speakrs 0.7.0

Fast Rust speaker diarization with pyannote-level accuracy and native CPU/CoreML/CUDA inference
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

//! Graph reuse and row ordering through the production embedding wrapper

use super::{Batches, CHUNK_MASKS_LEN, FBANK_LEN};
use crate::inference::cuda::batch_class_tests::{Case, load_case, parity, setup};
use crate::inference::cuda::fp16_range_tests::{SyntheticModel, runtime, scratch};
use crate::inference::cuda::{
    CudaError, CudaMath, CudaRuntime, CudaSession, DeviceTensor, EMBEDDING_DIM, ResNetEmbedding,
    SafetensorsFile,
};

fn session(weights: &SafetensorsFile, math: CudaMath) -> Result<CudaSession<Batches>, CudaError> {
    CudaSession::new(crate::inference::cuda::CudaRuntime::new(0)?, |runtime| {
        Ok(Batches {
            model: ResNetEmbedding::load(runtime, weights, math)?,
            graphs: true,
            plans: std::array::from_fn(|_| None),
            spare: None,
            next_spare: false,
            activations: None,
        })
    })
}

fn reordered(source: &Case, chunks: usize, phase: usize) -> Case {
    let mut fbank = Vec::with_capacity(chunks * FBANK_LEN);
    let mut masks = Vec::with_capacity(chunks * CHUNK_MASKS_LEN);
    let mut expected = Vec::with_capacity(chunks * 3 * 256);
    for index in 0..chunks {
        let row = (index + phase) % source.chunks;
        fbank.extend_from_slice(&source.fbank[row * FBANK_LEN..(row + 1) * FBANK_LEN]);
        for speaker in 0..3 {
            let source_row = row * 3 + (speaker + phase) % 3;
            masks.extend_from_slice(&source.masks[source_row * 589..(source_row + 1) * 589]);
            expected.extend_from_slice(&source.expected[source_row * 256..(source_row + 1) * 256]);
        }
    }
    Case {
        fbank,
        masks,
        expected,
        chunks,
    }
}

fn embed(session: &mut CudaSession<Batches>, case: &Case) -> Result<Vec<f32>, CudaError> {
    session.run(|runtime, batches| {
        batches.embed(
            runtime,
            case.chunks,
            &case.masks,
            |runtime, rows, target| {
                target.copy_from_host(
                    runtime.stream(),
                    &case.fbank[rows.start * FBANK_LEN..rows.end * FBANK_LEN],
                )
            },
        )
    })
}

#[test]
#[ignore = "GPU wrapper check; run under the GPU lock with reference tensors"]
fn embedding_wrapper_remainders_and_graph_reuse() -> Result<(), CudaError> {
    let Some((runtime, root)) = setup("embedding_wrapper_remainders_and_graph_reuse") else {
        return Ok(());
    };
    drop(runtime);
    let source = load_case(&root, "wespeaker-multimask-tail-b32", "test_and_short_b32");
    assert_eq!(source.chunks, 32);
    let weights = SafetensorsFile::open(std::env::var_os("TRUNK_WEIGHTS").map_or_else(
        || root.join("wespeaker-multimask-tail/wespeaker-multimask-tail.safetensors"),
        std::path::PathBuf::from,
    ))?;
    for math in [CudaMath::Tf32, CudaMath::Fp32] {
        let mut reused = session(&weights, math)?;
        for chunks in 1..=32 {
            let case = reordered(&source, chunks, 0);
            let actual = embed(&mut reused, &case)?;
            assert!(actual.iter().all(|value| value.is_finite()));
            assert!(parity(&actual, &case.expected, 256).min_cosine >= 0.999);
        }

        // change both window order and speaker order before reusing every captured class
        let mut fresh = session(&weights, math)?;
        for chunks in (1..=32).rev() {
            let case = reordered(&source, chunks, 7);
            let actual = embed(&mut reused, &case)?;
            let control = embed(&mut fresh, &case)?;
            assert!(actual.iter().all(|value| value.is_finite()));
            assert_eq!(
                actual, control,
                "{math:?} remainder {chunks}: stale graph state"
            );
            assert!(parity(&actual, &case.expected, 256).min_cosine >= 0.999);
        }
        eprintln!(
            "WRAPPER_REUSE math={math:?} remainders=32 changed_inputs=true fresh_session_parity=true"
        );
    }
    Ok(())
}

#[test]
#[ignore = "GPU retained-memory check; run under the GPU lock with reference tensors"]
fn embedding_32_then_31_retains_one_activation_allocation() -> Result<(), CudaError> {
    let Some((runtime, root)) = setup("embedding_32_then_31_retains_one_activation_allocation")
    else {
        return Ok(());
    };
    drop(runtime);
    let source = load_case(&root, "wespeaker-multimask-tail-b32", "test_and_short_b32");
    let weights = SafetensorsFile::open(std::env::var_os("TRUNK_WEIGHTS").map_or_else(
        || root.join("wespeaker-multimask-tail/wespeaker-multimask-tail.safetensors"),
        std::path::PathBuf::from,
    ))?;
    let mut reused = session(&weights, CudaMath::Tf32)?;
    for chunks in [32, 31] {
        let case = reordered(&source, chunks, 0);
        let actual = embed(&mut reused, &case)?;
        assert!(parity(&actual, &case.expected, 256).min_cosine >= 0.999);
        reused.run(|_, batches| {
            // measure real device storage, not the sum of each class's logical views
            assert_eq!(
                batches.activations.as_ref().unwrap().retained_bytes(),
                1_144_586_240
            );
            Ok::<_, CudaError>(())
        })?;
    }
    reused.run(|_, batches| {
        assert!(batches.plans.iter().all(Option::is_some));
        let allocation = batches.activations.as_ref().unwrap();
        assert!(
            batches
                .plans
                .iter()
                .flatten()
                .all(|batch| batch.shares_activations(allocation))
        );
        Ok::<_, CudaError>(())
    })?;
    Ok(())
}

/// A full batch whose download is queued keeps its result while the next full batch runs
/// on the other 32-chunk batch, including the FP16 range recompute that reruns the
/// queued batch from its own inputs after the next one has overwritten the shared
/// activations
///
/// The synthetic model passes the filterbank value plus 128 through `resnet.layer1.0`
/// to embedding element 0: an input of 0 gives 256 and saturates FP16 tiles where the
/// device selects them, and an input of -120 gives 16, in range
#[test]
fn queued_download_survives_the_next_full_batch() -> Result<(), CudaError> {
    let Some(runtime) = runtime("queued_download_survives_the_next_full_batch") else {
        return Ok(());
    };
    let mut model = SyntheticModel::channel_zero_path(128.0);
    model.center("resnet.conv1", 1.0);
    model.center("resnet.layer1.0.conv1", 1.0);
    model.center("resnet.layer1.0.conv2", 1.0);
    let path = scratch("queued-download");
    let weights = model.write(&path);
    let mut session = CudaSession::new(runtime, |runtime| {
        Ok(Batches {
            model: ResNetEmbedding::load(runtime, &weights, CudaMath::Tf32)?,
            graphs: true,
            plans: std::array::from_fn(|_| None),
            spare: None,
            next_spare: false,
            activations: None,
        })
    })?;
    let masks = vec![1.0; 32 * CHUNK_MASKS_LEN];
    let fill = |value: f32| {
        move |runtime: &CudaRuntime, rows: std::ops::Range<usize>, target: &mut DeviceTensor| {
            target.copy_from_host(runtime.stream(), &vec![value; rows.len() * FBANK_LEN])
        }
    };

    // three full batches, each launched before the previous one is read, so both
    // 32-chunk batches take a turn holding a queued download
    let inputs = [(0.0, 256.0), (-120.0, 16.0), (0.0, 256.0)];
    let (outputs, saturating, recomputed) = session.run(|runtime, batches| {
        let mut outputs = Vec::new();
        let mut queued: Option<super::Launch> = None;
        for (input, _) in inputs {
            let (done, launch) = batches.launch(runtime, 32, &masks, fill(input))?;
            assert!(done.is_empty());
            let launch = launch.expect("a full batch is left running");
            if let Some(previous) = queued.take() {
                assert_ne!(previous.spare, launch.spare);
                outputs.push(batches.finish_download(runtime, previous)?);
            }
            batches.enqueue_download(runtime, &launch)?;
            queued = Some(launch);
        }
        let last = queued.expect("the last batch is queued");
        outputs.push(batches.finish_download(runtime, last)?);

        let batch = batches.plans[4].as_ref().expect("planned batch");
        let saturating = batch.fp16_convs().contains(&"resnet.layer1.0.conv1");
        Ok::<_, CudaError>((outputs, saturating, batch.recomputed_without_fp16()))
    })?;
    eprintln!("QUEUED_DOWNLOAD saturating={saturating} recomputed={recomputed}");
    // the batches given 0 ran in the non-spare batch, and saturate where FP16 tiles run
    assert_eq!(recomputed, saturating);
    for ((_, expected), rows) in inputs.iter().zip(&outputs) {
        assert_eq!(rows.len(), 32 * 3 * EMBEDDING_DIM);
        for (row, embedding) in rows.chunks(EMBEDDING_DIM).enumerate() {
            assert_eq!(embedding[0], *expected, "row {row}");
            assert!(embedding[1..].iter().all(|&value| value == 0.0));
        }
    }
    drop(session);
    drop(weights);
    std::fs::remove_file(&path).expect("removed scratch model");
    Ok(())
}