speakrs 0.6.0

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

//! Native CUDA backend for [`ExecutionMode::Cuda`](super::ExecutionMode::Cuda) and
//! [`ExecutionMode::CudaFast`](super::ExecutionMode::CudaFast), without ONNX Runtime
//!
//! - `CudaRuntime` owns the device context, one stream, and the cuBLAS and cuDNN
//!   handles lazily bound to that stream; `CudaSession` keeps a runtime together with the
//!   models built on it so they can move between threads
//! - `KernelModule` embeds the committed cuda-oxide PTX, one module per area, as the
//!   best shipped variant for each enabled GPU target feature
//! - `DeviceTensor` is a shaped device buffer and `SafetensorsFile` uploads named
//!   weights with shape checks
//! - `Sgemm` describes a row-major matrix product; optional library plans use
//!   [`CudaMath`] defaults to FP32 with TF32 disabled
//! - `CudaFbank`, `ResNetEmbedding` and `CudaSegmentation` are the three models
//!
//! Public options and errors describe inference policy; [`tune_cuda`] measures
//! approved boundary choices on an explicitly requested device

/// Target-only builds contain no optional numerical libraries
const fn driver_only() -> bool {
    !cfg!(feature = "_cuda-libraries")
}

#[cfg(feature = "_cuda-libraries")]
mod blas;
mod buffer;
// kernel workers' candidates consume this interface; until one lands, parts of it
// and the candidate PTX areas are unused
#[allow(dead_code)]
mod candidate;
mod dense;
mod device;
#[cfg(feature = "_cuda-libraries")]
mod dnn;
mod embedding;
mod error;
mod fbank;
mod gemm;
mod geometry;
pub(crate) mod implementation;
#[cfg(test)]
mod kernel_inventory_tests;
mod kernels;
#[cfg(feature = "_cuda-libraries")]
mod libraries;
#[cfg(all(test, target_os = "linux"))]
mod loader_tests;
mod math;
mod options;
#[cfg(all(test, feature = "_cuda-libraries"))]
mod probe;
mod runtime;
mod segmentation;
mod session;
mod tier;
mod tuning;
mod weights;

#[cfg(test)]
pub(crate) mod batch_class_tests;
#[cfg(test)]
mod fp16_range_tests;
#[cfg(all(test, feature = "_cuda-libraries"))]
mod test_support;
#[cfg(all(test, feature = "_cuda-libraries"))]
mod tests;

pub(crate) use buffer::DeviceTensor;
#[cfg(all(test, feature = "_cuda-libraries"))]
use dnn::ConvPlanner;
pub(crate) use embedding::{
    EMBEDDING_DIM, EmbeddingBatch, EmbeddingBatchClass, ResNetEmbedding, SPEAKERS_PER_CHUNK,
    SharedEmbeddingActivations,
};
pub use error::{CudaError, CudaLibrary, GeometryError, WeightFault};
pub(crate) use fbank::{
    CudaFbank, FBANK_FRAMES, FBANK_MEL_BINS, FBANK_WINDOW_SAMPLES, FbankBuffers,
};
use gemm::Sgemm;
#[cfg(all(test, feature = "_cuda-libraries"))]
use geometry::Conv2d;
use kernels::{KernelModule, LoadedKernels};
pub use math::CudaMath;
pub use options::CudaGraphs;
#[cfg(feature = "_cuda-libraries")]
pub use options::CudaLstmAlgorithm;
pub(crate) use runtime::CudaRuntime;
pub(crate) use segmentation::{CudaSegmentation, SegmentationOptions};
pub(crate) use session::CudaSession;
pub use tier::{ComputeCapability, PtxTier};
pub use tuning::{
    CudaTuneError, CudaTuneMeasurement, CudaTuneOptions, CudaTuneReport, CudaTuneRow, tune_cuda,
};
pub(crate) use weights::SafetensorsFile;