Struct EmbeddingBackwardPlan 

pub struct EmbeddingBackwardPlan<T: Element> { /* private fields */ }

embedding_backward plan.

Adjoint of crate::EmbeddingPlan: dweight[indices[n], :] += dout[n, :] via atomicAdd. Rows where indices[n] == padding_idx (or negative / out-of-range) are skipped.

When to use: backward for EmbeddingPlan.

Dtypes: {f32, f64} only — atomicAdd is native-FP.

Shape limits: dweight is [V, D], dout is [N, D], indices is [N], all extents non-negative.

Workspace: none. Caller MUST zero dweight before launch (or pre-populate to accumulate into a running gradient).

Precision guarantee: non-deterministic — atomicAdd ordering varies between launches.

Implementations

impl<T: Element> EmbeddingBackwardPlan<T>

pub fn select( _stream: &Stream, desc: &EmbeddingBackwardDescriptor, _pref: PlanPreference, ) -> Result<Self>

Pick a kernel for desc.

pub fn can_implement<I: IndexElement>( &self, args: &EmbeddingBackwardArgs<'_, T, I>, ) -> Result<()>

Validate args.

pub fn workspace_size(&self) -> usize

Workspace size in bytes (zero).

pub fn sku(&self) -> KernelSku

Identity of the kernel this plan picked.

pub fn precision_guarantee(&self) -> PrecisionGuarantee

Numerical guarantees for this plan’s kernel.

pub fn run<I: IndexElement>( &self, stream: &Stream, _workspace: Workspace<'_>, args: EmbeddingBackwardArgs<'_, T, I>, ) -> Result<()>

Launch.

Phase 11.5: generic over I: IndexElement.