Cargo.toml

[workspace]
members = [
    "xtask",
    "crates/speakrs-ptx-sm75",
    "crates/speakrs-ptx-sm80",
    "crates/speakrs-ptx-sm120",
    "crates/speakrs-sdk",
    "crates/speakrs-python",
]
exclude = ["scripts/", "crates/speakrs-cuda-kernels"]

# shared by the kernel artifact crates, which release in lockstep with speakrs
[workspace.package]
version = "0.7.0"
edition = "2024"
rust-version = "1.89"
license = "Apache-2.0"
repository = "https://github.com/avencera/speakrs"
authors = ["Praveen Perera"]

[package]
name = "speakrs"
version = "0.7.0"
edition = "2024"
rust-version = "1.89"
description = "Fast Rust speaker diarization with pyannote-level accuracy and native CPU/CoreML/CUDA inference"
license = "Apache-2.0"
repository = "https://github.com/avencera/speakrs"
documentation = "https://docs.rs/speakrs"
readme = "README.md"
authors = ["Praveen Perera"]
keywords = ["speaker", "diarization", "audio", "pyannote", "onnx"]
categories = ["multimedia::audio", "science"]
include = ["/src/**", "/examples/**", "/README.md", "/LICENSE", "/Cargo.toml"]

[features]
default = ["online"]
online = ["dep:hf-hub", "hf-hub/ureq", "hf-hub/native-tls"]
cpu = ["dep:safetensors", "dep:rustfft", "dep:libloading", "_backend"]
coreml = [
    "dep:block2",
    "dep:objc2",
    "dep:objc2-core-ml",
    "dep:objc2-foundation",
    "_backend",
]
# Linux-only native NVIDIA backend; CUDA is never part of default
# cuda defaults to speakrs kernels on every GPU, with cuDNN/cuBLAS fallback
cuda = ["cuda-sm75", "cuda-sm80", "cuda-sm90", "cuda-sm120", "_cuda-libraries"]
# internal: optional numerical libraries, added only by cuda
_cuda-libraries = ["_cuda", "cudarc/cublas", "cudarc/cudnn"]
# internal: driver support, without optional numerical libraries or kernel artifacts
_cuda = ["dep:cudarc", "dep:libloading", "dep:safetensors", "dep:sha2", "dep:serde", "dep:serde_json", "dep:getrandom", "_backend"]
# each target embeds every area's best shipped variant for that target, so it pulls in
# the artifact crate of every tier it may use; areas without a newer variant use sm75
# default-features = false with cuda-rtx50 alone is a driver-only build
cuda-sm75 = ["_cuda", "dep:speakrs-ptx-sm75"]
cuda-sm80 = ["_cuda", "dep:speakrs-ptx-sm75", "dep:speakrs-ptx-sm80"]
cuda-sm90 = ["_cuda", "dep:speakrs-ptx-sm75", "dep:speakrs-ptx-sm80"]
cuda-sm120 = ["_cuda", "dep:speakrs-ptx-sm75", "dep:speakrs-ptx-sm80", "dep:speakrs-ptx-sm120"]
# RTX 20: Turing; RTX 30: Ampere; RTX 40: Ada; A100: Ampere
cuda-rtx20 = ["cuda-sm75"]
cuda-rtx30 = ["cuda-sm80"]
cuda-rtx40 = ["cuda-sm80"]
cuda-a100 = ["cuda-sm80"]
# RTX 50: consumer Blackwell
cuda-rtx50 = ["cuda-sm120"]
migraphx = ["_ort", "ort/migraphx", "_backend"]
# load ONNX Runtime at run time for MIGraphX or caller-created CPU sessions
# native CPU model execution does not use this runtime
load-dynamic = ["_ort", "ort/load-dynamic", "dep:libloading"]

# internal: ONNX Runtime support for MIGraphX and external session helpers
_ort = ["dep:ort"]
# internal: enabled by every inference backend; gates the crate items in `lib.rs`
_backend = []
_metrics = []

[package.metadata.docs.rs]
features = ["cpu", "cuda", "cuda-sm75", "cuda-sm80", "cuda-sm90", "cuda-sm120", "migraphx", "load-dynamic", "_metrics"]

[dependencies]
serde = { version = "1", features = ["derive"], optional = true }
serde_json = { version = "1", optional = true }
sha2 = { version = "0.10", optional = true }
getrandom = { version = "0.3", features = ["std"], optional = true }
hf-hub = { version = "0.5", optional = true, default-features = false }

# math
rustfft = { version = "6.4.1", optional = true }
kodama = "0.3.0"
ndarray = { version = "0.17.2", features = ["matrixmultiply-threading"] }
# the default npz feature pulls in zip 6; tests read .npz captures with zip 9
ndarray-npy = { version = "0.10.0", default-features = false }

ort = { version = "2.0.0-rc.13", features = ["ndarray"], optional = true }

# native CUDA: libcuda, cuBLAS and cuDNN are loaded at run time, so the crate builds
# without a CUDA toolkit (including on macOS). The 12.4 API floor keeps the CUDA 12.4
# canary image working; the PTX itself is JIT-compiled by the installed driver
cudarc = { version = "0.19.10", optional = true, default-features = false, features = [
    "std",
    "driver",
    # the precompiled `Ptx` type requires this feature but does not load NVRTC
    "nvrtc",
    "dynamic-loading",
    "cuda-12040",
] }
safetensors = { version = "0.8.0", optional = true, default-features = false, features = ["alloc"] }
# prebuilt PTX and cubins, one crate per tier so each stays under the crates.io upload
# limit; exact pins because the host code and the kernels change together
speakrs-ptx-sm75 = { version = "=0.7.0", path = "crates/speakrs-ptx-sm75", optional = true }
speakrs-ptx-sm80 = { version = "=0.7.0", path = "crates/speakrs-ptx-sm80", optional = true }
speakrs-ptx-sm120 = { version = "=0.7.0", path = "crates/speakrs-ptx-sm120", optional = true }
libloading = { version = "0.9.0", optional = true }

# logging
tracing = "0.1.44"

# error handling
thiserror = "2.0.18"

# concurrency
crossbeam-channel = "0.5"
rayon = "1.11.0"

[target.'cfg(target_os = "macos")'.dependencies]
objc2 = { version = "0.6", optional = true }
block2 = { version = "0.6", optional = true }
objc2-core-ml = { version = "0.3", optional = true, features = [
    "MLModel",
    "MLMultiArray",
    "MLModelConfiguration",
    "MLDictionaryFeatureProvider",
    "MLFeatureValue",
    "MLFeatureProvider",
    "MLPredictionOptions",
    "alloc",
    "block2",
] }

objc2-foundation = { version = "0.3", optional = true, features = [
    "NSArray",
    "NSDictionary",
    "NSError",
    "NSObject",
    "NSString",
    "NSURL",
    "NSValue",
    "NSSet",
    "NSEnumerator",
    "NSObjCRuntime",
] }

[[bin]]
name = "speakrs"
required-features = ["_cuda"]

# the examples load native model assets in `ExecutionMode::Cpu`
[[example]]
name = "assign_transcript_speakers"
required-features = ["cpu"]

[[example]]
name = "diarize_wav"
required-features = ["cpu"]

[[example]]
name = "print_turns"
required-features = ["cpu"]

[[example]]
name = "queued"
required-features = ["cpu"]

[[example]]
name = "speaker_airtime"
required-features = ["cpu"]

[dev-dependencies]
hound = "3.5.1"
sha2 = "0.10"
libloading = "0.9.0"
approx = "0.5.1"
zip = { version = "9", default-features = false, features = ["deflate-flate2-zlib-rs"] }
serde_json = "1"
tracing-subscriber = { version = "0.3.22", features = ["env-filter"] }

# native model tests need optimized kernels; test assertions remain enabled
[profile.test]
opt-level = 2