[package]
name = "rtx-distributed"
version = "1.0.0"
edition = "2024"
rust-version = "1.92"
authors = ["RustyTorch Team"]
license = "MIT OR Apache-2.0"
repository = "https://github.com/rustytorch/rustytorch"
description = "Distributed training support for RustyTorch++"

[dependencies]
# Core dependencies (explicit versions - crate excluded from workspace)
anyhow = "1.0"
thiserror = "2"
tracing = "0.1"
tokio = { version = "1.0", features = ["full"] }
serde = { version = "1.0", features = ["derive"] }
bincode = "1.3"
serde_json = "1.0"

# HPC-AI integration (optional) - path dep removed: hpc-channels not in this workspace
# hpc-channels = { path = "../../../../hpc-channels", features = ["rkyv-codec"], optional = true }

# RNCCL - Pure Rust collective communications (optional)
rnccl-core = { path = "../../vendor/rnccl/rnccl-core", optional = true }
rnccl-collectives = { path = "../../vendor/rnccl/rnccl-collectives", default-features = false, features = ["metal", "transport"], optional = true }
rnccl-transport = { path = "../../vendor/rnccl/rnccl-transport", optional = true }
rnccl-bootstrap = { path = "../../vendor/rnccl/rnccl-bootstrap", optional = true }

# Local crates
rtx-runtime = { path = "../../core/rtx-runtime" }
rtx-tensor = { path = "../../core/rtx-tensor" }

# NCCL binding via cudarc - safer and more comprehensive (optional for non-CUDA builds)
cudarc = { version = "0.18.2", features = ["std", "driver", "runtime", "nvrtc", "cublas", "cublaslt", "nccl", "cudnn", "cusparse", "cusolver", "cufile", "curand", "cuda-13000", "f16"], optional = true }

# Network and process management
mpi = { version = "0.6", optional = true }
crossbeam = "0.8"
parking_lot = "0.12"
uuid = { version = "1.0", features = ["v4", "serde"] }

# Configuration
config = "0.14"
clap = { version = "4.0", features = ["derive"] }

# Async trait support
async-trait = "0.1"

# Additional dependencies for multi-node support
rand = "0.8"
futures = "0.3"

# Half-precision floating point
half = "2.3"

# System info
hostname = "0.4"

# Time and date handling
chrono = { version = "0.4", features = ["serde"] }

[dev-dependencies]
proptest = "1.4"
criterion = { version = "0.5", features = ["html_reports", "csv_output"] }
tempfile = "3.8"
tracing-subscriber = "0.3"

[features]
default = []
cuda = ["dep:cudarc"]
nccl = ["cuda", "cudarc/nccl"]
rccl = []  # RCCL for AMD ROCm GPUs - requires librccl.so
rnccl = ["dep:rnccl-core", "dep:rnccl-collectives", "dep:rnccl-transport", "dep:rnccl-bootstrap"]  # Rust-native NCCL from rnccl
mpi = ["dep:mpi"]
# HPC-AI low-latency channels (optional) - removed: dep not available
# hpc-channels = ["dep:hpc-channels"]

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

[[bin]]
name = "benchmark_backends"
path = "src/bin/benchmark_backends.rs"

[lints]
workspace = true