[package]
name = "rtx-backend-cuda"
version = "1.0.0"
edition = "2024"
rust-version = "1.92"
authors = ["RustyTorch++ Team"]
license = "Apache-2.0"
description = "CUDA backend implementation for RustyTorch++ with hand-optimized kernels"

[dependencies]
# Backend trait definitions
rtx-backend = { workspace = true }

# Core utilities
anyhow = { workspace = true }
thiserror = "2"
tracing = { workspace = true }
parking_lot = { workspace = true }
once_cell = "1.19"

# Numeric types
half = { version = "2.3", features = ["serde"] }
serde = { version = "1.0", features = ["derive"] }

# CUDA acceleration via cudarc (optional - only builds on CUDA-capable systems)
cudarc = { workspace = true, optional = true }

# Flash Attention integration (optional)
rtx-flash-attention = { workspace = true, features = ["cuda"], optional = true }

# Random number generation
rand = { workspace = true }
rand_distr = { workspace = true }

[dev-dependencies]
proptest = "1.4"
tokio-test = "0.4"

[features]
# No default - this crate is CUDA-specific, only use on CUDA-capable systems
default = []
cuda = ["dep:cudarc"]
# Enable rtx-flash-attention integration for optimized attention
flash-attention = ["dep:rtx-flash-attention"]

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

[lints]
workspace = true