[package]
name = "tract"
version = "0.23.9"
license = "MIT OR Apache-2.0"
authors = ["Mathieu Poumeyrol <kali@zoy.org>"]
description = "Tiny, no-nonsense, self contained, TensorFlow and ONNX inference"
repository = "https://github.com/sonos/tract"
keywords = ["NeuralNetworks"]
categories = ["science"]
autobenches = false
edition = "2024"
rust-version.workspace = true
include = ["Cargo.toml", "src/**/*.rs", "LICENSE*", "tests"]

[dependencies]
anyhow.workspace = true
boow.workspace = true
erased-serde.workspace = true
flate2 = { workspace = true, optional = true }
half.workspace = true
tract-api.workspace = true
tract-core.workspace = true
tract-nnef.workspace = true
tract-onnx-opl.workspace = true
tract-onnx = { workspace = true, optional = true }
tract-extra.workspace = true
tract-pulse = { workspace = true, optional = true }
tract-pulse-opl.workspace = true
tract-libcli = { workspace = true, optional = true }
tract-transformers = { workspace = true, optional = true }
serde_json.workspace = true

[target.'cfg(any(target_vendor = "apple"))'.dependencies]
tract-metal = { workspace = true, optional = true }

[target.'cfg(any(target_os = "linux", target_os = "windows"))'.dependencies]
tract-cuda = { workspace = true, default-features = false, optional = true }


[dev-dependencies]
icu_normalizer.workspace = true
icu_properties.workspace = true
reqwest.workspace = true
rustls.workspace = true
tempfile.workspace = true
serde_json.workspace = true

[features]
default = ["cuda-13000", "metal", "flate2", "onnx", "profile", "pulse", "transformers"]
complex = []

# Gzip support for NNEF archives: reading `.nnef.tgz` models and writing them
# back. Enabled by default; consumers that only load uncompressed NNEF drop
# flate2 and miniz_oxide with `default-features = false`.
flate2 = ["dep:flate2", "tract-nnef/flate2"]

# The ONNX parser: `onnx()`, the `Onnx` loader and the `InferenceModel` /
# `InferenceFact` shape-inference types it produces. Enabled by default;
# consumers that only load NNEF drop tract-onnx and its operator register with
# `default-features = false`. `enable_onnx` is unaffected — the ONNX operators
# an NNEF graph may carry live in tract-onnx-opl, which stays either way.
onnx = ["dep:tract-onnx"]

# `Runnable::cost_json` and `Runnable::profile_json`, backed by tract-libcli.
# Enabled by default; with it off both return an error, and tract-libcli, with
# clap and its other CLI dependencies, stays out of the build.
profile = ["dep:tract-libcli"]

# Offline pulsification: the `pulse` and `blockify` model transforms, and the
# write side of pulsed NNEF. Enabled by default; consumers that only load and
# run an already pulsified model keep `enable_pulse` off tract-pulse-opl alone
# with `default-features = false`. Implies `transformers`, which tract-pulse
# needs for the DiagGather pulsifier and blockify.
pulse = ["dep:tract-pulse", "tract-extra/pulse", "transformers"]

# The tract_transformers NNEF extension: attention, KV cache and rope
# operators, and the transforms that detect them, plus the ONNX attention
# operators that lower onto them and libcli's causal-LLM tensor specs. Enabled
# by default; with it off `enable_tract_transformers` fails and a model using
# those operators does not load, so only drop it for a model known not to need
# them.
transformers = ["dep:tract-transformers", "tract-onnx?/transformers", "tract-libcli?/transformers"]
# default = [ "dylib" ]
# dylib = []
# staticlib = []

# Gates the `extern crate tract_metal` that registers the Metal runtime, on the
# Apple targets that have it. Enabled by default; CPU-only consumers drop the
# backend, and the shader sources it embeds, with `default-features = false`.
metal = ["dep:tract-metal", "tract-libcli?/metal"]

# Marker feature implicitly enabled by every cuda-XXXXX selector below.
# Gates the `extern crate tract_cuda` that registers the CUDA runtime.
# Not meant to be enabled directly — always pick a concrete cuda-XXXXX so
# cudarc has an API version to bind against.
cuda = []

# CUDA driver-API selectors (linux/windows targets only). Pass-through to
# tract-cuda. Exactly one must be active for CUDA support; the default pulls
# cuda-13000. CPU-only consumers drop the backend entirely with
# `default-features = false`.
cuda-12000 = ["cuda", "dep:tract-cuda", "tract-cuda/cuda-12000"]
cuda-12010 = ["cuda", "dep:tract-cuda", "tract-cuda/cuda-12010"]
cuda-12020 = ["cuda", "dep:tract-cuda", "tract-cuda/cuda-12020"]
cuda-12030 = ["cuda", "dep:tract-cuda", "tract-cuda/cuda-12030"]
cuda-12040 = ["cuda", "dep:tract-cuda", "tract-cuda/cuda-12040"]
cuda-12050 = ["cuda", "dep:tract-cuda", "tract-cuda/cuda-12050"]
cuda-12060 = ["cuda", "dep:tract-cuda", "tract-cuda/cuda-12060"]
cuda-12080 = ["cuda", "dep:tract-cuda", "tract-cuda/cuda-12080"]
cuda-12090 = ["cuda", "dep:tract-cuda", "tract-cuda/cuda-12090"]
cuda-13000 = ["cuda", "dep:tract-cuda", "tract-cuda/cuda-13000"]
cuda-13010 = ["cuda", "dep:tract-cuda", "tract-cuda/cuda-13010"]
cuda-13020 = ["cuda", "dep:tract-cuda", "tract-cuda/cuda-13020"]

# Serve many concurrent sessions off one prepared model (`Runnable::autobatch`),
# and the `deploy` runtime, which prepares models as a deployment config says.
# Unstable: the shape of the surface is still being settled against the server
# it is meant for, so it is not covered by the no-breaking-change commitment.
unstable-autobatch = ["tract-core/deploy"]