Enum Device 

pub enum Device {
    Cpu,
    Wgpu(Arc<WgpuContext>),
}

Where a tensor lives and where its ops execute.

Variants

Cpu

Reference implementation used for testing and verification.

Wgpu(Arc<WgpuContext>)

Production backend: WebGPU via wgpu (Vulkan/Metal/D3D12/Browser).

Implementations

impl Device

pub fn wgpu() -> Result<Device>

Available on non-WebAssembly only.

Create the default WebGPU device (highest-performance adapter). Sync facade — native only; on wasm use Device::wgpu_async.

pub async fn wgpu_async() -> Result<Device>

Async device creation — the primary form (roadmap v4, pitfall 14); works on native and wasm32/browser.

pub fn dispatch_scope(&self) -> Option<DispatchScope>

Batch every dispatch made while the returned guard is alive into one command buffer and one submit. None on the CPU backend, which has no queue to batch.

Hold it for the span of a logical step — a decode step issues ~100 kernels, and a submit each leaves the GPU waiting on the CPU between every one of them. Readbacks flush automatically, so a scope can wrap code that reads results without special care.

pub fn describe(&self) -> String

Human-readable adapter description.

pub fn max_binding_bytes(&self) -> usize

Largest single buffer binding this device supports, in bytes. Large weights (GPT-2’s wte) are row-chunked to stay under this.

pub fn same_as(&self, other: &Device) -> bool

Two devices are compatible when tensors on them can be combined.

Trait Implementations

impl Clone for Device

fn clone(&self) -> Device

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for Device

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.