Cargo.toml

[workspace]
members = [
    "crates/lumen-core",
    "crates/lumen-flags",
    "crates/lumen-testkit",
    "crates/turboquant-cache",
    "crates/lumen-mlx",
    "crates/lumen-diffusion",
    "crates/lumen-server",
    "crates/lumen-app",
    "xtask",
]
# The libFuzzer targets need nightly + sanitizer flags and would drag
# libfuzzer-sys into every workspace build graph. They are driven exclusively
# through `cargo xtask fuzz`; the deterministic twin of each target runs in
# tier 0 as `crates/lumen-mlx/tests/tool_surface_fuzz.rs`.
exclude = ["fuzz"]
# lumen-app pulls in the Tauri + webview stack (~hundreds of crates). Excluding
# it from `default-members` keeps `cargo build` / `cargo test` at the workspace
# root fast — build the desktop app explicitly with `cargo build -p lumen-app`
# (dev) or `cargo tauri build` (release bundle).
default-members = [
    "crates/lumen-core",
    "crates/lumen-flags",
    "crates/lumen-testkit",
    "crates/turboquant-cache",
    "crates/lumen-mlx",
    "crates/lumen-diffusion",
    "crates/lumen-server",
]
resolver = "2"

[workspace.package]
edition = "2024"
rust-version = "1.85"
license = "MIT"

# `coverage` is set by `cargo llvm-cov`, and `lumen_core::always!()` /
# `never!()` expand a `#[cfg(coverage)]` arm at every call site — including in
# other crates — so the cfg has to be known workspace-wide or every consumer
# warns. Declared here rather than allowed, so a genuine typo in a cfg name is
# still caught.
[workspace.lints.rust]
unexpected_cfgs = { level = "warn", check-cfg = ['cfg(coverage)', 'cfg(coverage_nightly)'] }

# Two clippy lints are allowed workspace-wide, on purpose and with a reason.
# Everything else is expected to be zero so `cargo xtask gate` can fail on a
# new warning; a lint left warning forever teaches people to skim the output,
# which is how the 330-warning backlog happened in the first place.
[workspace.lints.clippy]
# 27 sites. A kernel entry point takes what the kernel takes: queries, keys,
# values, scale, mask, cache, offsets, window, flags. Bundling them into a
# params struct does not reduce the arity, it hides it behind a constructor
# and adds a copy on a path where the copy is measurable. The threshold is a
# heuristic for business logic and this is not business logic.
too_many_arguments = "allow"
# 18 sites, almost all `HashMap<String, Vec<(usize, Array)>>`-shaped weight
# tables built once at load time. A type alias would name the shape without
# explaining it, and the shape is the documentation: what the loader returns
# is exactly what the checkpoint contains.
type_complexity = "allow"
# 12 sites, every one a matrix kernel where the loop index is the algorithm:
# `for i in 0..dim { let row = &m[i*dim..(i+1)*dim]; y[i] = row·x }` uses `i`
# to slice the operand AND to place the result. The iterator rewrite has to
# thread both, and reads worse than the arithmetic it replaces.
needless_range_loop = "allow"
# 2 sites, both runner enums that exist once per process for its lifetime.
# Boxing the big variant to even them up would add an indirection on the
# decode path to save bytes that are never allocated more than once.
large_enum_variant = "allow"

[workspace.dependencies]
# TurboQuant workspace crates
lumen-core = { path = "crates/lumen-core" }
# Test-only; add under `[dev-dependencies]`, never `[dependencies]`.
lumen-testkit = { path = "crates/lumen-testkit" }
turboquant-cache = { path = "crates/turboquant-cache" }

# Linear algebra & numerics
faer = "0.24"
rand = "0.10.1"
rand_distr = "0.6.0"

# Metal GPU (objc2 stack)
objc2-metal = "0.3.2"
objc2-metal-performance-shaders = "0.3.2"
objc2-metal-performance-shaders-graph = "0.3.2"

# Tokenizer & model loading
tokenizers = "0.23.1"
hf-hub = "0.5"
safetensors = "0.7.0"

# Serialization
serde = { version = "1.0.228", features = ["derive"] }
serde_json = "1.0.150"

# HTTP server
atomic_http = { version = "0.14.0", features = ["arena", "router"] }

# Async runtime
tokio = { version = "1.52.3", features = ["full"] }

# Error handling
anyhow = "1.0.102"
thiserror = "2.0.18"

# ── Local-dev override (uncomment when you want to edit the mlx-rs fork
#    in place instead of going through the git URL pin). A clean clone of
#    lumen-rs uses the git refs above; only enable this path override if
#    you have the sibling checkout.
# [patch."https://github.com/rabbitson87/mlx-rs"]
# mlx-rs  = { path = "../../../mlx-rs/mlx-rs" }
# mlx-sys = { path = "../../../mlx-rs/mlx-sys" }