[package]
name = "jouleclaw-omni"
version.workspace = true
edition.workspace = true
authors.workspace = true
license.workspace = true
repository.workspace = true
description = "JouleClaw omni-modal generation tier. Diffusion (40+ samplers), Stable Diffusion / SDXL / SD3 / Flux paths, MusicGen, Whisper, Gaussian3D, video, cross-modal fusion, MoE expert cache. Ported wholesale from efficient-genai; preserves the internal module structure for now (hal / inference / modalities / runtime / orchestrator / quality / server / tensor / weight_store / core). Future phases will split this into separate crates as the architecture stabilises."

[lib]
name = "jouleclaw_omni"
path = "src/lib.rs"

[features]
default = ["std"]
std = []

# Hardware backends
# cudarc refuses to build unless a CUDA version is chosen. Asking the build
# system (nvcc) beats pinning a version here: this fleet has 12.8, another
# host will have something else, and a pinned feature would break it.
cuda = ["dep:cudarc", "cudarc/cuda-version-from-build-system"]
metal = ["dep:metal", "dep:objc2"]
# HIP is opened at runtime with libloading, so no ROCm is needed to build:
# this compiles on every host and finds devices only where HIP is installed.
rocm = ["dep:libloading"]
vulkan = ["dep:ash", "dep:gpu-allocator"]

# JIT compilation
jit = ["dep:cranelift", "dep:cranelift-jit"]

# Model formats
safetensors = ["dep:safetensors"]
gguf = []

# Optimizations
simd = []
rayon = ["dep:rayon"]

# Server
server = ["dep:axum", "dep:tower-http", "dep:tracing-subscriber", "dep:image", "dep:uuid", "dep:base64"]

# Note: the jouledb feature from efficient-genai was dropped during the
# JouleClaw port. It path-depended on `joulesperbit/crates/joule-db-{weights,core}`
# (JoulesPerBit IP) which does not port into the public open standard.
# The `src/weight_store/` module remains under a `#[cfg(feature = "jouledb")]`
# gate so a downstream consumer who DOES have the JouleDB crates can re-enable
# the feature without forking jouleclaw-omni. The feature is declared to
# check-cfg below (after [features]) so the gate doesn't surface as an
# `unexpected_cfgs` warning when the feature is absent.

# Convenience: everything except cuda + jit (which often require external
# toolchains) so a developer can `cargo build --features dev` on bare
# macOS/Linux without provisioning extra SDKs.
dev = ["metal", "vulkan", "safetensors", "rayon", "simd", "server"]
full = ["cuda", "metal", "vulkan", "jit", "safetensors", "rayon", "simd", "server"]

[lints.rust]
unexpected_cfgs = { level = "warn", check-cfg = ['cfg(feature, values("jouledb"))'] }

[dependencies]
# JouleClaw foundation
jouleclaw-loader-gguf = { path = "../jouleclaw-loader-gguf" }

# Core
thiserror = "2.0"
tracing = "0.1"
parking_lot = "0.13"
dashmap = "6.2"
crossbeam = "0.8"
serde = { version = "1.0", features = ["derive"] }
serde_json = "1.0"

# Async runtime
tokio = { version = "1.54", features = ["rt-multi-thread", "sync", "macros", "time", "fs", "net", "io-util"] }
async-stream = "0.3"
futures = "0.3"

# Memory management
bumpalo = "3.20"
memmap2 = "0.9"

# Serialization (zero-copy)
rkyv = { version = "0.8" }
bytemuck = { version = "1.25", features = ["derive"] }

# Math
half = { version = "2.7", features = ["bytemuck", "num-traits"] }
num-traits = "0.2"
rand = "0.10"
rustfft = "6.4"

# Text processing
regex = "1.13"
tokenizers = "0.23"

# Native libc bindings
libc = "0.2.191"

# Optional: CUDA
cudarc = { version = "0.19", optional = true }
libloading = { version = "0.9", optional = true }

# Optional: Metal (macOS/iOS) - using objc2 for modern Rust bindings
metal = { version = "0.33", optional = true }
objc2 = { version = "0.6", optional = true }

# Optional: Vulkan
ash = { version = "0.38", optional = true }
gpu-allocator = { version = "0.28", optional = true, features = ["vulkan"] }

# Optional: JIT compiler
cranelift = { version = "0.137", optional = true }
cranelift-jit = { version = "0.137", optional = true }

# Optional: Model formats
safetensors = { version = "0.8", optional = true }

# Optional: Parallelism
rayon = { version = "1.12", optional = true }

# Optional: Server stack
axum = { version = "0.8", optional = true }
tower-http = { version = "0.7", features = ["trace", "cors", "fs"], optional = true }
tracing-subscriber = { version = "0.3", optional = true }
image = { version = "0.25", optional = true }
uuid = { version = "1.29", features = ["v4"], optional = true }
base64 = { version = "0.23", optional = true }

# HTTP client for model_fetch (HF Hub / S3 downloads).
reqwest = { version = "0.13", features = ["rustls", "stream", "json"], default-features = false }

[dev-dependencies]
proptest = "1.12"
tokio-test = "0.4"
# tests/metal_kernels.rs dispatches shaders directly and reads f16 back.
half = { version = "2.7", features = ["bytemuck"] }

# The metal crate is macOS-only, so it cannot be an unconditional
# dev-dependency: the linux fabric jobs would fail to resolve it.
[target.'cfg(target_os = "macos")'.dev-dependencies]
metal = "0.33"