Struct KernelEnvelope 

pub struct KernelEnvelope { /* private fields */ }

Builder/configuration for a single run_kernel invocation.

Existing actors that call the free run_kernel function are unaffected — that path still constructs a default envelope internally. Actors that want observability migrate to KernelEnvelope::new(lib).with_trace(..).with_nvtx(..).run_kernel(..).

Implementations

impl KernelEnvelope

pub fn new(lib_tag: &'static str) -> Self

Construct a trace-less, NVTX-less envelope tagged with the given library. op_name defaults to lib_tag and can be refined via Self::with_op_name.

pub fn with_op_name(self, op_name: &'static str) -> Self

Override the op identifier surfaced to the trace callback (e.g. "sgemm", "conv2d_forward").

pub fn with_dtype(self, dtype: &'static str) -> Self

Tag the envelope with a dtype name (e.g. "f32", "f16"). Surfaced to trace callbacks as KernelInfo::dtype.

pub fn with_trace(self, trace: Arc<dyn KernelTrace>) -> Self

Attach a KernelTrace callback. Cloning Arc<dyn KernelTrace> is cheap; the same object can be shared across many envelopes.

pub fn with_nvtx(self, name: &'static str) -> Self

Attach an NVTX range label. No-op unless the nvtx cargo feature is enabled.

pub fn run_kernel<O, KA, F>( self, stream: &Arc<CudaStream>, completion: &Arc<dyn CompletionStrategy>, output: O, reply: Sender<Result<O, GpuError>>, enqueue: F, )

where O: Send + 'static, KA: Send + 'static, F: FnOnce() -> Result<KA, GpuError>,

Builder-style equivalent of the free run_kernel function with observability hooks layered in.

Behaviour without a trace and without an NVTX range is byte-for-byte identical to run_kernel.

Trait Implementations

impl Clone for KernelEnvelope

fn clone(&self) -> KernelEnvelope

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for KernelEnvelope

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

Formats the value using the given formatter.