Struct KernelOpsDispatch 

pub struct KernelOpsDispatch<'a> { /* private fields */ }

Dispatch wrapper that tries KernelOps first, then falls back to TensorOps for operations that have a TensorOps equivalent.

This enables gradual migration: callers use the dispatch without caring which path actually runs.

Implementations

impl<'a> KernelOpsDispatch<'a>

pub fn new( kernel_ops: Option<&'a dyn KernelOps>, tensor_ops: &'a dyn TensorOps, ) -> Self

pub fn rms_norm( &self, input: &TensorRef, weight: &TensorRef, eps: f32, ) -> Result<TensorRef>

RMS norm: prefer KernelOps::NormOps, fall back to TensorOps::rms_norm.

pub fn gelu(&self, input: &TensorRef) -> Result<TensorRef>

GELU: prefer KernelOps::ActivationOps, fall back to TensorOps::gelu.

pub fn silu(&self, input: &TensorRef) -> Result<TensorRef>

SiLU: prefer KernelOps::ActivationOps::silu_mul is fused so there is no direct TensorOps equivalent. This helper exposes the non-fused TensorOps::silu for callers that only need plain SiLU.

pub fn silu_mul(&self, gate: &TensorRef, up: &TensorRef) -> Result<TensorRef>

Fused SiLU-multiply (SwiGLU building block). Falls back to silu(gate) * up via TensorOps when kernel is unavailable.

pub fn linear(&self, input: &TensorRef, weight: &TensorRef) -> Result<TensorRef>

Dense linear (no bias). Falls back to TensorOps::matmul.

pub fn softmax(&self, input: &TensorRef, dim: i32) -> Result<TensorRef>

Softmax: always via TensorOps (no kernel sub-trait for plain softmax).

pub fn kernel_ops(&self) -> Option<&'a dyn KernelOps>

Access the underlying KernelOps (if any) for ops that have no TensorOps fallback (e.g. rotary_embedding, attention, sampling).

pub fn tensor_ops(&self) -> &'a dyn TensorOps

Access the underlying TensorOps.