Struct ConcatPlan 

pub struct ConcatPlan<T: Element, const N: usize> { /* private fields */ }

2-input concat plan.

y = cat(a, b, dim=concat_dim) (PyTorch torch.cat).

When to use: residual joins, KV-cache concat in attention, any 2-input concatenation. Variable-arity (N-input) concat needs a separate plan with device-side pointer arrays — deferred. Pair with ConcatBackwardPlan.

Dtypes: {f32, f64, f16, bf16}.

Shape limits: rank in [1, 8]; a_shape and b_shape must match on every axis except concat_dim; concat_dim ∈ [0, N).

Workspace: none.

Precision guarantee: deterministic, bit-stable, bit-exact (pure load + store).

Implementations

impl<T: Element, const N: usize> ConcatPlan<T, N>

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

Pick a kernel for desc.

pub fn can_implement(&self, args: &ConcatArgs<'_, T, N>) -> Result<()>

Validate args.

pub fn workspace_size(&self) -> usize

Workspace size in bytes. Always 0 for the trailblazer.

pub fn sku(&self) -> KernelSku

Identity of the kernel this plan picked.

pub fn precision_guarantee(&self) -> PrecisionGuarantee

Numerical guarantees.

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

Launch.