[package]
name = "eoc-local"
version.workspace = true
edition.workspace = true
license.workspace = true
repository.workspace = true
authors.workspace = true
description = "EOC local-inference backends — llama.cpp (GGUF), MLX (Apple Silicon), MLC-LLM (TVM), ONNX Runtime — behind a single NeuralBackend trait with measured joule attribution."

[features]
# Default build pulls in *no* heavy native deps. The crate is usable as
# a pure-Rust GGUF / tokenizer / sampler / registry library out of the
# box; the actual inference backends are feature-gated because each one
# brings its own large native dependency.
default = []

# Backend feature flags. Each one enables the matching module and its
# associated native dependency. They are independent — operators turn
# on whichever stacks they have installed.
llamacpp = ["dep:llama-cpp-2"]
mlx      = ["dep:mlx-rs"]
# MLC enables the module surface. The tvm/mlc-llm runtime is *not*
# pulled in as a direct crates.io dependency because the published
# `tvm-rt = 0.1.0-alpha` has a broken transitive `heim` requirement
# that pins ancient `memchr`. Operators using this backend should add
# their preferred mlc-llm / tvm-rt crate to a workspace `[patch.crates-io]`
# section and wire the calls inside `src/mlc.rs` against it.
mlc      = []
onnx     = ["dep:ort", "dep:tokenizers", "dep:ndarray"]
gguf     = []
# Convenience — pull in just the tokenizer wrapper without an inference
# backend (handy for tooling).
tokenizer-only = ["dep:tokenizers"]

[dependencies]
eoc-core.workspace = true
eoc-neural.workspace = true
eoc-meter.workspace = true
async-trait.workspace = true
serde.workspace = true
serde_json = "1"
thiserror.workspace = true
tokio = { version = "1", default-features = false, features = ["macros", "sync", "rt", "rt-multi-thread"] }
tracing = "0.1"

# Optional native backends. Versions pinned to the latest published
# tags on crates.io as of the EOC 0.1 cut. Operators wishing to vendor
# a different revision can override via [patch.crates-io] in their
# workspace root.
llama-cpp-2  = { version = "0.1", optional = true }
mlx-rs       = { version = "0.34", optional = true }
ort          = { version = "2.0.0-rc.15", optional = true, default-features = false, features = ["ndarray", "std"] }
tokenizers   = { version = "0.23", optional = true, default-features = false, features = ["onig"] }
ndarray      = { version = "0.17", optional = true, features = ["std"] }

[dev-dependencies]
tokio = { version = "1", default-features = false, features = ["macros", "rt-multi-thread", "test-util"] }
tempfile = "3"
# Re-declared as dev-deps so integration tests (under `tests/`) can refer
# to them by their public name.
eoc-core.workspace = true
eoc-neural.workspace = true
eoc-meter.workspace = true

# MLX is hard-gated on macOS — surface a clear build error if someone
# tries `--features mlx` on Linux. The cfg also guards the module body.
[target.'cfg(not(target_os = "macos"))'.dependencies]
# (intentionally empty; mlx-rs is only depended on via the `mlx` feature
# above, and the `mlx` module is `#[cfg(target_os = "macos")]` gated.)

[package.metadata.docs.rs]
all-features = false
features = ["gguf", "tokenizer-only"]