Struct DequantizePerTokenBackwardPlan 

pub struct DequantizePerTokenBackwardPlan<TIn: Element, TOut: IntElement> { /* private fields */ }

dequantize_per_token backward plan.

Straight-through linear: dq_FP[n, d] = scale[n] * dy[n, d]. Int input is non-differentiable; FP gradient flows through.

When to use: backward for DequantizePerTokenPlan.

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

Shape limits: rank-2 [N, D].

Workspace: none.

Precision guarantee: deterministic, bit-stable.

Implementations

impl<TIn: Element, TOut: IntElement> DequantizePerTokenBackwardPlan<TIn, TOut>

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

Pick a kernel for desc.

pub fn can_implement( &self, args: &DequantizePerTokenBackwardArgs<'_, TIn, TOut>, ) -> Result<()>

Validate args.

pub fn workspace_size(&self) -> usize

Workspace bytes — none.

pub fn sku(&self) -> KernelSku

Identity.

pub fn precision_guarantee(&self) -> PrecisionGuarantee

Numerical guarantees.

pub fn run( &self, stream: &Stream, _workspace: Workspace<'_>, args: DequantizePerTokenBackwardArgs<'_, TIn, TOut>, ) -> Result<()>

Launch.