Enum Device 

pub enum Device {
    Gpu(Arc<Context>),
    Cpu,
    Remote(String),
    BrowserWorker(Arc<WsConn>),
    Npu(Arc<dyn NpuBackend>),
}

A compute device in the fabric.

Variants

Gpu(Arc<Context>)

Cpu

Remote(String)

A remote worker reached over TCP — a cloud node, or (with a WS bridge) a browser tab. Work crosses the boundary as host buffers, exactly like the GPU/CPU hop, so it’s the same fabric.

BrowserWorker(Arc<WsConn>)

A browser tab reached over a WebSocket, computing on the tab’s WebGPU. Same op frames as Remote, different transport — this is the fabric physically spanning cloud+local+browser.

Npu(Arc<dyn NpuBackend>)

An NPU reached through an execution-provider backend (CoreML/ANE, DirectML-QNN, OpenVINO, WebNN). WebGPU cannot target an NPU, so this only exists when a real backend is plugged in — it never falls back to the GPU/CPU silently.

Implementations

impl Device

pub fn name(&self) -> String

pub fn bmm( &self, a: &[f32], b: &[f32], batch: usize, m: usize, k: usize, n: usize, ) -> Vec<f32>

Batched matmul [batch,m,k] · [k,n] → [batch,m,n] on host data.

pub fn linear_relu( &self, x: &[f32], rows: usize, w: &[f32], in_: usize, out: usize, ) -> Vec<f32>

One MLP layer: relu(x[rows,in] · w[in,out]).