Struct StackedExpertLinear 

pub struct StackedExpertLinear<B: Backend + BackendQuantMarlin> { /* private fields */ }

View into a single column-slice of a shared stacked GPTQ store.

Phase 3e/2: backed by a Box<dyn Linear<B>> produced by B::make_stacked_expert_linear (CUDA: CudaMarlinStackedExpertLinear; CPU: CpuGptqLinear over a sliced row range). The store itself is Arc<B::GptqStore> so cloning a view is cheap; dropping all views drops the underlying store.

Implementations

impl<B: Backend + BackendQuantMarlin> StackedExpertLinear<B>

pub fn new( stack: Arc<dyn MarlinExpertStack<B>>, expert_offset: usize, expert_n: usize, ) -> Result<Self>

Phase C step 4b: takes the trait-object MarlinExpertStack directly (was Arc<B::GptqStore> + B::make_stacked_expert_linear).

pub fn new_with_bias( stack: Arc<dyn MarlinExpertStack<B>>, expert_offset: usize, expert_n: usize, bias: &[f32], ) -> Result<Self>

Trait Implementations

impl<B: Backend + BackendQuantMarlin> Linear<B> for StackedExpertLinear<B>

fn in_features(&self) -> usize

fn out_features(&self) -> usize

fn forward( &self, ctx: &mut B::Context, input: &B::Buffer, out: &mut B::Buffer, m: usize, )

Append GEMM work onto ctx. Caller flushes the context when results must be materialised.