Cargo.toml

[workspace]
members = [
    "candle-usage",
    "server",
    "models",
    "locals",
    "oneapi",
    "cudarust",
    "engine",
    "tile",
    "kv-manager",
]
# Excluded crates resolve their dependencies standalone:
# - rlx: rlx-cuda requires cudarc's `dynamic-loading` feature, candle requires
#   `dynamic-linking`, and cudarc forbids enabling both in one build graph.
# - oxide/first, kernels: depend on cuda-oxide crates that need its pinned
#   nightly toolchain (cuda-macros uses `#![feature]`), so they cannot build
#   with the stable toolchain the root workspace uses.
# - gpgpu: not a workspace member; excluded to keep it buildable standalone.
exclude = ["rlx-demo", "gpgpu", "oxide/first", "kernels"]
resolver = "2"

[workspace.dependencies]
axum = "0.7"
# One shared cudarc definition for every workspace member. Cargo unifies
# features across the graph, and candle-core (git) enables cudarc with
# `dynamic-linking` + no default features; any member enabling the default
# `dynamic-loading` feature makes cudarc's build script panic ("Both
# `dynamic-loading` and `dynamic-linking` features are active").
cudarc = { version = "0.19.8", default-features = false, features = [
    "std",
    "driver",
    "cublas",
    "cublaslt",
    "curand",
    "nvrtc",
    "f16",
    "cuda-version-from-build-system",
    "dynamic-linking",
] }
tokenizers = { version = "0.22.0", features = ["http"] }
tokio = { version = "1.48", features = ["full"] }
serde = { version = "1", features = ["derive"] }
serde_json = "1"

# candle
candle-transformers = { version = "0.11.0", git = "https://github.com/huggingface/candle.git", features = [
    "cuda",
] }
candle-nn = { version = "0.11.0", git = "https://github.com/huggingface/candle.git", features = [
    "cuda",
] }
candle-core = { version = "0.11.0", git = "https://github.com/huggingface/candle.git", features = [
    "cuda",
] }
log = "0.4"
env_logger = "0.11"