Struct WriteSlicePlan 

pub struct WriteSlicePlan<T: DeviceRepr + Copy + 'static, const N: usize> { /* private fields */ }

write_slice plan.

dest[start_0..end_0, ..., start_{N-1}..end_{N-1}] = source — assign (not accumulate). Drives Fuel team’s persistent KV-cache append.

When to use: in-place per-axis range write. Distinct from ScatterAddPlan (which accumulates per-index) and from PadPlan (which produces a larger output tensor). No backward — non-differentiable.

Dtypes: every byte-aligned element kind in baracuda’s element bank — f16, bf16, f32, F32Strict, f64, i32, i64, Bool, S8, U8, Fp8E4M3, Fp8E5M2, Complex32, Complex64. Plus nibble-packed S4 / U4 with the even-alignment constraint on the innermost axis. Bin (1-bit packed) is out of scope.

Shape limits: rank in [1, 8]; per-axis 0 ≤ start ≤ end ≤ dest_shape[d]; source_shape[d] = end - start.

Workspace: none.

Precision guarantee: deterministic, bit-stable, bit-exact (no arithmetic — pure memcpy / index + copy).

Implementations

impl<T: DeviceRepr + Copy + 'static, const N: usize> WriteSlicePlan<T, N>

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

Pick a kernel for desc. Validates rank, range bounds, source shape consistency, dtype coverage, and the nibble-axis-alignment constraint for S4 / U4. Detects the available fast path.

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

Validate args against the descriptor: shapes match, device buffers are large enough.

pub fn workspace_size(&self) -> usize

Workspace size in bytes. Always 0.

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( &self, stream: &Stream, _workspace: Workspace<'_>, args: WriteSliceArgs<'_, T, N>, ) -> Result<()>

Launch on stream. workspace is ignored (always zero).