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

//! The measured ResNet artifact for the 36-SM RTX 5060 Ti

use super::super::ModuleBinding;
use crate::inference::cuda::candidate::ConvOxide;
use crate::inference::cuda::kernels::{ArtifactHash, LoadedArtifact, ModuleRequest};
use crate::inference::cuda::{ComputeCapability, KernelModule, PtxTier};

/// Both modes use the same sm80 cubin; only TF32 selects the tensor-core entries
///
/// The FP32 entries match the legacy path within 0.04% in the b1/b32 trunk checks
/// All 144 sm80 checks pass against cuDNN and sampled f64 truth
pub(crate) const BINDING: ModuleBinding = ModuleBinding::new(
    ConvOxide::RTX50_SCOPE,
    ModuleRequest::new(
        KernelModule::Resnet,
        PtxTier::Sm80,
        LoadedArtifact::Cubin {
            arch: ComputeCapability::new(12, 0),
            sha256: ArtifactHash::from_hex(
                "f9cada62aab90b3a641ee5fb6232fa951bd6a14d4c38926d550fdded7b2a00c3",
            ),
        },
    ),
);