i32, t as i32]);

    let b_var = Variable::new("B", 1, cfg.max_batch_size as i64);
    let t_var = Variable::new("T", 1, cfg.max_mel_frames as i64);
    let b2 = b_var.bind(2).unwrap();
    let t_bound = t_var.bind(t as i64).unwrap();

    let mut out = model.encoder.forward_batch(&x, &lengths, &b2, &t_bound).unwrap();
    out.realize().unwrap();

    let buf = out.buffer().unwrap();
    let data = buf.as_array::<f32>().unwrap();
    for v in data.as_slice().unwrap() {
        assert!(v.is_finite(), "encode_batch produced non-finite value: {v}");
    }
}

#[test]
#[ignore = "heavy: symbolic seq-len threading through compiled kernels"]
fn test_encode_batch_respects_dynamic_seq_len() {
    let model = model_with_random_weights();
    let cfg = test_config();
    let t_dynamic = 64usize;

    let mut jit = GigaAmEncoderJit::new(model);
    jit.prepare(
        crate::jit::InputSpec::f32(&[cfg.max_batch_size, cfg.n_mels, cfg.max_mel_frames]),
        crate::jit::InputSpec::i32(&[cfg.max_batch_size]),
    )
    .unwrap();
    let profiles = jit.execute_with_vars_profiled(&[("b", 1), ("t", t_dynamic as i64)]).unwrap();

    assert!(!profiles.is_empty(), "expected kernels for profiled dynamic execute");
    assert!(
        profiles.iter().any(|p| p.kernel.var_names.iter().any(|name| name == "t")),
        "expected at least one kernel to keep dynamic seq var 't'"
    );
}

fn assert_runtime_bounds_err(err: crate::jit::JitError) {
    match err {
        crate::jit::JitError::Runtime { source: svod_runtime::Error::Execution { reason } } => {
            assert!(reason.contains("outside bounds"), "unexpected runtime error: {reason}");
        }
        other => panic!("expected runtime bounds error, got {other:?}"),
    }
}

#[test]
#[ignore = "heavy: variable upper-bound enforcement at execute time"]
fn test_with_b_bound_shrinks_upper_bound() {
    let model = model_with_random_weights();
    let cfg = test_config();
    let mut jit = GigaAmEncoderJit::new(model).with_b_bound(2);
    jit.prepare(crate::jit::InputSpec::f32(&[2, cfg.n_mels, 64]), crate::jit::InputSpec::i32(&[2])).unwrap();
    jit.execute_with_vars(&[("b", 2), ("t", 64)]).unwrap();
    assert_runtime_bounds_err(jit.execute_with_vars(&[("b", 3), ("t", 64)]).unwrap_err());
}

#[test]
#[ignore = "heavy: with_t_fixed should fold the symbolic dim out of compiled kernels"]
fn test_with_t_fixed_specializes_kernels() {
    let model = model_with_random_weights();
    let cfg = test_config();
    let pinned_t = 64usize;
    let mut jit = GigaAmEncoderJit::new(model).with_t_fixed(pinned_t);
    jit.prepare(crate::jit::InputSpec::f32(&[1, cfg.n_mels, pinned_t]), crate::jit::InputSpec::i32(&[1])).unwrap();
    let kernels = jit.prepared_kernels().unwrap();
    let any_kernel_keeps_t = kernels.iter().any(|k| k.kernel.var_names.iter().any(|n| n == "t"));
    assert!(!any_kernel_keeps_t, "with_t_fixed should fold `t` out of kernel var lists");
    let any_kernel_keeps_b = kernels.iter().any(|k| k.kernel.var_names.iter().any(|n| n == "b"));
    assert!(any_kernel_keeps_b, "`b` is still dynamic and should remain in some kernel's var list");

    jit.execute_with_vars(&[("b", 1), ("t", pinned_t as i64)]).unwrap();
}

#[test]
#[ignore = "heavy: variable lower-bound enforcement at execute time"]
fn test_with_b_min_bound_raises_lower_bound() {
    let model = model_with_random_weights();
    let cfg = test_config();
    let mut jit = GigaAmEncoderJit::new(model).with_b_min_bound(2);
    jit.prepare(crate::jit::InputSpec::f32(&[2, cfg.n_mels, 64]), crate::jit::InputSpec::i32(&[2])).unwrap();
    jit.execute_with_vars(&[("b", 2), ("t", 64)]).unwrap();
    assert_runtime_bounds_err(jit.execute_with_vars(&[("b", 1), ("t", 64)]).unwrap_err());
}