A high-performance inference engine for AI models
pub(crate) mod common;
#[cfg(backend = "metal")]
pub(crate) mod metal;
#[cfg(backend = "vulkan")]
pub mod vulkan;
#[cfg(backend = "webgpu")]
pub(crate) mod webgpu;
#[cfg(backend = "cpu")]
pub(crate) mod cpu;
pub use common::Backend;
pub trait BackendSelection {
type Output;
type Error;
fn select<B: Backend>(self) -> impl Future<Output = Result<Self::Output, Self::Error>> + Send;
}
pub async fn select_backend<S: BackendSelection>(
selection: S,
unknown: S::Error,
) -> Result<S::Output, S::Error> {
let default = if cfg!(backend = "metal") {
"metal"
} else if cfg!(backend = "webgpu") {
"webgpu"
} else if cfg!(backend = "cpu") {
"cpu"
} else {
unreachable!()
};
// TODO: remove magic env var
match std::env::var("UZU_BACKEND").map(|s| s.to_lowercase()).as_deref().unwrap_or(default) {
#[cfg(backend = "metal")]
"metal" => selection.select::<metal::Metal>().await,
#[cfg(backend = "webgpu")]
"webgpu" => selection.select::<webgpu::WebGPU>().await,
#[cfg(backend = "cpu")]
"cpu" => selection.select::<cpu::Cpu>().await,
_ => Err(unknown),
}
}